JP5027859B2 - Signal identification method and signal identification apparatus - Google Patents

Description

  The present invention relates to a signal identification method and a signal identification device used for inspecting the state of an inspection object.

  Conventionally, as a signal identification device for inspecting the state of an inspection object, whether or not the inspection object is in a normal state by inputting a plurality of feature amounts extracted from the inspection signal to a learned neural network There is known an apparatus for performing (see, for example, Patent Document 1). This signal identification device needs to learn a neural network before making the above determination. For this reason, the conventional signal identification device uses, as a learning signal, a normal signal (normal sample) detected when the inspection target is in a normal state and an abnormal signal (normal sample) detected when the inspection target is in an abnormal state ( Abnormal samples), a plurality of feature quantities are extracted from each of the normal signal and the abnormal signal, and the plurality of feature quantities are input to the neural network to learn the neural network.

  As another example, a signal identification device that learns a neural network using only normal signals as learning signals is known (see, for example, Patent Document 2). Since this signal identification device does not require an abnormal signal that is difficult to collect as compared with a normal signal, a learning signal can be easily prepared, and a neural network can be easily learned.

Japanese Patent No. 3778177 Japanese Patent No. 4241818

  However, in the conventional signal identification device, it is easy to determine which feature amount of the extraction range among the plurality of feature amounts extracted for each extraction range (extraction condition) from each learning signal has a large influence on the determination. Therefore, it was difficult to determine the above combination by trial and error in determining a combination of extraction ranges of feature amounts necessary for inspection.

  Further, in the conventional signal identification device, the inspection accuracy is not sufficient for the user's request, and in particular, when a signal completely different from any learning signal is input as the inspection signal, an erroneous determination may occur. .

  The present invention has been made in view of the above points, and an object of the present invention is to provide a signal identification method and a signal identification apparatus capable of automatically selecting a combination of extraction ranges as inspection regions and improving inspection accuracy. There is to do.

  The invention of the signal identification method according to claim 1 includes learning for creating an inspection reference for inspecting a state of an inspection object using a plurality of learning signals including a normal signal and an abnormal signal, and the inspection object A signal identification method for performing an inspection for determining the state of the inspection object using an inspection signal representing the state of the inspection and the inspection standard, and is preset from the plurality of learning signals at the time of learning A first step of extracting a feature amount for each extraction range; a second step of selecting a combination of extraction ranges used for the inspection as an inspection region using the feature amount of each extraction range of each learning signal; A third step of setting the inspection standard using the feature amount of each extraction range of the inspection region, and extracting a feature amount for each extraction range from the inspection signal at the time of inspection; From the feature amount, A fourth step of selecting a feature amount of each extraction range of the region, and inspecting the state of the inspection target using the selected feature amount and the inspection reference, and in the second step, each learning Each time a weight is set for a signal for use, and each time the weight is changed, the feature amount of each learning signal is used in each extraction range, and each feature amount is checked against the determination criterion while changing the criterion. Thus, it is determined whether or not each learning signal is a normal signal, and an erroneous determination rate is calculated using the weight of the erroneously determined learning signal for each of the determination criteria. A discriminator candidate to which a discriminator candidate to which a judgment criterion that minimizes the misjudgment rate is applied is created and the weight is changed, and the discriminator candidate having the smallest misjudgment rate among the discriminator candidates created for each extraction range. And select The reliability is calculated using the misjudgment rate for each separate device, and the combination of the extraction ranges of the discriminators is set as the inspection region. In the third step, each extraction range of the inspection region is included in each learning signal. Each time, the product of the feature value and the reliability is calculated as learning data, and each learning data is input to a competitive learning type neural network to learn the competitive learning type neural network, and the competitive learning A clustering map in which the output layer neurons of the neural network are associated with categories, the inspection standard is set using the created clustering map, and in the fourth step, the inspection region of the inspection signal in the inspection signal is set. The product of the feature value of each extraction range and the reliability is calculated as inspection data, and the inspection data and the inspection standard To determine whether or not the inspection object is in a normal state.

  According to the present invention, using the feature amount of each learning signal, it is determined whether each learning signal is a normal signal by collating each feature amount with the above determination criterion while changing the determination criterion. By calculating the misjudgment rate at that time and selecting a combination of extraction ranges that reduces the misjudgment rate as the inspection region, it is possible to automatically select a combination of extraction ranges to be the inspection region.

  Further, according to the present invention, the competitive learning type neural network is trained using the product of the feature amount of each learning signal and the reliability of each discriminator at the time of learning, and the feature amount of the inspection signal and each discrimination at the time of inspection Inspection accuracy can be improved by determining the state of the inspection object using the product of the reliability of the device.

  The signal identification method according to a second aspect of the present invention is the signal identification method according to the first aspect, wherein in the second step, each time the weight is changed, the feature amount of each learning signal is increased or decreased in each extraction range. Are sorted in order, each sorted feature quantity is set as a threshold value in order, the magnitude relationship between each feature quantity and the threshold value is compared, and each feature quantity is checked against the judgment criterion while changing the judgment criterion. To determine whether each learning signal is a normal signal, calculate the sum of the weights of erroneously determined learning signals as the erroneous determination rate for each of the determination criteria, and select the discriminator Each time, the weight of the learning signal erroneously determined by the discriminator is increased. In the third step, after learning by the competitive learning type neural network, each learning data is transferred to the competitive learning type neural network. For each learning data, in each output layer neuron on the clustering map, find the distance between the learning data and the weight vector, and for each output layer neuron on the clustering map Then, a variance determined by a plurality of distances obtained from each learning data and the weight vector, and a Gaussian function defined by an average value as the weight vector are set, and a Gaussian for each output layer neuron on the clustering map is set. After the function is set, each learning data is re-input to the competitive learning type neural network, and for each learning data, the sum of Gaussian function values of all output layer neurons on the clustering map is obtained, and all learning is performed. The inspection standard from the distribution of the sum of the Gaussian function values for the data In the fourth step, the inspection data is input to the Gaussian function to obtain a sum of Gaussian function values related to the inspection data, and the Gaussian function value related to the inspection data is calculated. When the sum is less than the inspection threshold, it is determined that the inspection target is not in a normal state.

  According to the present invention, when an inspection signal that is an abnormal signal or an inspection signal that is far from a learning signal is input, it is determined that the inspection target is not in a normal state, and it is erroneously determined that the inspection target is in a normal state. Since the determination can be prevented, the inspection accuracy can be improved.

  The signal identification method according to a third aspect of the present invention is the signal identification method according to the second aspect, wherein in the fourth step, when the sum of Gaussian function values related to the inspection data is less than the inspection threshold, the inspection object is It is determined that the inspection target is in an abnormal state, and when the sum of Gaussian function values related to the inspection data is equal to or greater than the inspection threshold value, it is determined that the inspection target is in a normal state.

  According to the present invention, when an inspection signal that is an abnormal signal or an inspection signal that is far from a learning signal is input, it is determined that the inspection target is in an abnormal state, and it is erroneously determined that the inspection target is in a normal state. Since the determination can be prevented, the inspection accuracy can be further improved.

  The signal identification method according to a fourth aspect of the present invention is the signal identification method according to the second aspect, wherein, in the fourth step, if the sum of Gaussian function values related to the inspection data is less than the inspection threshold, the inspection object is It is determined that the state is unknown, and in each discriminator, the inspection data is checked against the determination criterion to determine whether the inspection signal is a normal signal or an abnormal signal, and a Gaussian relating to the inspection data is obtained. The sum of the function values is equal to or greater than the inspection threshold, and the total reliability of the discriminator determined to be the normal signal is that of the discriminator determined to be that the inspection signal is an abnormal signal. If the reliability is greater than or equal to the sum of reliability, it is determined that the inspection target is in a normal state, the sum of Gaussian function values related to the inspection data is greater than or equal to the inspection threshold, and the inspection signal is a normal signal. Size If the sum of the reliability of the discriminator is less than the sum of the reliability of the discriminator determined that the inspection signal is an abnormal signal, it is determined that the inspection target is in an abnormal state. .

  According to the present invention, when an inspection signal that is far from the learning signal is input, it is erroneously determined that the inspection object is in a normal state or an abnormal state by determining that the inspection object is in an unknown state. Therefore, inspection accuracy can be improved. In addition, by distinguishing between the abnormal state and the unknown state, it is possible to capture the change in the state of the inspection object by examining the inspection signal determined to be the unknown state. Furthermore, by distinguishing the unknown state from the abnormal state, it is possible to specify a disturbance factor such as background noise, and to provide a guideline for improving the operating environment.

  The invention of the signal identification method according to claim 5 includes: learning for creating an inspection reference for inspecting a state of an inspection object using a plurality of learning signals including a normal signal and an abnormal signal; and the inspection object A signal identification method for performing an inspection for determining the state of the inspection object using an inspection signal representing the state of the inspection and the inspection standard, and is preset from the plurality of learning signals at the time of learning A first step of extracting a feature amount for each extraction range; a second step of selecting a combination of extraction ranges used for the inspection as an inspection region using the feature amount of each extraction range of each learning signal; A third step of setting the inspection standard using the feature amount of each extraction range of the inspection region, and extracting a feature amount for each extraction range from the inspection signal at the time of inspection; From the feature amount, A fourth step of selecting a feature amount of each extraction range of the region, and inspecting the state of the inspection target using the selected feature amount and the inspection reference, and in the second step, each learning The feature amount of the learning signal is used as learning data, and a weight is set for each learning signal. Each time the weight is changed, the feature amount of each learning signal is input to the competitive learning type neural network in each extraction range. And learning the competitive learning type neural network, creating a clustering map in which the output layer neurons of the competitive learning type neural network are associated with categories, setting thresholds, and determining the feature quantities of each learning signal and the thresholds To determine whether each learning signal is a normal signal, calculate the sum of the weights of the erroneously determined learning signals as an erroneous determination rate, and determine the threshold value Create a classifier candidate applied to the quasi, and each time the weight is changed, the classifier candidate having the smallest error determination rate is selected as a classifier from the classifier candidates created for each extraction range, The reliability is calculated using the misjudgment rate for each classifier, the combination of extraction ranges of each classifier is set as the inspection region, and in the third step, each learning data is input to each classifier, The sum of the reliability is obtained for each of the learning data, an inspection threshold is set as the inspection standard from the distribution of the reliability total for all the learning data, and in the fourth step, the inspection area The feature quantity of each extraction range is used as inspection data, and in each discriminator, the inspection data is checked against the determination criterion to determine whether the inspection signal is a normal signal, and all the discriminators Of reliability The sum of the reliability of the discriminator determined that the sum is equal to or greater than the inspection threshold and the inspection signal is a normal signal is the reliability of the discriminator determined that the inspection signal is an abnormal signal. When the sum is equal to or greater than the sum, it is determined that the inspection target is in a normal state.

  According to the present invention, using the feature amount of each learning signal, it is determined whether each learning signal is a normal signal by collating each feature amount with the above determination criterion while changing the determination criterion. By calculating the misjudgment rate at this time and using combinations of extraction ranges that reduce the misjudgment rate, it is possible to automatically select a combination of extraction ranges to be an inspection region.

  Moreover, according to this invention, a test | inspection precision can be improved by determining the state of a test object using the reliability of each discriminator at the time of a test | inspection.

  According to a sixth aspect of the present invention, there is provided the signal identification device according to the fifth aspect, wherein in the second step, after learning by the competitive learning type neural network, each learning data is re-input to the competitive learning type neural network. Then, for each learning data, in each output layer neuron on the clustering map, a distance between the learning data and a weight vector is obtained, and each learning is performed for each output layer neuron on the clustering map. After setting the Gaussian function for each output layer neuron on the clustering map, setting the variance determined by the plurality of distances obtained from the data for use and the weight vector, and the Gaussian function defined by the average value that is the weight vector , Re-enter each learning data into the competitive learning neural network, For each learning data, obtain the sum of Gaussian function values of all output layer neurons on the clustering map, set the threshold from the distribution of the sum of Gaussian function values for all learning data, and set the discriminator Each time the selection is performed, the weight of the learning signal erroneously determined by the discriminator is increased, and in the fourth step, when the sum of the reliability is less than the inspection threshold, the inspection target is in an unknown state The sum of the reliability is equal to or greater than the inspection threshold value, and the sum of the reliability of the discriminator determined that the inspection signal is a normal signal is the abnormal signal. If the reliability of the classifier determined to be less than the total sum of reliability, the test object is determined to be in an abnormal state.

  According to the present invention, when an inspection signal that is far from the learning signal is input, it is erroneously determined that the inspection object is in a normal state or an abnormal state by determining that the inspection object is in an unknown state. Therefore, inspection accuracy can be improved. In addition, by distinguishing between the abnormal state and the unknown state, it is possible to capture the change in the state of the inspection object by examining the inspection signal determined to be the unknown state. Furthermore, by distinguishing the unknown state from the abnormal state, it is possible to specify a disturbance factor such as background noise, and to provide a guideline for improving the operating environment.

  The invention of the signal identification device according to claim 7 includes: learning for creating an inspection standard for inspecting a state of an inspection object using a plurality of learning signals including a normal signal and an abnormal signal; and the inspection object A signal identification device for performing an inspection for determining the state of the inspection object using the inspection signal representing the state of the inspection and the inspection reference, wherein the plurality of learning signals are input during learning, An inspection signal is input, an extraction unit that extracts a feature amount for each preset extraction range from each of the plurality of learning signals and the inspection signal, and each extraction range of each learning signal during learning A selection unit that selects a combination of extraction ranges used for the inspection as an inspection region using a feature amount, a setting unit that sets the inspection reference using a feature amount of each extraction range of the inspection region, and the inspection signal An inspection unit that selects a feature amount of each extraction range of the inspection region from a feature amount of each extraction range, and inspects the state of the inspection target using the selected feature amount and the inspection reference; Is a weight setting change unit that sets and changes the weight for each learning signal, and changes the determination criterion using the feature amount of each learning signal in each extraction range each time the weight is changed. However, the learning determination unit that determines whether each learning signal is a normal signal by collating each feature amount with the determination criterion, and the learning determination unit for each determination criterion. A calculation unit that calculates an erroneous determination rate using the weight of the erroneously determined learning signal, and a discriminator candidate creation unit that creates a discriminator candidate to which the determination criterion that minimizes the erroneous determination rate is applied for each extraction range And the weight is changed Each of the classifier candidates created for each extraction range is selected as a classifier candidate having the smallest error determination rate as a classifier, and the reliability is determined using the error determination rate for each classifier. A classifier selection unit that calculates and uses a combination of extraction ranges of each classifier as the inspection region, and a learning signal that is erroneously determined by the classifier every time the classifier is selected by the classifier selection unit A weight instructing unit that instructs the weight setting changing unit to increase the weight of the feature, and the setting unit includes the feature amount and the reliability for each extraction range of the inspection region in each learning signal. A learning data creation unit that calculates the product of the degree and uses it as learning data, and inputs each learning data into a competitive learning type neural network to learn the competitive learning type neural network, A map creation unit that creates a clustering map that associates output layer neurons of a custom-type neural network with a category; and an inspection standard setting unit that sets the inspection standard using the clustering map, and the inspection unit includes: Whether the inspection object is in a normal state by calculating the product of the feature quantity of each extraction range of the inspection area and the reliability to obtain inspection data, and comparing the inspection data with the inspection standard It is characterized by determining.

  According to the present invention, using the feature amount of each learning signal, it is determined whether each learning signal is a normal signal by collating each feature amount with the above determination criterion while changing the determination criterion. By calculating the misjudgment rate at that time and selecting a combination of extraction ranges that reduces the misjudgment rate as the inspection region, it is possible to automatically select a combination of extraction ranges to be the inspection region.

  Further, according to the present invention, the competitive learning type neural network is trained using the product of the feature amount of each learning signal and the reliability of each discriminator at the time of learning, and the feature amount of the inspection signal and each discrimination at the time of inspection Inspection accuracy can be improved by determining the state of the inspection object using the product of the reliability of the device.

  The invention of the signal identification device according to claim 8 is the learning for creating an inspection standard for inspecting the state of the inspection object using a plurality of learning signals including a normal signal and an abnormal signal, and the inspection object A signal identification device for performing an inspection for determining the state of the inspection object using the inspection signal representing the state of the inspection and the inspection reference, wherein the plurality of learning signals are input during learning, An inspection signal is input, an extraction unit that extracts a feature amount for each preset extraction range from each of the plurality of learning signals and the inspection signal, and each extraction range of each learning signal during learning A selection unit that selects a combination of extraction ranges used for the inspection as an inspection region using a feature amount, a setting unit that sets the inspection reference using a feature amount of each extraction range of the inspection region, and the inspection signal An inspection unit that selects a feature amount of each extraction range of the inspection region from a feature amount of each extraction range, and inspects the state of the inspection target using the selected feature amount and the inspection reference; Uses the feature quantity of each learning signal as learning data, sets a weight for each learning signal, and sets and changes the weight, and for each learning range in each extraction range each time the weight is changed A map creation unit that inputs signal feature amounts to a competitive learning type neural network, trains the competitive learning type neural network, and creates a clustering map that associates output layer neurons of the competitive learning type neural network with categories. A threshold setting unit that sets a threshold using the clustering map, a feature amount of each learning signal and the threshold for each extraction range To determine whether or not each learning signal is a normal signal, and to calculate the sum of the weights of the erroneously determined learning signals as an erroneous determination rate, and to determine the threshold value for each extraction range A discriminator candidate creating unit that creates a discriminator candidate applied to a reference, and a discriminator having the smallest misjudgment rate among the discriminator candidates created for each extraction range each time the weight is changed A candidate is selected as a classifier, a reliability is calculated using the misjudgment rate for each classifier, and a classifier selector that uses a combination of extraction ranges of each classifier as the inspection region; and the classifier selector A weight instruction unit for instructing the weight setting change unit to increase the weight of the learning signal erroneously determined by the classifier every time the classifier is selected in Each learning data is input to each classifier. Obtaining the sum of the reliability for each learning data, setting a detection threshold as the inspection standard from the distribution of the sum of the reliability for all the learning data, and the inspection means each of the inspection regions The feature quantity of the extraction range is used as inspection data, and in each discriminator, the inspection data is checked against the determination criteria to determine whether the inspection signal is a normal signal, and The sum of the reliability of the discriminator determined that the total sum of reliability is equal to or greater than the inspection threshold and the test signal is a normal signal is the discriminator determined that the test signal is an abnormal signal. When the reliability is equal to or higher than the total sum, it is determined that the inspection target is in a normal state.

  According to the present invention, using the feature amount of each learning signal, it is determined whether each learning signal is a normal signal by collating each feature amount with the above determination criterion while changing the determination criterion. By calculating the misjudgment rate at this time and using combinations of extraction ranges that reduce the misjudgment rate, it is possible to automatically select a combination of extraction ranges to be an inspection region.

  Moreover, according to this invention, a test | inspection precision can be improved by determining the state of a test object using the reliability of each discriminator at the time of a test | inspection.

  According to the present invention, it is possible to automatically select a combination of extraction ranges serving as inspection regions and improve inspection accuracy.

1 is a block diagram illustrating an overall configuration of a signal identification device according to Embodiment 1. FIG. It is a figure which shows the structure of the competitive learning type | mold neural network which concerns on the same as the above. It is a block diagram which shows the structure of the selection part of the signal identification device which concerns on the same as the above. It is a block diagram which shows the structure of the setting part of the signal identification device which concerns on the same as the above. It is a block diagram which shows the structure of the test | inspection part of the signal identification device which concerns on the same as the above. It is a flowchart which shows a learning method among the signal identification methods which concern on the same as the above. It is a flowchart which shows the feature-value production | generation method in the learning method of the signal identification method which concerns on the same as the above. It is a flowchart which shows the preparation method of the discriminator candidate in the learning method of the signal identification method which concerns on the same as the above. It is a flowchart which shows an inspection method among the signal identification methods which concern on the same as the above. It is a block diagram which shows the structure of the test | inspection part of the signal identification device which concerns on Embodiment 2. FIG. It is a flowchart which shows an inspection method among the signal identification methods which concern on the same as the above. It is a block diagram which shows the structure of the selection part of the signal identification apparatus which concerns on Embodiment 3. FIG. It is a flowchart which shows a learning method among the signal identification methods which concern on the same as the above. It is a flowchart which shows the preparation method of the discriminator candidate in the learning method of the signal identification method which concerns on the same as the above. It is a flowchart which shows an inspection method among the signal identification methods which concern on the same as the above.

(Embodiment 1)
In the signal identification method according to the first embodiment, learning that creates an inspection standard for inspecting the state of the inspection object A illustrated in FIG. 1 using a plurality of learning signals including a normal signal and an abnormal signal; An inspection for determining the state of the inspection object A is performed using the inspection signal indicating the state of the inspection object A and the inspection standard.

  An example of a signal identification device used in the signal identification method according to the present embodiment will be described. The signal identification device according to the present embodiment, during learning, uses a plurality of learning data based on feature amounts extracted from learning signals collected in advance as unsupervised competitive learning neural networks (hereinafter referred to as “neural networks”). ) 1 (see FIG. 2) to learn the neural network 1. The signal identification device creates a clustering map from the learned neural network 1. On the other hand, at the time of inspection, the signal identification device inspects the state of the inspection object A using the clustering map.

  As shown in FIG. 2, the neural network 1 is a self-organization map (SOM) composed of two layers of an input layer 11 and an output layer 12. The input layer 11 includes a plurality of input layer neurons N1, N1,. The output layer 12 includes a plurality of output layer neurons N2, N2,. Each input layer neuron N1 and each output layer neuron N2 are mutually coupled, and one element of input data (input vector) is input to each input layer neuron N1. The number of input layer neurons N1 and output layer neurons N2 need not be the same. Although the neural network 1 of the present embodiment is assumed to be realized by executing an appropriate application program on a sequential processing type computer, it can also be realized by a dedicated neurocomputer.

  As shown in FIG. 1, the signal identification device of the present embodiment includes a signal input unit 2 having a plurality of signal sources, and an extraction unit (extraction unit) 3 that extracts a plurality of feature amounts from signals from the signal input unit 2. A selection unit (selection unit) 4 that selects a feature amount necessary for inspection from a plurality of feature amounts, and a setting unit that sets an inspection threshold value δ used for inspection using the feature amount selected by the selection unit 4 (Setting means) 5, a storage section 6 for storing various information, an inspection section (inspection means) 7 for determining the state of the inspection object A, the determination result of the inspection section 7 as an external device (not shown), etc. And an output unit 8 for outputting to the output. The operation of the signal identification device includes a learning mode used when learning the neural network 1 (during learning) and an inspection mode used when checking the state of the inspection target A (during inspection). The inspection target A is, for example, an apparatus or equipment that includes a rotating device. However, the inspection target A is not limited to the above devices and facilities.

  The signal input unit 2 includes a vibration sensor 20 and a microphone 21 as signal sources. The vibration sensor 20 detects vibration of the inspection object A at the time of inspection, and converts the detected vibration into an analog electric signal (inspection signal). The electrical signal converted by the vibration sensor 20 is a one-dimensional waveform signal representing a change in vibration of the inspection object A with respect to time. The microphone 21 detects the sound of the inspection object A at the time of inspection, and converts the detected sound into an analog electric signal (inspection signal). The electric signal converted by the microphone 21 is a one-dimensional waveform signal representing the change in sound of the inspection object A with respect to time. The inspection signals from the vibration sensor 20 and the microphone 21 are output to the extraction unit 3, respectively.

  On the other hand, at the time of learning, for each state (category) of the inspection object A, the vibration sensor 20 detects the vibration of the inspection object A, and the microphone 21 detects the sound of the inspection object A. That is, the vibration sensor 20 and the microphone 21 detect vibration and sound for each operation condition of the inspection target A within a range in which the inspection target A operates in a normal state. The vibration sensor 20 converts the detected vibration into an analog electrical signal (learning signal). The microphone 21 converts the detected sound into an analog electric signal (learning signal).

  A plurality of learning signals are input to the extraction unit 3 during learning. Each learning signal is assigned in advance with an individual sample number i (i = 0,..., M−1) to be distinguished from other learning signals. Each learning signal is a one-dimensional waveform and includes a normal signal and an abnormal signal (for example, a signal other than a normal signal such as an abnormal signal). The normal signal is a signal acquired from the actual inspection object A and is input from the signal input unit 2. On the other hand, the abnormal signal is a signal collected in advance regardless of the inspection target A, and is input from the storage unit 6. As a result, it is not necessary to collect abnormal signals (abnormal signals) that are very rarely generated at actual manufacturing sites, so that the operation burden on the operator can be reduced without reducing the inspection performance.

  The extraction unit 3 performs preprocessing for removing noise on the learning signal for each input learning signal. Thereafter, the extraction unit 3 performs a short-time Fourier transform on each learning signal using Equations 1 and 2, and features F (i) for each extraction range (a combination of time (center time of window function) b and frequency f). ) (B, f) is extracted. A set of feature values F (i) (b, f) of time (time coordinate) b and frequency (frequency coordinate) f extracted by the extraction unit 3 is obtained from the learning signal of sample number i. Use as learning data. For example, the feature quantity of sample number i = 2, time b3 and frequency f2 is expressed as F (2) (b3, f2). Equation 1 represents the feature amount F (i) (b, f) of the sample number i, time b, and frequency f, and Equation 2 represents the window function Rd (t−b) of Equation 1. The window function Rd (t−b) is a Gaussian window centered on time b. x (t) is an electrical signal.

  On the other hand, at the time of inspection, an inspection signal is input to the extraction unit 3 from the vibration sensor 20 or the microphone 21. After the inspection signal is input, the extraction unit 3 performs a short-time Fourier transform on the inspection signal in the same manner as at the time of learning, and extracts a feature quantity F (b, f) at time b and frequency f. A set of feature values F (b, f) at time b and frequency f is used as inspection data. For example, the feature quantity at time b2 and frequency f5 is represented as F (b2, f5) (see Equations 1 and 2; i is omitted).

  The selection unit 4 uses the AdaBoost method to determine the combination of the extraction ranges used for the inspection of the inspection target A from the feature amounts F (i) (b, f) of the respective extraction ranges of the learning signals at the time of learning. Select as. As illustrated in FIG. 3, the selection unit 4 includes a weight management unit 40, a learning determination unit 41, a first calculation unit 42, a classifier candidate creation unit 43, and a classifier selection unit 44. Yes.

  The weight management unit 40 sets and changes the weight W (i) for each learning data (learning signal) during learning, and the weight setting changing unit 400 weights each learning data with respect to the weight setting changing unit 400. A weight instruction unit 401 for instructing to change W (i). W (i) represents the weight of the learning data of sample number i. For example, the weight of learning data with sample number i = 1 is represented as W (1), and the weight of learning data with sample number i = m−1 is represented as W (m−1).

  The weight setting changing unit 400 sets the weight W (i) for each learning signal so that the sum ΣW (i) of the weights W (i) of each learning data becomes 1 (ΣW (i) = 1). And change. The initial value of the weight W (i) of each learning data is set evenly. When there are m pieces of learning data, the weights W (i) of the learning data are all set to 1 / m.

  As will be described later, the weight instruction unit 401 sets the weight W (i) of the learning signal erroneously determined by the classifier DEC (n) every time the classifier DEC (n) is selected by the classifier selection unit 44. The weight setting changing unit 400 is instructed to increase and decrease the weight W (i) of the learning data correctly determined by the classifier DEC (n).

  Each time the weight W (i) is changed, the learning determination unit 41 changes the feature amount F (i of time b / frequency f from all the learning data in each extraction range (combination of time b / frequency f). ) (B, f) is extracted, and the extracted feature values F (i) (b, f) are rearranged in ascending order. The rearranged feature quantities F (c) are expressed as F (0), F (1),..., F (m−1) in order from the largest. Thereafter, the learning determination unit 41 compares the feature amounts F (c) with the threshold values while sequentially setting the feature amounts F (c) as threshold values. If the feature amount F (c) is less than the threshold, the learning determination unit 41 determines that the learning signal from which the feature amount F (c) is extracted is a normal signal. If the feature amount F (c) is greater than or equal to the threshold, the learning determination unit 41 determines that the learning signal from which the feature amount F (c) is extracted is an abnormal signal. As described above, the learning determination unit 41 uses the feature amount F (c) of each learning signal to collate the feature amount F (c) with the determination criterion while changing the determination criterion. It can be determined whether or not each learning signal is a normal signal. Note that the learning determination unit 41 may rearrange the feature values F (i) (b, f) in descending order instead of ascending order.

  When the weight setting change unit 400 sets or changes the weight W (i) of each learning data for all combinations of time b and frequency f, the first calculation unit 42 sets a threshold value (determination criterion). Each time, the sum ΣW (i) of the weights W (i) of the learning signals erroneously determined by the learning determination unit 41 is calculated as the error determination rate ε. The misjudgment is determined as an abnormal signal although the learning signal from which the feature value F (c) is extracted is originally a normal signal, or is normal even though the learning signal from the extraction source is originally an abnormal signal. It means to judge it as a signal. Thereafter, when the calculated erroneous determination rate ε is greater than 0.5, the first calculation unit 42 sets (1−error determination rate ε) to a new erroneous determination rate ε.

  The discriminator candidate creation unit 43 obtains a discriminator candidate (weak discriminator candidate) DEC (b, f) to which a determination criterion that minimizes the erroneous determination rate ε is applied for each extraction range (combination of time b and frequency f). create. The determination criterion is that when the erroneous determination rate ε calculated by the first calculation unit 42 is 0.5 or less and the feature amount F (c) is less than the threshold, the extraction source of the feature amount F (c) The learning signal is determined to be a normal signal, and if the feature value F (c) is equal to or greater than a threshold value, the extraction source learning signal is determined to be an abnormal signal. On the other hand, when the misjudgment rate ε calculated by the first calculation unit 42 is larger than 0.5 and (1-wrong decision rate ε) becomes a new misjudgment rate ε, the judgment criterion is the feature amount F (c ) Is less than the threshold value, it is determined that the learning signal from which the feature value F (c) is extracted is an abnormal signal. The content is that the signal is determined to be a normal signal.

  Each time the weight W (i) is set or changed, the discriminator selection unit 44 selects the discriminator candidate DEC (b, f) created for each extraction range (combination of time b and frequency f). The classifier candidate DEC (b, f) having the smallest error determination rate ε is selected as the classifier (weak classifier) DEC (n). n is a classifier number (weak classifier number) for identifying each classifier DEC (n). For example, DEC (3) represents the third selected classifier. The discriminator selection unit 44 calculates the reliability α (n) by substituting the erroneous determination rate ε into Equation 3 for each selected discriminator DEC (n), and substitutes the reliability α (n) into Equation 4. The coefficient β (n) (β (n) <1) is calculated. The discriminator selection unit 44 selects the discriminator DEC (n) each time the weight W (i) is set or changed, and selects a plurality (n in total) of discriminators DEC (n). The plurality of discriminators DEC (n) are stored in the storage unit 6 in the order of selection. A combination of extraction ranges of each discriminator DEC (n) is set as an inspection region.

  The setting unit 5 sets the inspection standard using the feature amount F (i) (b, f) of each extraction range of the inspection region. The setting unit 5 sets the inspection threshold δ using the feature amounts F (i) (b, f) of each extraction range of the inspection region selected by the selection unit 4.

  As shown in FIG. 4, the setting unit 5 includes a learning data creation unit 50, a map creation unit 51, a Gaussian function setting unit 52, and a threshold setting unit 53, and sets an inspection standard using a clustering map. To do.

  In each learning signal, the learning data creation unit 50 performs the feature amount F (i) (b) corresponding to the discriminator DEC (n) for each discriminator DEC (n) (for each extraction range of the inspection region). , F) and the reliability α (n) of the discriminator DEC (n) are calculated as learning data.

  At the time of learning, the map creating unit 51 sequentially inputs each learning data to the neural network 1 as an input vector. Each element of the learning data is associated with the input layer neuron N1 (see FIG. 2). The map creation unit 51 inputs a plurality of learning data to the neural network 1 to learn the neural network 1 and determines weighting coefficients for each input layer neuron N1 and each output layer neuron N2 (see FIG. 2). . Each output layer neuron N2 is associated with a weight vector whose element is a weighting coefficient between each input layer neuron N1. If learning data has not been input to the neural network 1 so far, the map creation unit 51 distributes the average value for each element of the plurality of learning data input this time by random numbers at the start of learning. A value given variation within the range is given to each output layer neuron N2 as an initial value of the weight vector. Thereafter, the weight vector of each output layer neuron N2 is determined by learning of the neural network 1 described above.

  Thereafter, the map creation unit 51 creates a two-dimensional clustering map in which each region is associated with the output layer neuron N2 in a 1: 1 ratio. That is, in the clustering map, each output layer neuron N2 of the output layer 12 of the neural network 1 is associated with either a normal category or an unknown category representing the state of the inspection target A. From the above, if a category of learning data is associated with each region of the clustering map, the category corresponding to the output layer neuron N2 fired by the inspection data can be known. The clustering map created by the map creation unit 51 is stored in the storage unit 6.

  After the map creation unit 51 learns the neural network 1 and creates a clustering map, the Gaussian function setting unit 52 re-inputs each learning data to the neural network 1 and sets each of the learning data on the clustering map. The Euclidean distance between the weight vector of the output layer neuron N2 and the learning data (input vector) is obtained. The Gaussian function setting unit 52 is defined by a variance determined by a plurality of Euclidean distances obtained from each learning data and a weight vector, and an average value that is a weight vector for each output layer neuron N2 on the clustering map. Set the Gaussian function.

Specifically, the Gaussian function setting unit 52 gives all the learning data of the same category to the neural network 1 to create a list of Euclidean distances (hereinafter referred to as “distance list”), and the Euclidean distance is calculated from the distance list. Select the maximum value. The Gaussian function setting unit 52 defines, for each output layer neuron N2, a Gaussian function y = exp (− | [x] − [m] | in which the selected maximum value is defined as the distribution σ and the weight vector is the average vector [m]. ² / 2σ ² ) is set. [X] is an input vector, and | [x] − [m] | is a Euclidean distance between the input vector [x] and a weight vector (average vector) [m]. y is an output value when the input vector [x] is input to the Gaussian function. Instead of setting the maximum value of the Euclidean distance as the variance σ, the Euclidean distance corresponding to a rank of several percent (for example, the top 5%) in the distance list may be set as the variance σ.

  According to the Gaussian function, the output value y increases as the input vector [x] is closer to the weight vector [m], and the output value y is maximum when the input vector [x] matches the weight vector [m]. Become 1 This output value y is used as the degree of attribution. That is, the range of the output value y is 0 to 1, and the greater the numerical value (the closer to 1), the higher the degree of attribution.

  As described above, the Gaussian function setting unit 52 can set a Gaussian function for each output layer neuron N2 corresponding to the category of learning data. No Gaussian function is set for the output layer neuron N2 outside the category. When a Gaussian function is set for each output layer neuron N2 of the neural network 1, the degree of attribution can be obtained by giving the Gaussian function the Euclidean distance between the feature quantity and the weight vector. Therefore, after setting the Gaussian function, the feature quantity (input vector) [x] obtained from the inspection object A is input to the neural network 1, and the weight vector [m] of each output layer neuron N2 and the feature quantity [x] The Euclidean distance is substituted into the Gaussian function of each output layer neuron N2, and the output value y of the Gaussian function is obtained for each category associated with the output layer neuron N2. Since the output value y is the degree of attribution of the feature amount to each category as described above, it can be determined whether or not the feature amount belongs to the category by setting a threshold for the degree of attribution.

  After setting a Gaussian function for each output layer neuron N2 on the clustering map, the threshold setting unit 53 re-inputs each learning data to the neural network 1, and for every learning data, all output layer neurons on the clustering map. The sum of the Gaussian function values of N2 is obtained, and the inspection threshold δ is set from the distribution of the sum of the Gaussian function values for all the learning data.

  The threshold value for inspection δ may be, for example, a degree of belonging corresponding to three times the variance, or may be a sum of output values y when the learning data of each category is given to the neural network 1 after the degree of belonging is set. It may be a value obtained by multiplying an appropriate coefficient.

  The setting unit 5 re-inputs all the learning data to the neural network 1, thereby setting the output layer neuron N2 fired by the learning data as the output layer neuron N2 belonging to the category of the normal inspection target A.

  A region corresponding to the output layer neuron N2 fired by each learning data is a region corresponding to a normal category.

  The storage unit 6 is, for example, a semiconductor memory. The storage unit 6 stores non-stationary signals among the learning signals, and stores a plurality of learning data created by the selection unit 4.

  The inspection unit 7 selects the feature amount F (b, f) of each extraction range of the inspection region from the feature amount F (b, f) of each extraction range of the inspection signal, and the feature amount of each extraction range of the inspection region. The product of F (b, f) and reliability α (n) is calculated and used as inspection data, and the inspection data and the inspection standard are compared to determine whether or not the inspection object A is in a normal state. .

  As shown in FIG. 5, the inspection unit 7 includes a second calculation unit 70 and an inspection determination unit 71.

  The second calculation unit 70 creates the product of the feature value F (b, f) and the reliability α (n) of each extraction range of the inspection area as inspection data. The second calculation unit 70 that has created the inspection data inputs the inspection data to each Gaussian function and obtains the sum of the Gaussian function values related to the inspection data.

  The inspection determination unit 71 determines whether or not the sum of the Gaussian function values is equal to or greater than the inspection threshold δ. When the total of the Gaussian function values is equal to or greater than the inspection threshold δ, the inspection determination unit 71 determines that the inspection object A is in a normal state. On the other hand, when the total sum of the Gaussian function values is less than the inspection threshold δ, the inspection determination unit 71 determines that the inspection object A is in an unknown state.

  The determination result made by the inspection unit 7 is output from the inspection unit 7 to the storage unit 6. The storage unit 6 stores all the determination results so far. Moreover, the said determination result is output to the output part 8 from the test | inspection part 7, and it gives an instruction | indication to an appropriate alarm means from the output part 8 as needed, and alert | reports an alarm.

  When the determination result is input from the inspection unit 7, the output unit 8 sounds an alarm sound or displays a warning screen according to the input determination result. Further, when an external device (not shown) is connected, the output unit 8 can output information corresponding to the determination result to the external device.

  Next, a signal identification method according to the present embodiment will be described. First, a learning method for setting determination conditions for inspection will be described with reference to FIG.

  First, the extraction unit 3 extracts feature amounts F (i) (b, f) for each preset extraction range from a plurality of learning signals using a preset extraction method. Thereafter, the selection unit 4 uses the feature amounts F (i) (b, f) of each extraction range of each learning signal to select a combination of extraction ranges used for the inspection of the inspection target A as an inspection region (FIG. 6 S1).

  Step S1 will be described with reference to FIG. First, m learning signals (learning waveform samples) are input to the extraction unit 3 (S31 in FIG. 7). The sample number i is set to the initial value 0 (S32). Thereafter, the extraction unit 3 performs a short-time Fourier transform on the learning signal to extract the feature amount F (i) (b, f) (S33). The extraction unit 3 determines whether the short-time Fourier transform of all (m) learning signals has been completed (S34). If the short-time Fourier transform of all the learning signals has not been completed, the sample number i is incremented by 1 (S35), and the next learning signal is subjected to the short-time Fourier transform to obtain the feature amount F (i) (b, f ). Thereby, the extraction part 3 can extract the feature-value F (i) (b, f) of each extraction range from the some signal for learning.

  When the short-time Fourier transform of all the learning signals is completed, the weight setting changing unit 400 of the selection unit 4 sets the weight W (i) of each learning signal to an initial value (1 / m) (S36). . Thereafter, the discriminator number n is set to the initial value 0 (S37), the time (time coordinate) b is set to the initial value 0 (S38), and the frequency (frequency coordinate) f is set to the initial value 0 (S39). ).

  Thereafter, the classifier candidate creation unit 43 creates a classifier candidate DEC (b, f) at time b and frequency f (S40). At this time, an erroneous determination rate ε of the classifier candidate DEC (b, f) is calculated. Thereafter, the discriminator selection unit 44 determines whether or not the error determination rate ε of the current discriminator candidate DEC (b, f) is the smallest among the previous error determination rates ε (S41). When the current misjudgment rate ε is the minimum, the discriminator selection unit 44 calculates the reliability α (n) when the discriminator candidate DEC (b, f) is the discriminator DEC (n) ( S42). Thereafter, the discriminator selection unit 44 updates the discriminator DEC (n) with the discriminator number n to the current discriminator candidate DEC (b, f) (S43). The discriminator selection unit 44 determines whether or not the processing from step S40 to step S43 has been performed for all frequencies f at time b (S44). If the processing from step S40 to step S43 is not completed for all frequencies f at time b, the frequency f is increased by 1 (S45), and the processing from step S40 to step S43 is repeated for all frequencies f. Further, it is determined whether or not the processing from step S39 to step S44 has been completed for all times b (S46). If the processing from step S39 to step S44 has not been completed for all times b, time b Is incremented by 1 (S47), and the processing from step S39 to step S44 is repeated until the processing from step S39 to step S44 is completed for all times b. Thereafter, the distribution of the weight W (i) of the learning data is updated (S48). Thereafter, it is determined whether or not the processing from step S38 to step S48 has been completed for all the classifiers DEC (n) (S49). If the processing from step S38 to step S48 has not been completed, the classifier number is determined. n is incremented by 1 (S50), and the processing from step S38 to step S48 is repeated until the processing from step S38 to step S48 is completed for all the discriminators DEC (n). Thereby, every time the weight W (i) of each learning signal is set or changed, the selection unit 4 has the smallest error determination rate ε among the discriminator candidates DEC (b, f) created for each extraction range. Can be selected as the classifier DEC (n).

  Further, the reliability α (n) is calculated using the erroneous determination rate ε for each classifier DEC (n), and the combination of the extraction ranges corresponding to each classifier DEC (n) is set as the inspection region, and the classifier DEC ( Each time n) is selected, the weight W (i) of learning data erroneously determined by the discriminator DEC (n) can be increased.

  Here, the creation of the classifier candidate DEC (b, f) in step S40 will be described with reference to FIG. First, the feature amounts F (i) (b, f) of the learning data are rearranged in ascending order (or descending order) for each extraction range (S61 in FIG. 8). The threshold number c is set to an initial value 0 (S62). A discriminator candidate DEC (b, f) is created with normal values less than the threshold value and abnormal values above the threshold value (S63). Thereafter, an erroneous determination rate ε is calculated (S64). When the erroneous determination rate ε is larger than 0.5 (S65), the determination criterion is reversed, and (1− erroneous determination rate ε) is set as a new erroneous determination rate ε (S66). When the erroneous determination rate ε is 0.5 or less, the erroneous determination rate ε is left as it is. Thereafter, it is determined whether or not the erroneous determination rate ε is minimum (S67). If the erroneous determination rate ε is minimum, the classifier candidate DEC (b, f) is updated (S68). It is determined whether or not the processing from step S63 to step S68 has been completed for all threshold numbers c (S69). If the processing from step S63 to step S68 has not been completed, the threshold number c is incremented by 1 ( S70) The process from step S63 to step S68 is repeated until the process from step S63 to step S68 is completed for all threshold numbers c. When the processing is completed for all the threshold numbers c, the final candidate is determined as a classifier candidate DEC (b, f) (S71).

  Next, after step S1 in FIG. 6 is performed will be described. The setting unit 5 sets the sample number i to the initial value 0 (S2 in FIG. 6), and selects the feature value F (i) (b, f) of the short-time Fourier transform of the learning signal (S3). Further, the setting unit 5 sets the discriminator number n to the initial value 0 (S4), and selects the feature quantity F (i) (b, f) corresponding to the discriminator DEC (n) (S5). Thereafter, feature element calculation is performed to calculate a product of the feature amount F (i) (b, f) and the reliability α (n) (S6). It is determined whether or not the processing of step S5 and step S6 has been completed for all the classifiers DEC (n) (S7), and the processing of steps S5 and S6 must be completed for all the classifiers DEC (n). For example, the discriminator number n is incremented by 1 (S8), and the processing of step S5 and step S6 is repeated until the processing of step S5 and step S6 is completed for all the discriminators DEC (n). Subsequently, it is determined whether or not the processing from step S3 to step S7 has been completed for all the learning signals (S9). If the processing from step S3 to step S7 has not been completed for all the learning signals. The sample number i is incremented by 1 (S10), and the processing from step S3 to step S7 is repeated until the processing from step S3 to step S7 is completed for all the learning signals. Thereby, in each learning signal, the product of the feature amount F (i) (b, f) and the reliability α (n) is calculated for each extraction range of the inspection region to obtain learning data. it can.

  Thereafter, the map creation unit 51 inputs each learning data to the neural network 1 to learn the neural network 1 and creates a clustering map (S11). Thereafter, the Gaussian function setting unit 52 re-inputs each learning data to the neural network 1 (S12), and calculates the Euclidean distance between the learned weight vector and the corresponding learning data at each selected map position ( S13), a distance list is created (S14), and the maximum Euclidean distance is selected (S15). Thereafter, the Gaussian function setting unit 52 sets a Gaussian function for each position of the clustering map (S16).

  Thereafter, the threshold setting unit 53 re-inputs each learning data to the neural network 1 and inputs all the learning data to the Gaussian function group (S17). The sum of Gaussian function values is calculated for each learning data (S18), and a Gaussian function total value list is created (S19). Thereafter, the threshold value setting unit 53 sets, as the inspection standard, 50% of the average value of the lower 10% of the total value as the inspection threshold value δ from the distribution of the sum of the Gaussian function values for all the learning data ( S20).

  Next, an inspection method for inspecting the inspection object A using the determination conditions will be described with reference to FIG. First, an inspection signal is input to the extraction unit 3 (S81 in FIG. 9). The extraction unit 3 performs a short-time Fourier transform on the inspection signal, and extracts the feature amount F (b, f) of each extraction range from the inspection signal (S82). The inspection unit 7 sets the discriminator number n to the initial value 0 (S83). From the extracted feature values F (b, f) of each extraction range, the feature values F (i) (bn, fn) of the extraction range corresponding to the discriminator DEC (n) are selected (S84). Thereafter, the product of the feature value F (bn, fn) of the extraction range corresponding to the discriminator DEC (n) and the reliability α (n) is calculated (S85). It is determined whether or not the processing has been completed for all the discriminators DEC (n) (S86). If the processing has not been completed for all the discriminators DEC (n), the discriminator number n is incremented by 1 (S87). ), The process of step S84 and step S85 is performed for the next discriminator DEC (n). Thereby, the feature amount F (bn, fn) of each extraction range of the inspection region can be selected from the feature amount F (bn, fn) of each extraction range extracted in step S82.

  Thereafter, the second calculation unit 70 of the inspection unit 7 inputs the inspection data to the Gaussian function group, and calculates the sum of the Gaussian function values related to the inspection data (S88). The inspection determination unit 71 determines whether or not the sum of the Gaussian function values is equal to or greater than the inspection threshold δ (S89). When the sum of the Gaussian function values is equal to or greater than the inspection threshold δ, the inspection determination unit 71 determines that the inspection object A is in a normal state (S90). On the other hand, when the sum of the Gaussian function values is less than the inspection threshold δ, the inspection determination unit 71 determines that the inspection object A is in an abnormal state (S91).

  As described above, according to the present embodiment, the feature amount F (i) (b, f) of each learning signal is used to change the determination criterion and change the feature amount F (i) (b, f). By comparing the reference and calculating whether or not each learning signal is a normal signal, the misjudgment rate ε is calculated, and a combination of extraction ranges that makes the misjudgment rate ε small is selected as an inspection region. Thus, it is possible to automatically select a combination of extraction ranges to be an inspection area.

  Further, according to the present embodiment, a neural network is obtained by using the product of the feature amount F (i) (b, f) of each learning signal and the reliability α (n) of each discriminator DEC (n) during learning. 1 is learned, and the state of the inspection object A is determined by using the product of the characteristic amount F (b, f) of the inspection signal and the reliability α (n) of each discriminator DEC (n) at the time of inspection. Inspection accuracy can be improved.

  Furthermore, according to the present embodiment, when an inspection signal that is an abnormal signal or an inspection signal that is far from a learning signal is input, it is determined that the inspection object A is in an abnormal state, and the inspection object A is normal. Since it is possible to prevent erroneous determination as being in a state, inspection accuracy can be improved.

(Embodiment 2)
The signal identification method according to the second embodiment is the signal according to the first embodiment in that the state of the inspection object A is inspected using the sum of the Gaussian function values regarding the inspection data and using the AdaBoost method at the time of inspection. It is different from the identification method. In addition, about the component similar to Embodiment 1, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  As shown in FIG. 10, the inspection unit 7 of the present embodiment includes a second calculation unit 70 and an inspection determination unit 71 in the same manner as the inspection unit 7 of Embodiment 1, and the third calculation. A part 72 is further provided.

  At the time of inspection, the third calculation unit 72 uses the AdaBoost method to extract a plurality of discriminators DEC (n) from the storage unit 6 in descending order of reliability α (n), and the discriminator previously extracted In order from DEC (n), the feature amount F (b, f) extracted from the inspection signal by the extraction unit 3 is compared with a threshold value to determine whether or not the inspection target A is in a normal state. When the discriminator DEC (n) determines that the inspection target A is in the normal state, the third calculation unit 72 adds the reliability α (n) of the current discriminator DEC (n) to the reliability sum S so far. ) Is added as a new reliability sum S. When the discriminator DEC (n) determines that the inspection target A is in an abnormal state, the third calculator 72 calculates the reliability α (n) of the current discriminator DEC (n) from the total reliability S thus far. The value obtained by subtracting () is the new reliability sum S.

  After the determination similar to that in the first embodiment is performed in all the classifiers DEC (n), the inspection determination unit 71 according to the present embodiment determines that the inspection target A is in a normal state when the reliability sum S is 0 or more. It is finally determined that When the reliability sum S is less than 0, the inspection determination unit 71 finally determines that the inspection object A is in an abnormal state. The final determination result of the inspection determination unit 71 is output to the output unit 8.

  Next, a signal identification method according to the present embodiment will be described. Since the learning method is the same as the signal identification method according to the first embodiment, only the inspection method will be described with reference to FIG.

  First, an inspection signal is input to the extraction unit 3 (S101 in FIG. 11). The extraction unit 3 performs a short-time Fourier transform on the inspection signal (S102). The inspection unit 7 sets the reliability sum S to the initial value 0 (S103), and sets the discriminator number n to the initial value 0 (S104). The feature quantity F (bn, fn) of the extraction range corresponding to the discriminator DEC (n) is selected (S105).

  Thereafter, it is determined whether or not the inspection signal is a normal signal by using the discriminator DEC (n) (S106). That is, in each discriminator DEC (n), the inspection data is checked against the determination criterion to determine whether the inspection signal is a normal signal or an abnormal signal. When it is determined that the inspection signal is a normal signal (S107), the reliability α (n) of the discriminator DEC (n) is added to the reliability sum S so far to obtain a new reliability sum S. (S108). On the other hand, when the inspection signal is determined to be an abnormal signal (S107), the reliability α (n) of the discriminator DEC (n) is subtracted from the reliability sum S so far to obtain a new reliability sum S. (S109).

  Thereafter, the product of the feature value F (bn, fn) and the reliability α (n) of each extraction range of the inspection area is calculated and used as inspection data (S110). It is determined whether or not the processing from step S105 to step S110 has been completed for all the classifiers DEC (n) (S111), and the processing from step S105 to step S110 is completed for all the classifiers DEC (n). If not, the discriminator number n is incremented by 1 (S112), and the processing from step S105 to step S110 is performed for the next discriminator DEC (n).

  Thereafter, the second calculation unit 70 of the inspection unit 7 inputs the inspection data to the Gaussian function group, and calculates the sum of the Gaussian function values related to the inspection data (S113). The inspection determination unit 71 determines whether or not the sum of the Gaussian function values is equal to or greater than the inspection threshold δ (S114). When the sum of the Gaussian function values is less than the inspection threshold δ, the inspection determination unit 71 determines that the inspection object A is in an unknown state (S115). On the other hand, when the sum of the Gaussian function values is equal to or greater than the inspection threshold δ, the inspection determination unit 71 determines whether or not the reliability sum S is equal to or greater than 0 (S116). That is, the sum of the reliability α (n) of the discriminator DEC (n) that has been determined that the inspection signal is a normal signal is the reliability of the discriminator DEC (n) that has been determined that the inspection signal is an abnormal signal. It is determined whether or not the sum of α (n) is equal to or greater. When the reliability sum S is 0 or more, it is determined that the inspection target A is in a normal state (S117). On the other hand, when the reliability sum S is less than 0, the inspection determination unit 71 determines that the inspection object A is in an abnormal state (S118).

  As described above, according to the present embodiment, when a test signal that is far from the learning signal is input, the test target A is in a normal state or an abnormal state by determining that the test target A is in an unknown state. Can be prevented from being erroneously determined, so that the inspection accuracy can be improved. Further, by distinguishing between the abnormal state and the unknown state, it is possible to capture the state change of the inspection target A by considering the inspection signal determined to be the unknown state. Furthermore, by distinguishing the unknown state from the abnormal state, it is possible to specify a disturbance factor such as background noise, and to provide a guideline for improving the operating environment.

(Embodiment 3)
In the third embodiment, learning to create an inspection standard for inspecting the state of the inspection object A using a plurality of learning signals including a normal signal and an abnormal signal in a method different from that in the first embodiment; A signal identification method for performing an inspection for determining the state of the inspection object A using the inspection signal indicating the state of the inspection object A and the inspection standard will be described. In addition, about the component similar to Embodiment 1, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  The signal identification method according to the present embodiment is different from the signal identification method according to the first and second embodiments in that the neural network 1 is used to create the classifier candidate DEC (b, f).

  The selection unit 4 of the present embodiment uses the AdaBoost method to select a combination of extraction ranges to be an inspection region by a method different from that of the first embodiment. As shown in FIG. 12, the selection unit 4 includes a weight management unit 40, a first calculation unit 42, a discriminator candidate creation unit 43, and a discriminator selection unit 44 according to the selection unit 4 (FIG. 3) and a map creation unit 45, a Gaussian function setting unit 46, and a threshold setting unit 47. In the present embodiment, the feature amount F (i) (b, f) extracted from each learning signal by the extraction unit 3 is used as learning data, and the feature amount F (b, f) in each extraction range of the inspection region. Is used as inspection data.

  Each time the weight W (i) of each learning data is set or changed, the map creating unit 45 converts the feature amount F (i) (b, f) of each learning signal into a neural network in each extraction range. 1 is input to learn the neural network 1 to create a clustering map in which the output layer neuron N2 of the neural network 1 is associated with a category.

  The Gaussian function setting unit 46 re-inputs the learning data to the neural network 1 after learning of the neural network 1, and for each learning data, the learning data and the weight vector in each output layer neuron N2 on the clustering map. The Euclidean distance is determined with respect to each output layer neuron N2 on the clustering map, and is defined by a variance determined by a plurality of Euclidean distances obtained from each learning data and a weight vector, and an average value that is a weight vector Set the Gaussian function.

  After setting a Gaussian function for each output layer neuron N2 on the clustering map, the threshold setting unit 47 re-inputs each learning data to the neural network 1, and for each learning data, Gauss of all neurons on the clustering map. The sum of function values is obtained, and a threshold value is set from the distribution of the sum of Gauss function values for all learning data.

  The first calculation unit 42 according to the present embodiment uses the threshold value to check the feature amount F (i) (b, f) of each learning signal against the determination criterion while changing the determination criterion. It is determined whether or not the learning signal is a normal signal, and the sum of the weights W (i) of the erroneously determined learning signals is calculated as an erroneous determination rate ε.

  The classifier candidate creation unit 43 according to the present embodiment creates a classifier candidate DEC (b, f) in which a threshold is applied to a determination criterion.

  The discriminator selection unit 44 of the present embodiment misidentifies from the discriminator candidates DEC (b, f) created for each extraction range every time the weight W (i) of each learning signal is set or changed. The classifier candidate DEC (b, f) having the smallest constant rate ε is selected as the classifier DEC (n), and the reliability α (n) is calculated using the erroneous determination rate ε for each classifier DEC (n). A combination of extraction ranges of each discriminator DEC (n) is set as an inspection region.

  The setting unit 5 of this embodiment inputs each learning data to each discriminator DEC (n), obtains the sum of the reliability α (n) for each learning data, and the reliability α for all the learning data. The sum total of (n) is obtained to create a summation list, and the detection threshold δ can be set as an inspection standard from the distribution of the sum of the reliability α (n).

  The inspection unit 7 of the present embodiment uses the feature amount F (b, f) of each extraction range of the inspection region as inspection data, and performs inspection using the AdaBoost method. That is, the inspection unit 7 determines whether or not the inspection signal is a normal signal by collating the inspection data with the determination criterion in each discriminator DEC (n). The inspection unit 7 determines that the inspection target A is unknown when the sum of the reliability α (n) of all the classifiers DEC (n) (hereinafter referred to as “second reliability total”) T is less than the inspection threshold δ. It is determined that it is in a state. The sum of the reliability α (n) of the discriminator DEC (n) determined that the second reliability total T is equal to or greater than the inspection threshold δ and the inspection signal is a normal signal, and the inspection signal is abnormal. If it is equal to or greater than the sum of the reliability α (n) of the discriminator DEC (n) determined to be a signal, the inspection unit 7 determines that the inspection target A is in a normal state. The sum of the reliability α (n) of the discriminator DEC (n) determined that the reliability sum S is equal to or greater than the inspection threshold δ and the inspection target A is in the normal state, and the inspection target A is in the abnormal state. If it is less than the sum of the reliability α (n) of the discriminator DEC (n) determined as follows, the inspection unit 7 determines that the inspection target A is in an abnormal state.

  Next, a signal identification method according to the present embodiment will be described. First, the learning method will be described with reference to FIG.

  First, m learning signals (learning waveform samples) are input to the extraction unit 3 (S131 in FIG. 13). The sample number i is set to the initial value 0 (S132). Thereafter, the extraction unit 3 performs a short-time Fourier transform on the learning signal to extract the feature amount F (i) (b, f) (S133). The extraction unit 3 determines whether the short-time Fourier transform of all (m) learning signals has been completed (S134). If the short-time Fourier transform of all the learning signals has not been completed, the sample number i is incremented by 1 (S135), and the next learning signal is short-time Fourier transformed. Thereby, the extraction part 3 can extract the feature-value F (i) (b, f) for every extraction range from the some signal for learning. The feature amount F (i) (b, f) of each learning signal is used as learning data.

  When the short-time Fourier transform of all the learning signals is completed, the weight setting changing unit 400 of the selection unit 4 sets the weight W (i) of each learning signal to an initial value (1 / m) (S136). . Thereafter, the discriminator number n is set to the initial value 0 (S137), the time b is set to the initial value 0 (S138), and the frequency f is set to the initial value 0 (S139).

  Thereafter, the classifier candidate creation unit 43 creates a classifier candidate DEC (b, f) at time b and frequency f (S140). At this time, an erroneous determination rate ε of the classifier candidate DEC (b, f) is calculated. Thereby, a threshold value is set, and the feature amount F (i) (b, f) of each learning signal is compared with the threshold value to determine whether or not each learning signal is a normal signal. The sum ΣW (i) of the weights W (i) of the learning signals thus calculated is calculated as the error determination rate ε, and the classifier candidate DEC (b, f) in which the threshold is applied to the determination criterion can be created.

  Thereafter, the classifier selection unit 44 determines whether or not the erroneous determination rate ε of the current classifier candidate DEC (b, f) is the smallest among the previous erroneous determination rates ε (S141). When the current misjudgment rate ε is the minimum, the discriminator selection unit 44 calculates the reliability α (n) when the discriminator candidate DEC (b, f) is the discriminator DEC (n) ( S142). Thereafter, the classifier selector 44 updates the classifier DEC (n) with the classifier number n to the current classifier candidate DEC (b, f) (S143). The discriminator selection unit 44 determines whether or not the processing from step S140 to step S143 has been performed for all frequencies f at time b (S144). If the processing from step S140 to step S143 has not been completed for all frequencies f at time b, the frequency f is incremented by 1 (S145), and the steps from step S140 to step S143 are completed for all frequencies f. The processing from S140 to step S143 is repeated. Further, it is determined whether or not the processing from step S139 to step S144 has been completed for all times b (S146). If the processing from step S139 to step S144 has not been completed, time b is increased by 1 ( S147) The processing from step S139 to step S144 is repeated until the processing from step S139 to step S144 is completed for all times b. Thereafter, the distribution of the weight W (i) of the learning data is updated (S148). Each time the discriminator DEC (n) is selected, the weight W (i) of the learning signal erroneously determined by the discriminator DEC (n) can be increased. Thereafter, it is determined whether or not the process from step S138 to step S148 has been completed for all the classifiers DEC (n) (S149). If the process from step S138 to step S148 has not been completed, the classifier number is determined. n is incremented by 1 (S150), and the processing from step S138 to step S148 is repeated until the processing from step S138 to step S148 is completed for all the discriminators DEC (n).

  Here, the creation of the classifier candidate DEC (b, f) in step S140 will be described with reference to FIG. First, the map creation unit 45 inputs the feature values F (i) (b, f) of time b and frequency f in the normal signal among the plurality of learning signals to the neural network 1 to learn the neural network 1. Thus, a clustering map is created (S161 in FIG. 14).

  After learning the neural network 1, the Gaussian function setting unit 46 re-inputs the learning data of each normal signal to the neural network 1 (S162), and for each learning data, in each output layer neuron N2 on the clustering map, The Euclidean distance between the learning data and the weight vector is obtained (S163), a distance list is created (S164), and the maximum value of the Euclidean distance is selected (S165). Thereafter, the Gaussian function setting unit 46 sets a Gaussian function for each position of the clustering map (S166).

  After setting the Gaussian function for each output layer neuron N2 on the clustering map, the threshold setting unit 47 re-inputs each learning data to the neural network 1 and inputs all the learning data to the Gaussian function group (S167). . For each learning data, the sum of Gaussian function values of all output layer neurons N2 on the clustering map is calculated (S168), and a Gaussian function sum list is created (S169). Thereafter, the threshold setting unit 47 sets a value of 50% of the average value of the lower 10% of the total value from the distribution of the sum of the Gaussian function values for all the learning data (S170). Thereafter, an erroneous determination rate ε is calculated (S171). When the misjudgment rate ε is larger than 0.5 (S172), the judgment criterion is reversed to set (1-misjudgment rate ε) as a new misjudgment rate ε (S173). When the erroneous determination rate ε is 0.5 or less, the erroneous determination rate ε is left as it is. Thus, the discriminator candidate DEC (b, f) at time b and frequency f is determined (S174).

  The process after step S149 of FIG. 13 will be described. The threshold setting unit 47 inputs each learning data to each discriminator DEC (n) (S151 in FIG. 13). A total sum of reliability α (n) is obtained for each learning data (S152), and a reliability total list is created (S153). After that, the threshold setting unit 47 uses, as the inspection standard, 50% of the average value of the lower 10% of the total value as the inspection threshold δ from the distribution of the sum of the reliability α (n) regarding all the learning data. Setting is made (S154).

  Next, the inspection method will be described with reference to FIG. First, an inspection signal is input to the extraction unit 3 (S181 in FIG. 15). The extraction unit 3 performs a short-time Fourier transform on the inspection signal, and extracts a feature amount F (b, f) for each extraction range from the inspection signal (S182). The inspection unit 7 sets the reliability sum S to the initial value 0 (S183), sets the second reliability sum T to the initial value 0 (S184), and sets the discriminator number n to the initial value 0 (S184). S185). From the extracted feature value F (b, f) of each extraction range, the feature value F (bn, fn) of the extraction range corresponding to the discriminator DEC (n) is selected (S186). The feature amount F (bn, fn) of each extraction range of the inspection area is used as inspection data.

  Thereafter, it is determined using the discriminator DEC (n) whether the inspection signal is a normal signal (S187). That is, in each discriminator DEC (n), the inspection data is checked against the determination criterion to determine whether the inspection signal is a normal signal or an abnormal signal. When it is determined that the inspection signal is a normal signal (S188), a value obtained by adding the reliability α (n) of the discriminator DEC (n) to the previous reliability total S is set as a new reliability total S. (S189). On the other hand, when the inspection signal is determined to be an abnormal signal (S188), a value obtained by subtracting the reliability α (n) of the discriminator DEC (n) from the previous reliability total S is set as a new reliability total S. (S190).

  Thereafter, a value obtained by adding the reliability α (n) of the discriminator DEC (n) to the previous second reliability total T is set as a new second reliability total T (S191). It is determined whether or not the processing from step S186 to step S191 has been completed for all the classifiers DEC (n) (S192), and the processing from step S186 to step S191 has been completed for all the classifiers DEC (n). If not, the discriminator number n is incremented by 1 (S193), and the processing from step S186 to step S191 for the next discriminator DEC (n) is performed.

  Thereafter, the inspection unit 7 determines whether or not the second reliability sum T is equal to or greater than the inspection threshold δ (S194). If the second reliability sum T is less than the inspection threshold δ, the inspection unit 7 determines that the inspection object A is in an unknown state (S195). When the second reliability sum T is equal to or greater than the inspection threshold δ, the inspection determination unit 71 determines whether or not the reliability sum S is equal to or greater than 0 (S196). In other words, the inspection determining unit 71 determines that the sum of the reliability α (n) of the classifier DEC (n) determined that the inspection signal is a normal signal is the classifier determined that the inspection signal is an abnormal signal. It is determined whether or not the total of the reliability α (n) of DEC (n) is greater than or equal to. When the reliability sum S is 0 or more, it is determined that the inspection target A is in a normal state (S197). On the other hand, when the reliability sum S is less than 0, the inspection determination unit 71 determines that the inspection object A is in an abnormal state (S198).

  As described above, according to the present embodiment, the feature amount F (i) (b, f) of each learning signal is used to change the determination criterion and change the feature amount F (i) (b, f). By comparing the reference and calculating whether or not each learning signal is a normal signal, an erroneous determination rate ε is calculated, and by using a combination of extraction ranges such that the erroneous determination rate ε becomes small, an inspection region is obtained. A combination of extraction ranges can be automatically selected.

  Further, according to the present embodiment, it is possible to improve the inspection accuracy by determining the state of the inspection object A using the reliability α (n) of each discriminator DEC (n) at the time of inspection.

  Furthermore, according to the present embodiment, when an inspection signal that is far from the learning signal is input, the inspection object A is in a normal state or an abnormal state by determining that the inspection object A is in an unknown state. Can be prevented from being erroneously determined, so that the inspection accuracy can be improved. Further, by distinguishing between the abnormal state and the unknown state, it is possible to capture the state change of the inspection target A by considering the inspection signal determined to be the unknown state. Furthermore, by distinguishing the unknown state from the abnormal state, it is possible to specify a disturbance factor such as background noise, and to provide a guideline for improving the operating environment.

  As a modification of the first to third embodiments, the signal identification device used in the signal identification method is the same as the first to third embodiments even if the electrical signal is not subjected to the short-time Fourier transform, but the continuous wavelet transform is performed. There is an effect.

  The extraction unit 3 of the modified example performs the preprocessing for removing noise on the learning signal for each learning signal including the normal signal and the abnormal signal, and then uses the equations 5 and 6 for the learning signal. Are subjected to continuous wavelet transform, and feature values Y (i) (b, a) in a combination of time b and parameter a are extracted. Equation 5 represents the feature quantity Y (i) (b, a) of the sample number i, time b, and parameter a, and Equation 6 represents the mother wavelet Ψ (t) of Equation 5. x (t) is an electrical signal.

  On the other hand, at the time of inspection, as in the case of learning, the extraction unit 3 performs continuous wavelet transform on the inspection signal, and extracts inspection data that is a set of feature quantities Y (b, a) of time b and parameter a. To do.

1 Neural network 3 Extraction unit (extraction means)
4 selection part (selection means)
5 Setting part (setting means)
7 Inspection part (inspection means)
A Inspection object DEC (b, f) Discriminator candidate DEC (n) Discriminator δ Inspection threshold

Claims (8)

Learning to create an inspection standard for inspecting the state of the inspection object using a plurality of learning signals including a normal signal and an abnormal signal, an inspection signal representing the state of the inspection object, and the inspection reference And a signal identification method for performing an inspection for determining the state of the inspection object,
A first step of extracting a feature amount for each preset extraction range from the plurality of learning signals during learning;
A second step of selecting a combination of extraction ranges to be used for the inspection as an inspection region using a feature amount of each extraction range of each learning signal;
A third step of setting the inspection standard using a feature amount of each extraction range of the inspection region;
At the time of inspection, a feature amount is extracted for each extraction range from the inspection signal, a feature amount of each extraction range of the inspection region is selected from the extracted feature amount of each extraction range, and the selected feature amount and the inspection standard And a fourth step of inspecting the state of the inspection object using
In the second step,
Each time a weight is set for each learning signal, and each time the weight is changed, the feature amount of each learning signal is used in each extraction range, and each feature amount is collated with the determination criterion while changing the determination criterion. To determine whether or not each learning signal is a normal signal, calculate an erroneous determination rate using the weight of the erroneously determined learning signal for each of the determination criteria, and for each extraction range , Creating a classifier candidate applying a criterion for minimizing the erroneous determination rate,
Each time the weight is changed, the classifier candidate having the smallest error determination rate is selected as a classifier from the classifier candidates created for each extraction range, and the error determination rate is used for each classifier. The reliability is calculated, and the combination of the extraction ranges of each classifier is set as the inspection area,
In the third step,
In each learning signal, for each extraction range of the inspection region, a product of the feature amount and the reliability is calculated and used as learning data.
Each learning data is input to a competitive learning type neural network to learn the competitive learning type neural network, and a clustering map is created by associating the output layer neurons of the competitive learning type neural network with categories. Set the inspection criteria using a clustering map,
In the fourth step,
The product of the feature value of each extraction range of the inspection area in the inspection signal and the reliability is calculated as inspection data, and the inspection object is in a normal state by comparing the inspection data with the inspection standard It is judged whether it is. The signal identification method characterized by the above-mentioned. In the second step,
Each time the weight is changed, the feature amount of each learning signal is rearranged in ascending order or descending order in each extraction range, the rearranged feature amounts are sequentially set as threshold values, and the size of each feature amount and the threshold value is changed. By comparing the relations and changing each of the determination criteria, each feature amount is checked against the determination criteria to determine whether each learning signal is a normal signal. , Calculating the sum of the weights of the erroneously determined learning signals as the erroneous determination rate,
Each time the classifier is selected, the weight of the learning signal erroneously determined by the classifier is increased,
In the third step,
After learning by the competitive learning type neural network, each learning data is re-input to the competitive learning type neural network, and the learning data and weights are output in each output layer neuron on the clustering map for each learning data. The distance between the vectors, the variance determined by a plurality of distances obtained from each learning data and the weight vector for each output layer neuron on the clustering map, and the average value that is the weight vector Set the Gaussian function defined by
After setting a Gaussian function for each output layer neuron on the clustering map, each learning data is re-input to the competitive learning type neural network, and for each learning data, all output layer neurons on the clustering map are input. Find the sum of Gaussian function values, set the inspection threshold as the inspection criterion from the distribution of the sum of the Gaussian function values for all learning data,
In the fourth step,
The test data is input to the Gaussian function to obtain the sum of Gaussian function values related to the test data,
The signal identification method according to claim 1, wherein if the total sum of Gaussian function values related to the inspection data is less than the inspection threshold, the inspection object is determined not to be in a normal state. In the fourth step,
If the sum of Gaussian function values for the inspection data is less than the inspection threshold, determine that the inspection target is in an abnormal state;
The signal identification method according to claim 2, wherein when the total sum of Gaussian function values related to the inspection data is equal to or greater than the inspection threshold, the inspection object is determined to be in a normal state. In the fourth step,
If the sum of the Gaussian function values for the inspection data is less than the inspection threshold, determine that the inspection object is in an unknown state;
In each discriminator, the inspection data is checked against the determination criteria to determine whether the inspection signal is a normal signal or an abnormal signal,
The sum of the Gaussian function values related to the inspection data is equal to or greater than the inspection threshold, and the total reliability of the discriminators determined that the inspection signal is a normal signal, the inspection signal is an abnormal signal. If the reliability of the determined classifier is greater than or equal to the sum of reliability, it is determined that the inspection target is in a normal state,
The sum of the Gaussian function values related to the inspection data is equal to or greater than the inspection threshold, and the total reliability of the discriminators determined that the inspection signal is a normal signal, the inspection signal is an abnormal signal. The signal identification method according to claim 2, wherein the inspection target is determined to be in an abnormal state when the reliability of the determined classifier is less than a total sum of reliability. Learning to create an inspection standard for inspecting the state of the inspection object using a plurality of learning signals including a normal signal and an abnormal signal, an inspection signal representing the state of the inspection object, and the inspection reference And a signal identification method for performing an inspection for determining the state of the inspection object,
A first step of extracting a feature amount for each preset extraction range from the plurality of learning signals during learning;
A second step of selecting a combination of extraction ranges to be used for the inspection as an inspection region using a feature amount of each extraction range of each learning signal;
A third step of setting the inspection standard using a feature amount of each extraction range of the inspection region;
At the time of inspection, a feature amount is extracted for each extraction range from the inspection signal, a feature amount of each extraction range of the inspection region is selected from the extracted feature amount of each extraction range, and the selected feature amount and the inspection standard And a fourth step of inspecting the state of the inspection object using
In the second step,
The feature amount of each learning signal is used as learning data, a weight is set for each learning signal, and each time the weight is changed, the feature amount of each learning signal is changed to a competitive learning type neural network in each extraction range. To learn the competitive learning type neural network, create a clustering map in which the output layer neurons of the competitive learning type neural network are associated with categories, set a threshold value, The threshold is compared to determine whether each learning signal is a normal signal, the sum of the weights of the erroneously determined learning signals is calculated as an erroneous determination rate, and the threshold is applied to the determination criterion Create a vessel candidate,
Each time the weight is changed, the classifier candidate having the smallest error determination rate is selected as a classifier from the classifier candidates created for each extraction range, and the error determination rate is used for each classifier. The reliability is calculated, and the combination of the extraction ranges of each classifier is set as the inspection area,
In the third step,
Each learning data is input to each discriminator, the sum of the reliability is obtained for each of the learning data, and an inspection threshold is set as the inspection standard from the distribution of the sum of the reliability regarding all the learning data. ,
In the fourth step,
The feature amount of each extraction range of the inspection region is used as inspection data, and in each discriminator, the inspection data is checked against the determination criterion to determine whether the inspection signal is a normal signal,
The sum of reliability of all discriminators is equal to or greater than the inspection threshold, and the sum of reliability of the discriminators determined that the inspection signal is a normal signal is that the inspection signal is an abnormal signal. A signal identification method characterized by determining that the inspection target is in a normal state when the reliability of the determined classifier is equal to or greater than a sum of reliability. In the second step,
After learning of the competitive learning type neural network, each learning data is re-input to the competitive learning type neural network, and the learning data and weights are output in each output layer neuron on the clustering map for each learning data. The distance between the vectors, the variance determined by a plurality of distances obtained from each learning data and the weight vector for each output layer neuron on the clustering map, and the average value that is the weight vector Set the Gaussian function defined by
After setting a Gaussian function for each output layer neuron on the clustering map, each learning data is re-input to the competitive learning type neural network, and for each learning data, all output layer neurons on the clustering map are input. Find the sum of Gaussian function values, set the threshold from the distribution of the sum of Gaussian function values for all learning data,
Each time the classifier is selected, the weight of the learning signal erroneously determined by the classifier is increased,
In the fourth step,
If the total reliability is less than the inspection threshold, determine that the inspection target is in an unknown state;
The discriminator in which the sum of the reliability is equal to or greater than the inspection threshold and the sum of the reliability of the discriminator determined that the inspection signal is a normal signal determines that the inspection signal is an abnormal signal 6. The signal identification method according to claim 5, wherein the inspection object is determined to be in an abnormal state when the reliability is less than a total sum of reliability. Learning to create an inspection standard for inspecting the state of the inspection object using a plurality of learning signals including a normal signal and an abnormal signal, an inspection signal representing the state of the inspection object, and the inspection reference A signal identification device that performs an inspection to determine the state of the inspection object using
The plurality of learning signals are input at the time of learning, the inspection signals are input at the time of inspection, and feature amounts are extracted from the plurality of learning signals and the inspection signal for each preset extraction range. Extraction means;
A selection means for selecting a combination of extraction ranges used for the inspection as an inspection region using a feature amount of each extraction range of each learning signal during learning;
Setting means for setting the inspection standard using the feature amount of each extraction range of the inspection region;
An inspection unit that selects a feature amount of each extraction range of the inspection region from a feature amount of each extraction range of the inspection signal, and inspects a state of the inspection target using the selected feature amount and the inspection reference; Prepared,
The selection means includes
A weight setting changing unit for setting and changing the weight for each learning signal;
Each time the weight is changed, in each extraction range, each learning signal is collated with the determination criterion while changing the determination criterion using the characteristic amount of each learning signal. And a learning determination unit for determining whether or not each is a normal signal,
For each determination criterion, a calculation unit that calculates an erroneous determination rate using the weight of the learning signal erroneously determined by the learning determination unit;
For each extraction range, a classifier candidate creating unit that creates a classifier candidate to which a determination criterion that minimizes the erroneous determination rate is applied;
Each time the weight is changed, the classifier candidate having the smallest error determination rate is selected as a classifier from the classifier candidates created for each extraction range, and the error determination rate is determined for each classifier. A classifier selector that calculates the reliability using the combination of extraction ranges of each classifier as the inspection region;
A weight instruction unit that instructs the weight setting change unit to increase the weight of a learning signal erroneously determined by the classifier each time the classifier is selected by the classifier selection unit;
The setting means includes
In each learning signal, for each extraction range of the inspection region, a learning data creation unit that calculates a product of the feature amount and the reliability and uses it as learning data;
A map creation unit that inputs each learning data to a competitive learning type neural network, trains the competitive learning type neural network, and creates a clustering map in which the output layer neurons of the competitive learning type neural network are associated with categories. And
Set the inspection standard using the clustering map,
The inspection means includes
Whether or not the inspection object is in a normal state by calculating the product of the feature value of each extraction range of the inspection area and the reliability to obtain inspection data, and comparing the inspection data with the inspection standard A signal identification device characterized by determining. Learning to create an inspection standard for inspecting the state of the inspection object using a plurality of learning signals including a normal signal and an abnormal signal, an inspection signal representing the state of the inspection object, and the inspection reference A signal identification device that performs an inspection to determine the state of the inspection object using
The plurality of learning signals are input at the time of learning, the inspection signals are input at the time of inspection, and feature amounts are extracted from the plurality of learning signals and the inspection signal for each preset extraction range. Extraction means;
A selection means for selecting a combination of extraction ranges used for the inspection as an inspection region using a feature amount of each extraction range of each learning signal during learning;
Setting means for setting the inspection standard using the feature amount of each extraction range of the inspection region;
An inspection unit that selects a feature amount of each extraction range of the inspection region from a feature amount of each extraction range of the inspection signal, and inspects a state of the inspection target using the selected feature amount and the inspection reference; Prepared,
The selection means includes
A feature setting of each learning signal is used as learning data, and a weight setting changing unit that sets and changes a weight for each learning signal;
Each time the weight is changed, in each extraction range, the feature quantity of each learning signal is input to the competitive learning type neural network to learn the competitive learning type neural network, and the output of the competitive learning type neural network is output. A map creation unit for creating a clustering map in which layer neurons are associated with categories;
A threshold setting unit that sets a threshold using the clustering map;
For each extraction range, the feature amount of each learning signal is compared with the threshold value to determine whether each learning signal is a normal signal, and the sum of the weights of the erroneously determined learning signals is determined as an erroneous determination rate. A calculation unit for calculating as
A discriminator candidate creating unit for creating a discriminator candidate applying the threshold value to the determination criterion for each extraction range;
Each time the weight is changed, the classifier candidate having the smallest error determination rate is selected as a classifier from the classifier candidates created for each extraction range, and the error determination rate is determined for each classifier. A classifier selector that calculates the reliability using the combination of extraction ranges of each classifier as the inspection region;
A weight instruction unit that instructs the weight setting change unit to increase the weight of a learning signal erroneously determined by the classifier each time the classifier is selected by the classifier selection unit;
The setting means includes
Each learning data is input to each discriminator, the sum of the reliability is obtained for each of the learning data, and a detection threshold is set as the inspection standard from the distribution of the reliability sums for all the learning data. ,
The inspection means includes
The feature amount of each extraction range of the inspection region is used as inspection data, and in each discriminator, the inspection data is checked against the determination criterion to determine whether the inspection signal is a normal signal,
The sum of reliability of all discriminators is equal to or greater than the inspection threshold, and the sum of reliability of the discriminators determined that the inspection signal is a normal signal is that the inspection signal is an abnormal signal. The signal discriminating apparatus, wherein if the sum of the reliability of the discriminator determined is equal to or higher than the sum, the signal to be inspected is determined to be in a normal state.

Images