JP7013965B2 - Information processing equipment, information processing methods, and programs - Google Patents

Description

The present invention relates to an information processing apparatus, an information processing method, and a program.

In recent years, with the development of technology, various sensor devices for detecting the state of an object have been developed. In addition, many methods have been proposed for classifying the state of the observation target using the sensor data acquired by the sensor device as described above. For example, a classification method by performing non-negative matrix factorization on sensor data is known.

For example, Patent Document 1 below shows whether the observation target of the sensor data is normal or abnormal, using a coefficient matrix obtained by applying non-negative matrix factorization to the sensor data as a feature quantity. It is described to classify into two classes. Further, in Patent Document 2 below, the original data is approximately expressed with a smaller order by providing a constraint condition so that a plurality of similar basis vectors cannot be obtained by decomposition in non-negative matrix factorization. Is described.

Japanese Unexamined Patent Publication No. 2017-151872 Japanese Unexamined Patent Publication No. 2015-135574

The smaller the number of bases obtained by non-negative matrix factorization, the more the calculation cost or communication cost can be suppressed. If set, there is a risk that the classification performance will deteriorate.

Therefore, the present invention has been made in view of the above problems, and an object of the present invention is a new and improved information processing apparatus capable of achieving both improvement of classification performance and cost reduction. , Information processing methods, and programs.

In order to solve the above problem, according to a certain viewpoint of the present invention, a preprocessing unit that acquires non-negative teacher data based on time-series data and a non-negative matrix factor decomposition are applied to the teacher data. By doing so, the basis vector is acquired , and based on the basis vector, the information indicating the classification performance corresponding to the basis vector is acquired as the basis vector performance information, and based on the basis vector performance information, the candidate of the teacher basis matrix is obtained. The basis vector included in the teacher basis matrix candidate is specified, and based on the teacher basis matrix candidate, information indicating the classification performance corresponding to the teacher basis matrix candidate is acquired as the basis matrix performance information, and the teacher basis matrix candidate is obtained. Based on the number of basis vectors included in, information indicating the calculation cost corresponding to the teacher basis matrix candidate is acquired as cost information, and the basis matrix performance information and the cost information corresponding to the teacher basis matrix candidate are obtained. A teacher data processing unit that calculates a combined evaluation value, identifies a teacher basis matrix from the teacher basis matrix candidates based on the evaluation value, and generates a teacher coefficient matrix based on the teacher basis matrix and the teacher data. An information processing apparatus comprising the above is provided .

The basis matrix performance information may be information based on the classification accuracy when the teacher data is classified using a plurality of coefficient vectors for the plurality of the basis vectors included in the teacher basis matrix candidate as feature quantities.

All the data included in the teacher data are associated with labels of the same type, and the basis vector performance information is included in each of a plurality of coefficient vectors for the plurality of the basis vectors included in the teacher basis matrix candidate. The information may be based on the sum of Mahalanobis distances calculated using the coefficients as a sample.

The basis vector performance information may be information based on the classification accuracy when the teacher data is classified using the coefficient vector for the basis vector as a feature amount.

All the data included in the teacher data are associated with labels of the same type, and the basis vector performance information is the sum of the Mahalanobis distances calculated using each coefficient included in the coefficient vector for the basis vector as a sample. It may be based information.

The information processing apparatus may further include a model building unit that generates classification model parameters based on the teacher coefficient matrix and the correct label associated with the teacher data.

The preprocessing unit further acquires non-negative evaluation data based on further time-series data, and the information processing apparatus generates an evaluation coefficient matrix from the evaluation data based on the teacher base matrix. And a classification unit that classifies the evaluation coefficient matrix based on the classification model parameters.

The feature extraction unit generates an evaluation coefficient matrix candidate from the evaluation data, and when the residual to the evaluation data calculated based on the teacher base matrix and the evaluation coefficient matrix candidate satisfies a predetermined condition, the feature extraction unit generates the evaluation coefficient matrix candidate. The evaluation coefficient matrix candidate may be used as the evaluation coefficient matrix.

The teacher data processing unit adds a basis vector obtained by applying a non-negative matrix factor decomposition to the residual when the residual does not satisfy the predetermined condition to the teacher base matrix. The teacher coefficient matrix may be further generated based on the above, and the model building unit may update the classification model parameters based on the further generated teacher coefficient matrix.

Further, in order to solve the above problems, according to another viewpoint of the present invention, acquisition of non-negative teacher data based on time-series data and application of non-negative matrix factor decomposition to the teacher data. By doing so, the base vector is acquired , and based on the base vector, information indicating the classification performance corresponding to the base vector is acquired as the base vector performance information, and the candidate of the teacher base matrix is obtained based on the base vector performance information. The base vector included in the teacher base matrix candidate is specified, and based on the teacher base matrix candidate, information indicating the classification performance corresponding to the teacher base matrix candidate is acquired as the base matrix performance information, and the teacher base matrix is obtained. Based on the number of base vectors included in the candidate, information indicating the calculation cost corresponding to the teacher base matrix candidate is acquired as cost information, and the base matrix performance information and the cost information corresponding to the teacher base matrix candidate are obtained . The evaluation value obtained by combining the above is calculated, the teacher base matrix is specified from the teacher base matrix candidates based on the evaluation value, and the teacher coefficient matrix is generated based on the teacher base matrix and the teacher data . Information processing methods performed by the information processing apparatus are provided, including.

Further, in order to solve the above-mentioned problems, according to another viewpoint of the present invention, a function of acquiring non-negative teacher data based on time-series data and a non-negative matrix factor for the teacher data are obtained. The basis vector is acquired by applying the decomposition, and the information indicating the classification performance corresponding to the basis vector is acquired as the basis vector performance information based on the basis vector, and the teacher basis is obtained based on the basis vector performance information. The basis vector included in the teacher basis matrix candidate that is a candidate for the matrix is specified, and based on the teacher basis matrix candidate, information indicating the classification performance corresponding to the teacher basis matrix candidate is acquired as the basis matrix performance information. Based on the number of basis vectors included in the teacher basis matrix candidate, information indicating the calculation cost corresponding to the teacher basis matrix candidate is acquired as cost information, and the basis matrix performance information corresponding to the teacher basis matrix candidate and the said A function that calculates an evaluation value in combination with cost information, identifies a teacher basis matrix from the teacher basis matrix candidates based on the evaluation value, and generates a teacher coefficient matrix based on the teacher basis matrix and the teacher data. And, a program to realize is provided.

As described above, according to the present invention, it is possible to achieve both improvement in classification performance and cost reduction.

It is a figure which shows an example of the system configuration which concerns on one Embodiment of this invention. It is a block diagram which shows the functional structure example of the sensor terminal 2 which concerns on the same embodiment. It is explanatory drawing which shows typically the processing example of NMF. It is a flowchart which shows the operation example of the sensor terminal 2 which concerns on this Embodiment in a learning stage. It is a flowchart which shows the flow of the process of step S110. It is a flowchart which shows the operation example of this Embodiment in a classification stage. It is a flowchart which shows the flow of the process of step S210. It is a figure which shows the hardware configuration example.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the present specification and the drawings, components having substantially the same functional configuration are designated by the same reference numerals, so that duplicate description will be omitted.

Further, in the present specification and the drawings, a plurality of components having substantially the same functional configuration may be distinguished by adding different alphabets after the same reference numerals. However, if it is not necessary to particularly distinguish each of the plurality of components having substantially the same functional configuration, only the same reference numerals are given.

<< 1. Background >>
Before explaining one embodiment of the present invention, first, the background leading to the creation of the present embodiment will be described. In recent years, various sensor terminals for observing the state of an observation target have been developed. In addition, we propose many methods to classify (including discrimination) the state of the observation target based on the time-series sensor data (hereinafter sometimes referred to as time-series data) acquired by the sensor terminal as described above. Has been done.

For example, in factories that manufacture goods, various equipment, devices, products, etc. are used, but the above-mentioned observation targets often cannot directly acquire control signals related to the operation of the equipment. , It is also widely used to detect the state by an external sensor terminal.

In this case, for example, by acquiring sensor data such as vibration data and acoustic data from a plurality of sensor terminals arranged around the observation target, it is possible to dynamically classify the state of the observation target. In addition, by analyzing and evaluating the obtained sensor information, it is possible to acquire movement loci and operating conditions, detect abnormal signs, etc., and improve production and work efficiency and ensure safety. can. In the present specification, an example in which a normal state (hereinafter, also simply referred to as normal) or an abnormal state (hereinafter, also simply referred to as abnormal) exists as an observation target state and is classified into two classes, normal or abnormal, will be described. However, the present invention is not limited to such an example.

As a classification method using time-series data such as sensor data, a classification method by performing non-negative matrix factorization is known. Non-negative matrix factorization
Matrix Factorization (hereinafter also referred to as NMF) is an algorithm that decomposes a non-negative matrix Y into a product (for example, WH) of a coefficient matrix W and a base matrix H, which are two non-negative matrices.

In the classification method using NMF, teacher data, which is non-negative time-series data associated with correct labels such as normal and abnormal, is learned in advance, and the time-series data to be classified using the classification parameters obtained by learning. Classify the evaluation data that is. At the time of learning, it is possible to obtain classification parameters for classification by learning the teacher coefficient matrix obtained by applying NMF to the teacher data. At the time of classification, a coefficient matrix is generated from the evaluation data using the teacher base matrix obtained by applying NMF to the teacher data, and the generated coefficient matrix is used as a feature quantity for classification using the above-mentioned classification parameters. It is possible to do.

By the way, when data processing such as learning and classification is assumed on the terminal side where the calculation processing capacity is high, it is desirable to suppress the calculation cost for learning and classification. In the classification method using NMF described above, the calculation cost increases according to the order of NMF (that is, the number of basis vectors included in the teacher basis matrix).

The order of NMF is a parameter that can be appropriately set in advance, for example, and generally, the classification performance is improved by setting a large order. However, the larger the order of NMF, the higher the calculation cost. Furthermore, the calculation cost for extracting the combination with the optimum classification performance from the preselected base vector and coefficient matrix increases exponentially as the order of NMF increases, so the total search time is sufficiently short. It may not end with.

Further, in an observation environment where a relatively high sampling frequency or data transfer such as vibration data or acoustic data is required, it is more desirable to suppress communication costs. For example, in the above-mentioned classification method using NMF, when a coefficient matrix used as a feature quantity is transmitted from a plurality of terminals to a server and classification processing is performed on the server side, the order of the above-mentioned NMF is increased. Communication costs will also increase.

Generally, a sensor terminal that transfers data wirelessly is driven by a battery, so that redundant wireless data transfer increases power consumption and tends to cause premature battery exhaustion. When the battery runs out, it becomes impossible to perform sensing, so it is desirable to suppress the communication cost as described above. Further, even in a system that can supply external power by wired or wireless power supply, it is not desirable to consume more power than necessary, so it is also desirable to suppress communication costs. Further, since the cost of holding the data in the storage of the sensor terminal or the server increases by transmitting and receiving more data than necessary, it is also desirable to suppress the communication cost.

As described above, the classification performance is improved by setting a large order of NMF, but the calculation cost and the communication cost (hereinafter, may be simply referred to as a cost) also increase according to the order of NMF. Therefore, in the above-mentioned classification method using NMF, the classification performance and the communication cost may have a trade-off relationship.

By the way, as a premise in this specification, it is desirable to detect the state in real time. That is, it is desirable to output the detection result corresponding to the change of the state as soon as possible after the change of the state occurs. Since the observable range of the sensor terminal is generally limited physically and spatially, the sensor terminal is arranged in a plurality of places, and the types of sensors are also diverse. In addition, in order to detect (classify) the state of equipment and environment using the sensor data obtained by the sensor, it is necessary to collect sufficient sensor data.

In general, the more prominent features corresponding to each detected state can be extracted, the higher the state discrimination accuracy becomes. For example, it is possible for a person with specialized knowledge to specify an appropriate observation location, that is, the location of the sensor for observation, but the internal structure of the equipment is complicated and it is not possible to make an installation decision unless actual data is acquired. In some cases, or a person without specialized knowledge may place the sensor terminal. In such a case, it is not always appropriate to set the order of NMF and select the combination of features, and it is desirable to automatically perform them based on an objective index.

One embodiment of the present invention was conceived by paying attention to the above points, and efficiently reduces costs such as calculation cost and communication cost while maintaining high classification performance related to state classification of observation objects. It is possible to suppress it. Hereinafter, an embodiment of the present invention that exhibits such an effect will be described in detail.

<< 2. Configuration example >>
<2-1. System configuration>
First, the configuration of the information processing system according to the embodiment of the present invention will be described. As described above, the information processing system according to the present embodiment has both maintenance of classification performance and cost control in the state classification of the observation target based on the sensor data (an example of time series data) acquired by the sensor. Is possible.

FIG. 1 is a diagram showing an example of a system configuration according to an embodiment of the present invention. Referring to FIG. 1, the information processing system 9 according to the present embodiment may include an observation target 1, a plurality of sensor terminals 2, and an aggregation device 3. Further, the sensor terminal 2 and the aggregation device 3 are connected via the network 5.

The observation target 1 is the target of the state classification in the present embodiment. In the present embodiment, the observation target 1 may be, for example, various devices and products in a factory, electronic devices installed in a company or a home, and the like. Further, the observation target 1 may include buildings such as buildings, bridges, and roads. Further, the observation object 1 according to the present embodiment may include one or more internal devices for which the sensor data is acquired by the sensor terminal 2.

The sensor terminal 2 is an information processing device that collects various sensor data from the observation target 1. Since the observable range of the sensor terminal is generally limited physically and spatially, even if a plurality of sensor terminals 2 according to the present embodiment are arranged for one observation target 1 as shown in FIG. good. In the example shown in FIG. 1, three sensor terminals 2A to 2C are arranged for one observation target 1.

Further, the sensor terminal 2 according to the present embodiment collects various sensor data related to the observation target 1, and classifies the operating state of the observation target 1 based on the collected sensor data. In the present embodiment, the classification result of the operating state of the observation target 1 by the sensor terminal 2 may be, for example, normal or abnormal. Further, the sensor terminal 2 transmits the classification result of the operating state of the observation target 1 to the aggregation device 3. A more detailed configuration of the sensor terminal 2 will be described later with reference to FIG.

The aggregation device 3 is a server (information processing device) that aggregates the classification results of the operating states of the observation target 1 and outputs the final classification results. As shown in FIG. 1, the aggregation device 3 according to the present embodiment includes a communication unit 32, an aggregation unit 34, and an output unit 36.

The communication unit 32 according to the present embodiment receives the classification result of the operating state of the observation target 1 from the plurality of sensor terminals 2. The aggregation unit 34 according to the present embodiment performs aggregation processing based on the classification results received from the plurality of sensor terminals 2 by the communication unit 32, and specifies the final classification result. The output unit 36 outputs the final classification result obtained by the aggregation process of the aggregation unit 34. The output unit 36 may be, for example, a display device, and in such a case, the output unit 36 displays and outputs the final classification result. The output unit 36 may output each of the classification results received from each sensor terminal 2 in addition to or instead of the final classification result.

The system configuration according to this embodiment has been described above. The system configuration described with reference to FIG. 1 is merely an example, and the system configuration according to the present embodiment is not limited to such an example. For example, in FIG. 1, a case where three sensor terminals 2A to 2C are arranged for one observation target 1 is shown as an example, but the number of sensor terminals 2 is not limited to this example. Further, there may be a plurality of sets of the observation target 1 and the sensor terminal 2 according to the present embodiment. The system configuration according to the present embodiment can be flexibly modified according to the characteristics of the observation target, the specifications of the network 5, and the like.

<2-2. Sensor terminal configuration>
Subsequently, a functional configuration example of the sensor terminal 2 according to the present embodiment will be described. FIG. 2 is a block diagram showing a functional configuration example of the sensor terminal 2 according to the present embodiment. Referring to FIG. 2, the sensor terminal 2 according to the present embodiment has a sensor 202, an AD (analog digital) conversion unit 204, a preprocessing unit 206, a teacher data processing unit 208, a classification model construction unit 212, a storage unit 214, and features. It includes an extraction unit 216, a classification unit 218, and a communication unit 220.

Of the parts shown in FIG. 2, some of the functional blocks are in the learning stage where learning is performed based on the sensor data and the classification stage where classification is performed based on the sensor data and the state of the observation target 1 is evaluated. Functions can be different. Therefore, in the following, if necessary, the functions in the learning stage and the functions in the classification stage may be described separately.

The sensor 202 has a function of collecting sensor data (an example of time-series data) related to the observation target 1 by sensing. Examples of the sensor 202 include a vibration sensor, an acoustic sensor, a heat sensor, an illuminance sensor, an image pickup sensor, and the like. The above is only an example, and the sensor terminal 2 according to the present embodiment may be provided with various sensors 202 according to the characteristics of the observation target 1.

Further, although FIG. 2 shows an example in which the sensor terminal 2 includes one sensor 202, the sensor terminal 2 may include a plurality of types of sensors 202. By providing the sensor terminal 2 with a plurality of types of sensors 202, it is possible to capture different physical phenomena according to the operating state of the observation target 1.

The AD conversion unit 204 converts analog sensor data into digital sensor data. When the sensor data output from the sensor 202 is digital sensor data, the sensor terminal 2 does not have to include the AD conversion unit 204.

The preprocessing unit 206 performs various preprocessing such as filtering processing for removing noise, measurement value conversion processing, and frequency conversion processing on the sensor data. In the frequency conversion process by the preprocessing unit 206, at least a part of the sensor data is converted into a frequency domain representation by performing a conversion such as a Fourier transform, a fast Fourier transform, a short-time Fourier transform, and a wavelet transform. It may be a process to be performed.

By performing the above-mentioned preprocessing on the sensor data, the preprocessing unit 206 acquires a non-negative observation matrix including a non-negative observation vector (data converted into a frequency domain representation). The observation matrix acquired by the preprocessing unit 206 is a matrix in which one or a plurality of observation vectors are arranged. Further, the observation vector may be a power spectrum obtained by the Fourier transform, a spectrogram obtained by the short-time Fourier transform, a scalogram obtained by the wavelet transform, or the like. Further, the preprocessing unit 206 may acquire an observation matrix for each predetermined period based on the sensor data for the predetermined period, or acquire the observation matrix based on the sensor data for the period specified by an arbitrary method. You may.

The preprocessing unit 206 may have a partial function different between the learning stage and the classification stage. In the following, the non-negative observation matrix acquired by the preprocessing unit 206 in the learning stage is called the teacher data or the teacher observation matrix, and the non-negative observation matrix acquired by the preprocessing unit 206 in the classification stage is called the evaluation data or the evaluation observation matrix. May be called.

The teacher data is a non-negative observation matrix acquired by the preprocessing unit 206 in the learning stage where the sensor terminal 2 learns, and in the present embodiment, each observation vector included in the teacher data is either normal or abnormal. The correct answer label is associated. The correct label is a label indicating the class to which each observation vector included in the teacher observation matrix (teacher data) belongs. To the above, the class to which each observation vector belongs in this embodiment is either normal or abnormal.

The method of associating the correct label with each observation vector included in the teacher data is not particularly limited, but for example, the user inputs class information indicating the class via an operation unit (not shown), and the preprocessing unit 206 is used as the class information. Each correct answer label may be associated with the observation vector based on the above. The preprocessing unit 206 outputs the observation matrix (teacher observation matrix) acquired as teacher data to the teacher data processing unit 208.

The evaluation data is a non-negative observation matrix acquired by the preprocessing unit 206 in the classification stage where the sensor terminal 2 classifies. The correct answer label does not have to be associated with the evaluation data. The preprocessing unit 206 outputs the observation matrix (evaluation observation matrix) acquired as evaluation data to the feature extraction unit 216.

The function of the teacher data processing unit 208 may differ between the learning stage and the classification stage. Therefore, in the following, first, the function of the teacher data processing unit 208 in the learning stage will be described, and then the function of the teacher data processing unit 208 in the classification stage will be described.

The teacher data processing unit 208 applies non-negative matrix factorization (NMF) to the teacher observation matrix (teacher data) output from the preprocessing unit 206 in the learning stage, and applies the teacher basis matrix and the teacher coefficient matrix. To generate. Hereinafter, NMF will be described in detail with reference to FIG.

As described above, NMF is an algorithm that decomposes the non-negative matrix Y of 1 into the product of two non-negative matrices W and H (for example, WH). Since it is difficult to analytically obtain the two non-negative matrices W and H, iteratively obtains an approximate solution so that the error between Y and WH becomes the local optimum solution by giving an initial value. .. In general, the local optimum solution changes depending on the initial value.

FIG. 3 is an explanatory diagram schematically showing a processing example of NMF. As shown in FIG. 3, for example, the non-negative observation matrix Y is decomposed by the NMF into the product of the non-negative coefficient matrix W and the non-negative basis matrix H.

As shown in FIG. 3, in the observation matrix Y according to the present embodiment, for example, the m-dimensional observation vector y becomes a row vector, and n rows m composed of only a few rows of the target section in the time series data. It is a matrix of columns.

In the basis matrix H according to the present embodiment, for example, NMF is applied to the observation matrix Y, and the decomposed m-dimensional basis vector h becomes a row vector, and k rows composed of rows of only a few k of the basis vectors. It is a matrix of m columns.

The coefficient matrix W according to the present embodiment is a matrix of n rows and k columns, for example, in which a k-dimensional coefficient vector w becomes a row vector and is composed of rows of only the number n of observation vectors. Here, the coefficient vector w is a row vector indicating a weighted value indicating how many components of each basis vector included in the basis matrix H are included in a certain observation vector.

As shown in FIG. 3, the above matrix has the relationship of the following mathematical formula (1).

By the way, as described above, the calculation cost for generating the classification model parameters in the classification model construction unit 212, which will be described later, increases according to the order of the NMF. Further, as described above, the classification performance by the classification unit 218 can be improved according to the order of NMF.

Here, the order of NMF means the number of basis vectors included in the teacher basis matrix, and in the example shown in FIG. 3, the number of basis vectors included in the teacher basis matrix is k. As shown in FIG. 3, the number k of the basis vectors included in the teacher basis matrix can be set regardless of the order m of the observation vector y included in the observation matrix Y or the number n of the observation vectors y.

Therefore, the teacher data processing unit 208 according to the present embodiment determines the number of basis vectors included in the teacher basis matrix so as to achieve both improvement in classification performance and suppression of calculation cost. Further, even if the number of basis vectors included in the teacher basis matrix is the same, the classification performance by the classification unit 218 can be influenced by the basis vectors included in the teacher basis matrix. Therefore, it is desirable that the teacher basis matrix includes a basis vector that can contribute to the improvement of classification performance.

Therefore, the teacher data processing unit 208 according to the present embodiment has a teacher basis matrix and a teacher based on the basis vector obtained by applying NMF to the teacher data, the performance information regarding the classification performance, and the cost information regarding the cost. Generate a coefficient matrix. Hereinafter, the function of the teacher data processing unit 208 will be described in more detail.

As described above, the teacher data processing unit 208 according to the present embodiment generates a teacher coefficient matrix based on a teacher basis matrix including a plurality of basis vectors and teacher data (teacher observation matrix). Here, the procedure for generating the teacher coefficient matrix _WL will be described with the basis matrix and the coefficient matrix obtained by decomposing the teacher observation matrix Y _L as the teacher basis matrix _HL and the teacher coefficient matrix _WL , respectively.

Assuming that the coefficient matrix generated from the teacher observation matrix Y _L based on the teacher base matrix _HL is the teacher coefficient matrix _{WL, the teacher observation matrix Y L ,} the teacher coefficient matrix _LL , and the teacher base are obtained from the equation (1). The relationship of the matrix _HL is expressed by the following equation (2).

From the above equation (2), it is necessary to use the inverse matrix of the teacher basis matrix _HL in order to obtain the teacher coefficient matrix _HL . However, since the teacher basis matrix _HL is not always a regular matrix in general, it may not have an inverse matrix. Therefore, the teacher data processing unit 208 may generate the teacher coefficient matrix _WL based on, for example, the pseudo-inverse matrix of the teacher basis matrix (the pseudo-inverse matrix of Moore-Penrose). Since the pseudo-inverse matrix _HL ⁺ of the teacher base matrix _HL is a matrix of k × m and generally k <m, the pseudo-inverse matrix _HL ⁺ decomposes the teacher base matrix _HL into singular values ( _HL) . = _{UL Σ L VL ) The} orthogonal matrix is _{UL, VL, the diagonal matrix is Σ L, and the reciprocal of the non-zero singular value is Σ L + among the} ^components of Σ _L. , Obtained by the following equation.

Using the pseudo-inverse matrix _HL ⁺ obtained in the above equation (3), the teacher coefficient matrix _WL is expressed by the following equation (4). However, the teacher coefficient matrix WL of the formula (4) is not the teacher coefficient matrix obtained by applying NMF as in the formula (2), but minimizes || Y _L - _{W L HL} || ² . Since it is a minimum squared solution, it does not always completely match the _WL of the equation (2).

Further, the teacher data processing unit 208 according to the present embodiment generates a teacher base matrix candidate as a candidate for the teacher base matrix in order to obtain a teacher base matrix that achieves both improvement in classification performance and suppression of calculation cost. .. A teacher basis matrix candidate is a matrix containing a plurality of basis vectors, like the teacher basis matrix.

The teacher data processing unit 208 according to the present embodiment may specify a basis vector included in the teacher base matrix candidate from a plurality of basis vectors obtained by applying NMF to the teacher data. For example, the teacher data processing unit 208 may specify the basis vector included in the teacher basis matrix candidate based on the basis vector performance information (an example of the performance information) indicating the classification performance for each basis vector. The basis vector performance information may be a value calculated for each basis vector, and hereinafter, the value of the basis vector performance information is referred to as a first evaluation value I ₁ .

For example, in the present embodiment, the classification accuracy when the teacher data is classified using the coefficient vector for each basis vector as the feature amount may be used as the basis vector performance information. In the present specification, the coefficient vector for the basis vector is obtained by multiplying the teacher data by the pseudo-inverse matrix of the basis vector, and the coefficients obtained for each observation vector included in the teacher data are arranged by the number of observation vectors. It is a vector. In the following, the classification accuracy when the teacher data is classified using the coefficient vector for _each basis vector as the feature quantity is defined as P1. That is, in this embodiment, I ₁ = P ₁ .

As described above, since the teacher data is associated with normal / abnormal correct answer labels, it is possible to classify the teacher data and perform cross-validation, for example. Therefore, as the classification accuracy P1, the _average classification accuracy obtained by performing cross-validation on the normal / abnormal data in the teacher data may be used. The present invention is not limited to such _an example, and the case where the worst classification accuracy is obtained may be emphasized, and the worst classification accuracy may be used as the classification accuracy P1.

For example, the teacher data processing unit 208 according to the present embodiment calculates the first evaluation value I ₁ for each basis vector, and is the most thirst from a plurality of basis vectors obtained by applying NMF to the teacher data. A basis vector having a high evaluation value I ₁ of 1 may be extracted. Then, the basis vector extracted (specified) in this way may be included in the teacher basis matrix candidate.

Further, the teacher data processing unit 208 generates a teacher basis matrix based on the basis matrix performance information (an example of the performance information) indicating the classification performance of the teacher basis matrix candidates including the basis vector specified as described above. You may. For example, the teacher data processing unit ₂₀₈ may use the classification accuracy P2 when the teacher data is classified using a plurality of coefficient vectors for a plurality of basis vectors included in the teacher basis matrix candidate as feature quantities as the basis matrix performance information. good.

Further, the teacher data processing unit 208 may generate a teacher basis matrix based on the basis matrix performance information corresponding to the teacher basis matrix candidate and the cost information corresponding to the teacher basis matrix candidate. The teacher data processing unit 208 in the present embodiment may use the calculation cost based on the number of basis vectors included in the teacher basis matrix candidate as cost information. For example, the calculation cost C ₂ is expressed by the following mathematical formula (5), where N is the data length of the teacher data and m is the number of basis vectors included in the current teacher basis matrix candidate.

For example, the teacher data processing unit 208 according to the present embodiment shows the classification performance of the current teacher basis matrix candidate with respect to the calculation cost based on the basis matrix performance information and the cost information corresponding to the teacher basis matrix candidate. The evaluation value I ₂ of may be calculated. For example, the second evaluation value I ₂ is expressed by the following mathematical formula (6) using P ₂ and C ₂ described above.

The second evaluation value I ₂ becomes larger as the classification performance of the teacher basis matrix candidate is higher, and becomes larger as the calculation amount is smaller. Therefore, for example, the teacher data processing unit 208 according to the present embodiment may use a teacher basis matrix candidate such that the second evaluation value I ₂ becomes larger as the teacher basis matrix.

For example, the teacher data processing unit 208 according to the present embodiment may calculate a second evaluation value I ₂ each time a teacher basis matrix candidate is generated. Then, if the second evaluation value I ₂ does not satisfy the condition, the teacher data processing unit 208 according to the present embodiment calculates a residual based on the teacher basis matrix candidate, and described above based on the residual. The process may be repeated. Assuming that the data is D, the extracted basis vector is H ₁ , and the coefficient vector for the basis vector H ₁ is W ₁ , the residual R for the data D is expressed by the following mathematical formula (7).

For example, the teacher data processing unit 208 according to the present embodiment further obtains 1 from a plurality of basis vectors obtained by applying NMF to the residual R if the second evaluation value I ₂ does not satisfy the condition. The basis vector of may be extracted. The method of extracting the basis vector may be a method using the first evaluation value I ₁ calculated for each basis vector as described above. Then, the teacher data processing unit 208 according to the present embodiment may add the extracted basis vector to the teacher basis matrix candidate to generate a new teacher basis matrix candidate including the basis vector.

As described above, the teacher data processing unit 208 according to the present embodiment may repeatedly generate teacher basis matrix candidates until the second evaluation value I ₂ satisfies the condition.

Here, in the present embodiment, the condition that the second evaluation value I ₂ should be satisfied is that the second evaluation value I ₂ calculated this time is better than the second evaluation value I ₂ calculated last time, for example. It may be small. When the generation of the teacher basis matrix candidate is repeated as described above, the basis vector included in the teacher basis matrix candidate increases each time the generation of the teacher basis matrix candidate is repeated. Therefore, it is expected that the classification accuracy P2 when the teacher data is classified using a plurality of coefficient vectors for a plurality of basis vectors included in the teacher basis matrix candidate as a feature quantity increases according to the _number of repetitions. The pace of increase can gradually slow down. On the other hand, since the number m of the basis vectors included in the teacher basis matrix candidate increases according to the number of repetitions and the data length N of the teacher data does not change according to the number of repetitions, the calculation cost C ₂ is constant according to the number of repetitions. Increases at the pace of. Therefore, if the second evaluation value I ₂ calculated this time is smaller than the second evaluation value I ₂ calculated last time, the previous teacher corresponding to the second evaluation value I ₂ calculated last time. Basis matrix candidates are considered desirable as teacher basis matrices. Therefore, the teacher data processing unit 208 according to the present embodiment is the current teacher basis matrix when the second evaluation value I ₂ calculated this time is smaller than the second evaluation value I ₂ calculated last time. The teacher basis matrix may be generated by deleting the last added basis vector from the candidates.

The condition to be satisfied by the second evaluation value I ₂ and the method of generating the teacher basis matrix when the condition is satisfied are not limited to the above-mentioned example. For example, the condition to be satisfied by the second evaluation value I ₂ may include being larger than a predetermined threshold value. Then, the teacher data processing unit 208 may use the current teacher basis matrix candidate as the teacher basis matrix when the second evaluation value I ₂ is larger than a predetermined threshold value.

Alternatively, the teacher data processing unit 208 repeats the generation of the teacher base matrix candidate a predetermined number of times, and among the teacher base matrix candidates generated so far, the teacher base matrix candidate having the largest second evaluation value I ₂ is selected. , May be a teacher basis matrix.

The function of the teacher data processing unit 208 in the learning stage has been described above. Next, the function of the teacher data processing unit 208 in the classification stage will be described.

The teacher data processing unit 208 is a basis obtained by applying _NMF to the residual RT when the residual _RT for the evaluation data is output from the feature extraction unit 216 described later in the classification stage. Vectors may be added to the teacher basis matrix. Then, the teacher data processing unit 208 may generate a teacher coefficient matrix by using the teacher data (teacher basis matrix) and the pseudo-inverse matrix of the teacher basis matrix to which the basis vector is added.

The teacher data processing unit 208 stores the generated teacher basis matrix in the storage unit 214 in both the learning stage and the classification stage, and stores the teacher coefficient matrix generated based on the teacher basis matrix in the classification model construction unit. Output to 212.

The classification model construction unit 212 generates a classification model parameter used by the classification unit 218 described later in the classification stage based on the teacher coefficient matrix generated by the teacher data processing unit 208 and the correct answer data associated with the teacher data. do. The classification model parameter generated by the classification model construction unit 212 may be a parameter corresponding to the classification model in the classification unit 218.

The classification model construction unit 212 generates classification model parameters based on the teacher coefficient matrix generated by the teacher data processing unit 208 in both the learning stage and the classification stage, and stores the classification model parameters in the storage unit 214. .. When the classification model parameters are already stored in the storage unit 214, the classification model parameters stored in the storage unit 214 are updated with the classification model parameters newly generated by the classification model construction unit 212. That is, the classification model construction unit 212 updates the classification parameters based on the teacher coefficient matrix generated by the teacher data processing unit 208 in the classification stage.

The storage unit 214 stores programs and parameters for each configuration of the sensor terminal 2 to function. Further, the storage unit 214 stores the teacher basis matrix generated by the teacher data processing unit 208 applying NMF to the teacher data, and the classification model parameters generated by the classification model construction unit 212.

The feature extraction unit 216 generates an evaluation coefficient matrix from the evaluation data (evaluation observation matrix) output from the preprocessing unit 206 in the classification stage based on the teacher basis matrix stored in the storage unit 214, and transfers the evaluation coefficient matrix to the classification unit 218. Output. Hereinafter, an example of generating an evaluation coefficient matrix by the feature extraction unit 216 will be described.

In the following, the evaluation observation matrix output from the preprocessing unit 206 is Y _T , the teacher base matrix stored in the storage unit 214 is _HL , and the coefficients generated from the evaluation observation matrix Y _T based on the teacher base matrix _HL . Let the matrix be the evaluation coefficient matrix candidate _WT , which is a candidate for the evaluation coefficient matrix. From the formula (1), the relationship between the evaluation observation matrix Y _T , the evaluation coefficient matrix candidate _WT , and the teacher basis matrix _HL is expressed by the following formula (8).

Using the pseudo-inverse matrix _HL ⁺ obtained by the same method as the above formula (3), the evaluation coefficient matrix candidate _WT is expressed by the following formula (9). However, the evaluation coefficient matrix candidate _WT of the formula (9) is not the teacher coefficient matrix obtained by applying NMF as in the formula (2), but || Y _L -W _{L HL} || ² is minimized. Since it is the least squared solution, it does not always exactly match the _WL of the equation (2).

The feature extraction unit 216 can generate an evaluation coefficient matrix candidate _WT from the evaluation observation matrix Y _T as in the above mathematical formula (9). Further, the feature extraction unit 216 according to the present embodiment may calculate the residual with respect to the evaluation observation matrix Y _T (evaluation data) based on the teacher basis matrix _HL and the evaluation coefficient matrix candidate _WT . The residual _RT with respect to the evaluation observation matrix Y _T is expressed by the following equation (10) using the teacher basis matrix _HL and the evaluation coefficient matrix candidate _WT .

The feature extraction unit 216 according to the present embodiment determines whether or not the residual _RT with respect to the evaluation observation matrix Y _T calculated as described above satisfies a predetermined condition, and the residual _RT is a predetermined value. When the condition is satisfied, the evaluation coefficient matrix candidate may be output to the classification unit 218 as an evaluation coefficient matrix. The predetermined condition that the residual _RT should satisfy may be, for example, that the residual _RT is sufficiently small, or that the residual _RT is equal to or less than a predetermined threshold value. It is desirable that a sufficiently small value is used for such a predetermined threshold value. With this configuration, if the evaluation observation matrix Y _T (evaluation data) can be represented sufficiently accurately by the currently used teacher basis matrix _HL , the evaluation generated using the teacher basis matrix _HL is used. The coefficient matrix candidate _WT is output to the classification unit 218 as the evaluation coefficient matrix _WT .

On the other hand, if the residual _{RT does not satisfy a predetermined condition, for example, if the residual RT is larger than a predetermined threshold in the above-mentioned example of the condition, the evaluation observation matrix Y T is the} currently used teacher basis matrix H. There is a possibility that _L cannot be expressed accurately enough. As a result, the classification performance of the classification unit 218 may deteriorate.

Therefore, the feature extraction unit 216 according to the present embodiment may output the residual _RT to the teacher data processing unit 208 when the residual _RT does not satisfy a predetermined condition. As a result, as described above, the teacher data processing unit 208 adds the basis vector to the teacher basis matrix based on the residual _RT , and updates the teacher coefficient matrix based on the teacher basis matrix. Further, as described above, the classification model construction unit 212 updates the classification parameters based on the teacher coefficient matrix updated by the teacher data processing unit 208.

If the residual _RT does not satisfy a predetermined condition, the feature extraction unit 216 may generate an evaluation coefficient matrix based on the teacher basis matrix to which the basis vector is added as described above. The evaluation coefficient matrix in such a case can be generated by setting the teacher basis matrix to which the basis vector is added to _HL and the evaluation coefficient matrix to _WT in the equations (8) and (9). Further, the feature extraction unit 216 outputs the generated evaluation coefficient matrix _WT to the classification unit 218.

With this configuration, if the evaluation observation matrix Y _T cannot be represented sufficiently accurately by the currently used teacher basis matrix _HL , a new basis vector is converted to the teacher basis matrix _HL based on the residual _RT . By adding it, it is possible to further improve the classification performance of the classification unit 218.

The classification unit 218 classifies the evaluation coefficient matrix _WT output from the feature extraction unit 216 based on the classification model parameters stored in the storage unit 214. The classification unit 218 outputs the classification result of the evaluation coefficient matrix _WT to the communication unit 220.

The classification unit 218 uses the coefficient vector included in the evaluation coefficient matrix _WT as a feature vector, and can classify based on the classification model parameters according to the classification model. The classification model used by the classification unit 218 is not particularly limited, but the classification unit 218 may use, for example, a classification model based on a multi-class SVM (Support Vector Machine) or the like.

As described above, by classifying using the coefficient vector indicating the weight for the basis matrix as the feature vector, it is possible to improve the classification performance as compared with the classification based on the magnitude of the power of a specific frequency, for example. ..

In addition, compared to the case where the classification process is performed using the evaluation data itself, by performing the classification process using the evaluation coefficient matrix _WT generated from the evaluation data, the dimension of the feature vector in the classification process is reduced and the classification is performed. It is possible to suppress the processing amount of processing.

The communication unit 220 transmits the classification result by the classification unit 218 to the aggregation device 3 shown in FIG.

<< 3. Operation example >>
The system configuration of one embodiment of the present invention and the configuration example of the sensor terminal 2 have been described above. Subsequently, an operation example of the present embodiment will be sequentially described by dividing it into a learning stage and a classification stage.

(Example of operation in the learning stage)
FIG. 4 is a flowchart showing an operation example of the sensor terminal 2 according to the present embodiment in the learning stage. As shown in FIG. 4, first, the sensor 202 collects the sensor data (an example of time-series data) related to the observation target 1 by sensing (S102), and the AD conversion unit 204 converts the analog sensor data into digital sensor data. Convert to (S104).

Subsequently, the preprocessing unit 206 performs various preprocessing such as filtering processing for removing noise, measurement value conversion processing, and frequency conversion processing on the sensor data (S106). In step S106, the observation vectors obtained by various preprocessings are added to the teacher data (teacher observation matrix), and normal or abnormal correct labels are associated with them.

If the acquisition of teacher data has not been completed (NO in S108), the process returns to step S102. The determination (S108) as to whether or not the acquisition of the teacher data is completed may be performed by the preprocessing unit 206 based on, for example, a user input via an operation unit (not shown).

When the acquisition of the teacher data is completed (YES in S108), the teacher data processing unit 208 generates a teacher basis matrix and a teacher coefficient matrix based on the teacher data (teacher observation matrix) (S110). Here, the flow of processing in step S110 will be described in more detail with reference to FIG.

FIG. 5 is a flowchart showing the flow of processing in step S110. As shown in FIG. 5, first, the teacher data processing unit 208 applies NMF to the teacher data and generates a plurality of basis vectors (S112). In step S112, the teacher data processing unit 208 sets the order of the NMF to a sufficiently large number k _max , generates k _max basis vectors, and calculates a coefficient vector for each basis vector.

In the following step S114, the teacher data processing unit 208 calculates a first evaluation value I ₁ indicating the classification performance for each basis vector based on the coefficient vector for each basis vector. As described above, in the present embodiment, the first evaluation value I ₁ is a classification calculated for each basis vector by performing cross-validation when the teacher data is classified using the coefficient vector for each basis vector as a feature quantity. _The accuracy may be P1.

Subsequently, the teacher data processing unit 208 extracts the basis vector H ₁ having the highest first evaluation value I ₁ calculated in step S114 from the plurality of basis vectors generated in step S112 (S116).

Subsequently, the teacher data processing unit 208 adds the _basis vector H1 extracted in step S118 to the teacher basis matrix candidates, and shows the classification performance of the teacher basis matrix candidates with respect to the cost (calculation cost in this embodiment). The evaluation value I ₂ of 2 is calculated (S118). As described above, the second evaluation value I ₂ is calculated by, for example, the mathematical formula (6).

In the following step S120, the second evaluation value I ₂ calculated this time and the second evaluation value I ₂ calculated last time are compared. Since the second evaluation value is calculated for the first time when step S120 is processed first, the second evaluation value I ₂ calculated this time is smaller than the second evaluation value I ₂ calculated last time. It may be determined that there is no such thing. If it is determined that the second evaluation value I 2 calculated this time is not smaller than the _{second evaluation value I 2 calculated} last time (NO in step S120), the process proceeds to step S122.

In step S122, the teacher data processing unit 208 calculates the residual R based _on the _basis vector H1 extracted in step S116 this time and the coefficient vector W1 for the _basis vector H1. As described above, the residual R is calculated by, for example, the mathematical formula (7). The residual R calculated here will be used as the data D when the processing of the next step S122 is performed.

Subsequently, the teacher data processing unit 208 applies NMF to the residual R calculated in step 122, generates a plurality of basis vectors, and calculates a coefficient vector for each basis vector (S124). The order of NMF applied to the residual R in step S122 may be set to the order k _max of the NMF used in step S110. Alternatively, one less order may be set each time step S122 is repeated.

Then, the processes of steps S114 to S124 described above are repeated until it is determined in step S120 that the second evaluation value I 2 calculated this time is smaller than the _{second evaluation value I 2 calculated} last time. ..

If it is determined that the second evaluation value I _{2 calculated this time is smaller than the second evaluation value I 2 calculated} last time (YES in step S120), the process proceeds to step S130. In step S130, the teacher data processing unit 208 deletes the last added basis vector from the current teacher basis matrix candidates to generate a teacher basis matrix, and generates a teacher coefficient matrix based on the teacher basis matrix. In step S130, the teacher coefficient matrix can be generated as described with reference to mathematical formulas (2) to (5). Further, the teacher basis matrix generated in step S130 is stored in the storage unit 214.

The process flow of step S110 in FIG. 4 has been described in detail above. Returning to FIG. 4, the explanation of the operation in the learning stage is continued. In the following step S150, the classification model building unit 212 generates the classification model parameters based on the teacher coefficient matrix generated in step S110 and the correct answer label.

The operation example of this embodiment in the learning stage has been described above. Subsequently, an operation example of the present embodiment in the classification stage will be described.

(Operation example at the classification stage)
FIG. 6 is a flowchart showing an operation example of the present embodiment in the classification stage. As shown in FIG. 6, first, the sensor 202 of the sensor terminal 2 collects sensor data (an example of time-series data) related to the observation target 1 by sensing (S202), and the AD conversion unit 204 of the sensor terminal 2 is analog. The sensor data of the above is converted into digital sensor data (S204).

Subsequently, the preprocessing unit 206 of the sensor terminal 2 performs various preprocessing such as filtering processing for removing noise, measurement value conversion processing, and frequency conversion processing on the sensor data (S206). In step S206, the observation vectors obtained by various preprocessing are added to the evaluation data (evaluation observation matrix).

In the classification stage, the evaluation data (evaluation observation matrix) may contain one observation vector or a plurality of observation vectors. When the evaluation data includes a plurality of observation vectors, the above steps S202 to S206 may be repeated.

Subsequently, the sensor terminal 2 generates an evaluation coefficient matrix from the evaluation data based on the teacher basis matrix (S210). Here, the flow of processing in step S110 will be described in more detail with reference to FIG. 7.

FIG. 7 is a flowchart showing the flow of processing in step S210. As shown in FIG. 7, first, the feature extraction unit 216 generates an evaluation coefficient matrix candidate from the evaluation data (evaluation observation matrix) based on the teacher basis matrix stored in the storage unit 214 (S212).

Subsequently, the feature extraction unit 216 calculates the residual _RT for the evaluation data (evaluation observation matrix) based on the teacher basis matrix and the evaluation coefficient matrix candidates generated in step S212 (S212). As described above, the residual _RT is calculated by, for example, the mathematical formula (10).

Subsequently, the feature extraction unit 216 compares the residual _RT calculated in step S216 with a predetermined threshold value (S216). When the residual _RT is equal to or less than a predetermined threshold (NO in S216), the feature extraction unit 216 determines that the evaluation data can be sufficiently accurately represented by the currently used teacher basis matrix, and evaluates it. The coefficient matrix candidate is output to the classification unit 218 as an evaluation coefficient matrix.

On the other hand, when the residual _RT is larger than a predetermined threshold (YES in S216), the feature extraction unit 216 determines that the evaluation data cannot be sufficiently accurately represented by the currently used teacher basis matrix, and the residual. The _RT is output to the teacher data processing unit 208. Then, the teacher data processing unit 208 adds the basis vector obtained by applying NMF to the residual _RT output from the feature extraction unit 216 to the teacher basis matrix (S230). The teacher basis matrix stored in the storage unit 214 is updated by the teacher basis matrix to which the basis vector is newly added in step S230.

Further, the teacher data processing unit 208 generates (updates) a teacher coefficient matrix based on the teacher basis matrix to which the basis vector is newly added in step S230 and the teacher data (S232). Then, the classification model construction unit 212 generates a classification model parameter based on the teacher coefficient matrix updated in step S232 (S234). The classification model parameters stored in the storage unit 214 are updated by the classification model parameters generated in step S234.

Subsequently, the feature extraction unit 216 generates an evaluation coefficient matrix from the evaluation data based on the teacher basis matrix to which the basis vector is newly added in step S230 (S236).

The process flow of step S210 in FIG. 6 has been described in detail above. Returning to FIG. 6, the description of the operation in the classification stage will be continued. In the subsequent step S250, the classification unit 218 of the sensor terminal 2 classifies the evaluation coefficient matrix generated in step S210 based on the classification model parameters stored in the storage unit 214.

Subsequently, the communication unit 220 of the sensor terminal 2 transmits the classification result of the evaluation coefficient matrix obtained in step S250 to the aggregation device 3 (S252). The processes of steps S202 to S252 may be performed in parallel by each of the plurality of sensor terminals 2 included in the information processing system 9 according to the present embodiment.

Then, the communication unit 32 of the aggregation device 3 receives the classification result from each sensor terminal 2 (S254), and the aggregation unit 34 of the aggregation device 3 performs aggregation processing based on the plurality of classification results received in step S254. , The final classification result is specified (S256). Then, the output unit 36 of the aggregation device 3 outputs the final classification result specified in step S256 (S258).

<< 4. Modification example >>
The embodiment of the present invention has been described above. Hereinafter, some modifications of the above embodiment will be described. In addition, each modification described below may be applied to the said embodiment alone, or may be applied to the said embodiment in combination. Further, each modification may be applied in place of the configuration described in the above embodiment, or may be additionally applied to the configuration described in the above embodiment.

<4-1. Modification 1>
In the above embodiment, learning is performed using teacher data (teacher observation matrix) including data (observation vector) associated with one of the correct answer labels of two types of classes (operating state), normal and abnormal. Although examples have been described, the present invention is not limited to such examples. For example, all the data included in the teacher data may be associated with the correct answer label of the same type.

For example, depending on the observation target, it may be difficult to prepare both the sensor data in the normal state and the sensor data in the abnormal state. In particular, in order to prepare the sensor data in the abnormal state, it is necessary to operate the observation target in the abnormal state. Therefore, it is not possible to prepare the sensor data in the abnormal state in the observation target in which the occurrence of the abnormal state is not so frequent. Have difficulty. On the other hand, preparing only the sensor data in the normal state can be realized, for example, by performing sensing and collecting the sensor data while the observation target is operating normally without any problem. It is easier than preparing both sensor data in an abnormal state. Therefore, in the following, a case where all the data included in the teacher data are associated with the correct label of the same type as normal will be described as a modification 1. In addition, since the functional configuration of each part described in the above-described embodiment is only partially different in such a modification, only the differences from the above-described embodiment will be described below.

The preprocessing unit 206 according to this modification outputs the teacher data to the teacher data processing unit 208 in which all the included data (observation vectors) are associated with the correct label of the same type as normal. It is different from the pretreatment unit 206 described in the above.

Further, the teacher data processing unit 208 according to this modification is different from the teacher data processing unit 208 described in the above embodiment in the calculation method of the first evaluation value I ₁ and the second evaluation value I ₂ . In this modification, all the data (observation vectors) included in the teacher data are associated with the correct label of the same type as normal. Therefore, in the calculation of the first evaluation value I ₁ and the second evaluation value I ₂ in the above embodiment, the classification accuracy P ₁ and P ₂ used as the performance information indicating the classification performance cannot be calculated. ..

Therefore, the teacher data processing unit 208 according to this modification uses the information based on the sum of the Mahalanobis distances calculated by using each coefficient included in the coefficient vector with respect to the base vector as a sample as the performance information, and the first evaluation value I ₁ and the second evaluation value I ₂ may be calculated.

First, the teacher data processing unit 208 according to this modification is the first basis vector performance information based on the total Mahalanobis distance M ₁ calculated by using each coefficient included in the coefficient vector for each basis vector as a sample. The evaluation value I ₁ may be calculated. Sum of Mahalanobis distance M ₁ is the sum of Mahalanobis distances obtained by calculating the Mahalanobis distance for each sample from the remaining samples of one sample out of the whole sample, when each coefficient included in the coefficient vector is used as a sample. It can be calculated by.

Here, the total Mahalanobis distance M ₁ indicates a variation within the class of normality, and the smaller the total Mahalanobis distance M ₁ , the smaller the variation, and the better the classification performance when classifying with other classes. Indicates a high probability. Therefore, the teacher data processing unit 208 according to this modification uses the total Mahalanobis distance M ₁ to calculate the first evaluation value I ₁ which is the base vector performance information as shown in the following mathematical formula (11). May be good.

Further, the teacher data processing unit 208 according to this modification bases the total Mahalanobis distance _M2 calculated by using each coefficient included in the plurality of coefficient vectors for the plurality of basis vectors included in the teacher basis matrix candidate as a sample. It may be used as matrix performance information.

Then, the teacher data processing unit 208 according to this modification uses the total Mahalanobis distance M ₂ which is the basis matrix performance information and the calculation cost C ₂ described in the above embodiment to obtain the second evaluation value I ₂ . It may be calculated by the following formula (12).

The second evaluation value I ₂ expressed as in the formula (12) becomes larger as the total Mahalanobis distance M ₂ is smaller, and becomes larger as the calculation amount is smaller. In addition, the total Mahalanobis distance M ₂ shows a variation within the class of normal, and the smaller the total Mahalanobis distance M ₂ , the smaller the variation, and the classification performance can be improved when classifying with other classes. Indicates that the sex is high. Therefore, for example, the teacher data processing unit 208 according to this modification may use a teacher basis matrix candidate such that the second evaluation value I ₂ becomes larger as the teacher basis matrix, as in the example described in the above embodiment. good.

The classification model construction unit 212 and the classification unit 218 according to this modification are different from the classification model construction unit 212 and the classification unit 218 according to the above embodiment in that the classification model used by the classification unit 218 is different from the above embodiment. You may. For example, the classification unit 218 according to the present modification may use a classification model based on the one-class SVM, and the classification model construction unit 212 may generate classification model parameters according to the classification model. Further, the classification result by the classification unit 218 may not be a normal / abnormal classification result but a classification score indicating the degree of abnormality.

As described above, according to this modification, it is possible to perform learning and classification even when only sensor data in a normal state can be prepared.

<4-2. Modification 2>
In the above embodiment, an example in which the sensor terminal 2 performs classification has been described, but the present invention is not limited to such an example. For example, the function of the classification unit 218 described above may be provided in the aggregation device 3 shown in FIG. 1 instead of the sensor terminal 2. Such an example will be described as a modification 2.

The communication unit 220 according to this modification transmits the evaluation coefficient matrix generated by the feature extraction unit 216 to the aggregation device 3. Then, in the aggregation device 3, after the classification processing based on the evaluation coefficient matrix received from the sensor terminal 2 is performed, the aggregation processing is performed and the final classification result is output.

Therefore, in this modification, as the number of coefficient vectors included in the evaluation coefficient matrix increases, not only the calculation cost but also the communication cost from the sensor terminal 2 to the aggregation device 3 increases. Here, the number of coefficient vectors included in the evaluation coefficient matrix increases with the number of basis vectors included in the teacher basis matrix.

In the above embodiment, the calculation cost C ₂ is used as the cost information, but in this modification, the C ₂ can be used as the cost information indicating not only the calculation cost but also the communication cost. However, since the calculation method of C ₂ may be the same as that of the above-described embodiment, the function of the teacher data processing unit 208 according to the present modification may be the same as that of the teacher data processing unit 208 according to the above embodiment.

According to this modification, when the evaluation coefficient matrix is transmitted from the sensor terminal 2 to an external server such as the aggregation device 3 and the classification process is performed, not only the calculation cost but also the calculation cost is maintained while maintaining high classification performance. It is also possible to suppress communication costs. By suppressing the communication cost, the power consumption of the sensor terminal 2 is suppressed, the battery life is improved when the sensor terminal 2 is battery-powered, and the system can be operated for a longer period of time. Further, by suppressing the communication cost, it is possible to transmit the evaluation coefficient matrix related to the sensor data having a high sampling frequency even if the radio link has a low band.

<< 5. Hardware configuration >>
Each embodiment of the present invention has been described above. Information processing such as the pre-processing, teacher data processing, classification model construction processing, feature extraction processing, classification processing, and aggregation processing described above is realized by the cooperation between the software and the hardware of the sensor terminal 2 and the aggregation device 3. Hereinafter, the hardware configuration of the information processing device 1000 will be described as a hardware configuration example of the sensor terminal 2 and the aggregation device 3 which are the information processing devices according to the embodiment of the present invention.

FIG. 8 is an explanatory diagram showing a hardware configuration of the information processing apparatus 1000 according to the embodiment of the present invention. As shown in FIG. 8, the information processing apparatus 1000 includes a CPU (Central Processing Unit) 1001, a ROM (Read Only Memory) 1002, a RAM (Random Access Memory) 1003, an input device 1004, and an output device 1005. , A storage device 1006 and a communication device 1007.

The CPU 1001 functions as an arithmetic processing device and a control device, and controls the overall operation in the information processing device 1000 according to various programs. Further, the CPU 1001 may be a microprocessor. The ROM 1002 stores programs, calculation parameters, and the like used by the CPU 1001. The RAM 1003 temporarily stores a program used in the execution of the CPU 1001 and parameters that are appropriately changed in the execution. These are connected to each other by a host bus composed of a CPU bus or the like. Mainly, the functions of the preprocessing unit 206, the teacher data processing unit 208, the classification model construction unit 212, the feature extraction unit 216, the classification unit 218, and the aggregation unit 34 are mainly performed by the collaboration between the CPU 1001, ROM 1002 and RAM 1003 and the software. It will be realized.

The input device 1004 is an input means for a user to input information such as a mouse, a keyboard, a touch panel, a button, a microphone, a switch, and a lever, and an input control circuit that generates an input signal based on the input by the user and outputs the input signal to the CPU 1001. It is composed of such things. By operating the input device 1004, the user of the information processing device 1000 can input various data to the information processing device 1000 and instruct the processing operation.

The output device 1005 includes, for example, a liquid crystal display (LCD) device, an OLED device, and a display device such as a lamp. Further, the output device 1005 includes an audio output device such as a speaker and headphones. For example, the display device displays an captured image, a generated image, or the like. On the other hand, the voice output device converts voice data and the like into voice and outputs it. The output device 1005 corresponds to the output unit 36.

The storage device 1006 is a device for storing data. The storage device 1006 may include a storage medium, a recording device for recording data on the storage medium, a reading device for reading data from the storage medium, a deleting device for deleting data recorded on the storage medium, and the like. The storage device 1006 stores programs and various data executed by the CPU 1001. The storage device 1006 corresponds to the storage unit 214.

The communication device 1007 is a communication interface composed of, for example, a communication device for connecting to a communication network. Further, the communication device 1007 may include a wireless LAN (Local Area Network) compatible communication device, an LTE (Long Term Evolution) compatible communication device, a wire communication device for wired communication, or a Bluetooth (registered trademark) communication device. The communication device 1007 corresponds to the communication unit 220 and the communication unit 32.

<< 6. Conclusion >>
As described above, according to the embodiment of the present invention, it is possible to achieve both improvement in classification performance and cost reduction.

Although the preferred embodiments of the present invention have been described in detail with reference to the accompanying drawings, the present invention is not limited to these examples. It is clear that a person having ordinary knowledge in the field of technology to which the present invention belongs can come up with various modifications or modifications within the scope of the technical ideas described in the claims. , These are also naturally understood to belong to the technical scope of the present invention.

For example, in the above embodiment, the classification result of each sensor terminal 2 is transmitted to the aggregation device 3, aggregated and output, but the present invention is not limited to this example. For example, each sensor terminal 2 may have an output unit such as a display unit, and the classification result of each sensor terminal 2 may be output from the output unit of each sensor terminal 2. Such a configuration is effective when it is not necessary to comprehensively judge the classification results of a plurality of sensor terminals 2 and it is desired to individually confirm the classification results of each sensor terminal 2.

Further, in the present specification, as an observation target, equipment, devices, products, etc. of a factory are taken as an example, but the present invention is not limited to such an example. For example, the above-described embodiment can be applied to a general device having a mechanical mechanism such as an office printer or a store ATM as an observation target.

Further, in the above embodiment, vibration data, acoustic data, and the like are given as examples as time-series data and sensor data, but the present invention is not limited to such examples. For example, when the sensor data is image data, the above-described embodiment can be applied to the operation state detection using image recognition.

Further, each step in the above embodiment does not necessarily have to be processed in chronological order in the order described as a flowchart. For example, each step in the processing of the above embodiment may be processed in an order different from the order described in the flowchart, or may be processed in parallel.

Further, according to the above embodiment, it is possible to provide a computer program for causing hardware such as the CPU 1001, ROM 1002, and RAM 1003 to exhibit the same functions as the configurations of the sensor terminal 2 and the aggregation device 3 described above. Also provided is a recording medium on which the computer program is recorded.

1 Observation target 2 Sensor terminal 3 Aggregator 5 Network 9 Information processing system 32 Communication unit 34 Aggregation unit 36 Output unit 202 Sensor 204 AD conversion unit 206 Preprocessing unit 208 Teacher data processing unit 212 Classification model construction unit 214 Storage unit 216 Feature extraction Part 218 Classification part 220 Communication part

Claims (11)

A pre-processing unit that acquires non-negative teacher data based on time-series data,
The base vector is acquired by applying the non-negative matrix factor decomposition to the teacher data, and the information indicating the classification performance corresponding to the base vector is acquired as the base vector performance information based on the base vector. Based on the base vector performance information, the base vector included in the teacher base matrix candidate that is a candidate for the teacher base matrix is specified, and the information indicating the classification performance corresponding to the teacher base matrix candidate based on the teacher base matrix candidate. Is acquired as base matrix performance information, and based on the number of base vectors included in the teacher base matrix candidate, information indicating the calculation cost corresponding to the teacher base matrix candidate is acquired as cost information, and the teacher base matrix candidate is acquired. An evaluation value is calculated by combining the base matrix performance information and the cost information corresponding to the above , the teacher base matrix is specified from the teacher base matrix candidates based on the evaluation value, and the teacher base matrix and the teacher data are used. A teacher data processing unit that generates a teacher coefficient matrix based on
Information processing device equipped with. The basis matrix performance information is the information according to claim 1 , which is information based on the classification accuracy when the teacher data is classified using a plurality of coefficient vectors for a plurality of the basis vectors included in the teacher basis matrix candidates as feature quantities. Information processing equipment. All the data included in the teacher data are associated with labels of the same type.
The basis matrix performance information is information based on the sum of Mahalanobis distances calculated by using each coefficient included in a plurality of coefficient vectors for a plurality of the basis vectors included in the teacher basis matrix candidate as a sample, according to claim 1 . The information processing device described. The information processing apparatus according to any one of claims 1 to 3, wherein the basis vector performance information is information based on classification accuracy when the teacher data is classified using a coefficient vector for the basis vector as a feature amount. All the data included in the teacher data are associated with labels of the same type.
The information processing apparatus according to any one of claims 1 to 3, wherein the basis vector performance information is information based on the sum of Mahalanobis distances calculated by using each coefficient included in the coefficient vector for the basis vector as a sample. .. The information processing apparatus according to any one of claims 1 to 5 , further comprising a model building unit that generates a classification model parameter based on the teacher coefficient matrix and a correct answer label associated with the teacher data. Information processing equipment. The preprocessing unit further acquires non-negative evaluation data based on the further time series data.
The information processing device is
A feature extraction unit that generates an evaluation coefficient matrix from the evaluation data based on the teacher basis matrix,
A classification unit that classifies the evaluation coefficient matrix based on the classification model parameters,
The information processing apparatus according to claim 6 , further comprising. The feature extraction unit generates an evaluation coefficient matrix candidate from the evaluation data, and when the residual to the evaluation data calculated based on the teacher base matrix and the evaluation coefficient matrix candidate satisfies a predetermined condition, the feature extraction unit generates the evaluation coefficient matrix candidate. The information processing apparatus according to claim 7 , wherein the evaluation coefficient matrix candidate is the evaluation coefficient matrix. The teacher data processing unit adds a basis vector obtained by applying non-negative matrix factorization to the residual when the residual does not satisfy the predetermined condition to the teacher base matrix. Based on this, the teacher coefficient matrix is further generated.
The information processing apparatus according to claim 8 , wherein the model building unit updates the classification model parameters based on the generated teacher coefficient matrix. Obtaining non-negative teacher data based on time series data,
The base vector is acquired by applying the non-negative matrix factor decomposition to the teacher data, and the information indicating the classification performance corresponding to the base vector is acquired as the base vector performance information based on the base vector. Based on the base vector performance information, the base vector included in the teacher base matrix candidate that is a candidate for the teacher base matrix is specified, and the information indicating the classification performance corresponding to the teacher base matrix candidate based on the teacher base matrix candidate. Is acquired as base matrix performance information, and based on the number of base vectors included in the teacher base matrix candidate, information indicating the calculation cost corresponding to the teacher base matrix candidate is acquired as cost information, and the teacher base matrix candidate is acquired. An evaluation value is calculated by combining the base matrix performance information and the cost information corresponding to the above , the teacher base matrix is specified from the teacher base matrix candidates based on the evaluation value, and the teacher base matrix and the teacher data are used. To generate a teacher coefficient matrix based on
Information processing methods performed by an information processing device , including. On the computer
Ability to acquire non-negative teacher data based on time series data,
The base vector is acquired by applying the non-negative matrix factor decomposition to the teacher data, and the information indicating the classification performance corresponding to the base vector is acquired as the base vector performance information based on the base vector. Based on the base vector performance information, the base vector included in the teacher base matrix candidate that is a candidate for the teacher base matrix is specified, and the information indicating the classification performance corresponding to the teacher base matrix candidate based on the teacher base matrix candidate. Is acquired as base matrix performance information, and based on the number of base vectors included in the teacher base matrix candidate, information indicating the calculation cost corresponding to the teacher base matrix candidate is acquired as cost information, and the teacher base matrix candidate is acquired. An evaluation value is calculated by combining the base matrix performance information and the cost information corresponding to the above , the teacher base matrix is specified from the teacher base matrix candidates based on the evaluation value, and the teacher base matrix and the teacher data are used. With the ability to generate a teacher coefficient matrix based on
A program to realize.

Images