JP7485302B2 - Reliability Evaluation Equipment - Google Patents

Description

The present disclosure relates to a reliability evaluation device that evaluates the reliability of inferences made by a learning means.

For example, in order to improve the performance and functionality of mechanical systems such as industrial robots, industrial machines, or machine tools, there is a demand for high-performance command generation units and servo control units that apply artificial intelligence. In particular, there is a demand for high-precision modeling using artificial intelligence, and for real-time control of mechanical systems using a control device using the high-precision model constructed by artificial intelligence. In some cases, relatively high performance can be obtained by using a model generated by artificial intelligence, but until now, there has been no appropriate method for evaluating the reliability of the inference results of artificial intelligence, and there is a risk that performance will deteriorate if the output of artificial intelligence is used as is for unlearned data. Therefore, there is a demand for technology that applies artificial intelligence to the control of mechanical systems while evaluating the output of artificial intelligence.

Patent Document 1 discloses a technology that uses a first learning device that performs supervised learning using training data and correct answer data that indicates the correct answer of the inference result for the training data, and a second learning device that performs unsupervised learning using the training data, in which input data used when inferring using the learning results of the first learning device is also input when inferring using the second learning device, and the validity of the output of the first learning device is evaluated based on the output of the second learning device.

JP 2019-139277 A

In conventional technology, the accuracy of inference varies for each piece of learned data, so it is not possible to properly evaluate the reliability of the inference made by a learning machine simply by determining whether the input at the time of inference is close to the learned data.

The present disclosure has been made in consideration of the above, and aims to provide a reliability evaluation device that appropriately evaluates the reliability of learning performed by a learning means.

In order to solve the above-mentioned problems and achieve the objective, the reliability evaluation device according to the present disclosure has a first learning means for learning using input data and correct answer data or target specifications as input, a second learning means for performing clustering or class classification on the input data, an evaluation result storage means for storing an evaluation result of the learning performed by the first learning means, a neighborhood group extraction means for performing clustering or class classification on the input data at the time of inference based on the result of the clustering or class classification performed by the second learning means to extract multiple clusters or classes, and an evaluation value determination means for determining a reliability evaluation value for evaluating the reliability of the learning performed by the first learning means based on the evaluation result stored by the evaluation result storage means and information on the multiple clusters or classes extracted by the neighborhood group extraction means.

The reliability evaluation device according to the present disclosure has the effect of being able to appropriately evaluate the reliability of the inferences made by the learning means.

FIG. 1 is a diagram showing a configuration during learning of a reliability evaluation device according to a first embodiment; FIG. 1 is a diagram showing a configuration of a first learning means included in a reliability evaluation apparatus according to a first embodiment; FIG. 1 is a diagram showing a configuration of a reliability evaluation device according to a first embodiment at the time of inference; FIG. 13 is a diagram showing a configuration of a reliability evaluation device according to a second embodiment at the time of inference. FIG. 13 is a diagram showing a configuration of a reliability evaluation device according to a third embodiment at the time of inference. FIG. 13 is a diagram showing a configuration of a reliability evaluation device according to a fourth embodiment at the time of inference. FIG. 13 is a diagram showing a configuration during learning of a reliability evaluation device according to a fifth embodiment. FIG. 13 is a diagram showing a configuration of a reliability evaluation device according to a fifth embodiment at the time of inference. FIG. 23 is a diagram showing a configuration during learning of a reliability evaluation device according to a sixth embodiment. FIG. 23 is a diagram showing a configuration during learning of a reliability evaluation device according to a seventh embodiment. FIG. 1 is a diagram showing a processor in the case where at least a part of a first learning means, a second learning means, a neighborhood group extraction means, an evaluation value estimation means, and an evaluation value determination means included in a reliability evaluation device according to a first embodiment is realized by a processor. FIG. 1 is a diagram showing a processing circuit in the case where at least a part of a first learning means, a second learning means, a neighborhood group extraction means, an evaluation value estimation means, and an evaluation value determination means of a reliability evaluation device according to a first embodiment is realized by the processing circuit.

The reliability evaluation device according to the embodiment is described in detail below with reference to the drawings.

Embodiment 1.
1 is a diagram showing a configuration of a reliability evaluation apparatus 1 according to a first embodiment during learning. The reliability evaluation apparatus 1 has a first learning means 2 that receives input data and teacher data or a target specification and performs learning. In this embodiment, the reliability evaluation apparatus 1 receives the input data and teacher data. That is, the reliability evaluation apparatus 1 has a first learning means 2 that receives the input data and supervised data and learns the relationship between the input data and the supervised data. The first learning means 2 has a first learning device 21 that learns the relationship between the input data and the teacher data. The teacher data is supervised data.

For example, the teacher data is the parameters of the controlled object, or the parameters of a control device that controls the controlled object. For example, the controlled object is a robot, a machine tool, or an industrial machine. For example, the input data is data indicating the position, speed, acceleration, or current of the motor of each axis that drives the controlled object, or the output of a sensor attached to the controlled object. The input data does not have to be, for example, the measured current data itself, but may be processed data, such as the difference between the current and a predicted current predicted by some means.

FIG. 2 is a diagram showing the configuration of the first learning means 2 included in the reliability evaluation device 1 according to the first embodiment. As described above, the first learning means 2 includes a first learning device 21. For example, the first learning device 21 performs learning using a feedforward neural network. The first learning device 21 may also perform learning using a recurrent network.

The first learning means 2 further includes a learning result evaluation means 22 that receives the learning result and correct answer data performed by the first learning device 21. The learning result evaluation means 22 evaluates the result of learning performed by the first learning device 21 based on a combination of various input data and teacher data. For example, the learning result evaluation means 22 calculates the absolute value of the difference between the teacher data and the output of the first learning device 21. The output of the first learning device 21 is the output when an input paired with the teacher data is input to the first learning device 21 after the first learning device 21 has performed learning. The learning result evaluation means 22 may not calculate the absolute value, but may only calculate the difference between the teacher data and the output of the first learning device 21. The learning result evaluation means 22 may divide the absolute value of the difference between the teacher data and the output of the first learning device 21 by the magnitude of the teacher data.

The reliability evaluation device 1 further has an evaluation result storage means 3 that stores the evaluation results obtained by the learning result evaluation means 22. For example, the evaluation result storage means 3 is realized by a semiconductor memory. The learning result evaluation means 22 outputs the calculated evaluation results at the time of learning to the evaluation result storage means 3. The evaluation result storage means 3 receives the evaluation results at the time of learning output from the learning result evaluation means 22 and stores the evaluation results at the time of learning. In other words, the evaluation result storage means 3 stores the evaluation results of the learning performed by the first learning means 2.

The reliability evaluation device 1 further has a second learning means 4 that receives information about the input data received by the first learning means 2. In this embodiment, the input data itself is used as the information about the input data. In other words, the input data input to the first learning means 2 is also input to the second learning means 4. The second learning means 4 performs clustering on the input data. The second learning means 4 has a clustering means 41 that performs clustering on the input data. The clustering means 41 performs clustering on the input data by, for example, executing the k-means method or the kSHAPE method.

When the clustering means 41 uses the k-means method, if the input data is time-series data, the clustering means 41 may use the k-means method after extracting features of the time-series data, such as the maximum or average value. For example, the time-series data is time-series data indicating the position of a motor or the current of a motor. Even if the input data is time-series data, the clustering means 41 may perform clustering on the input data, for example, using the kSHAPE method, without extracting features. The second learning means 4 outputs the result of the clustering to the evaluation result storage means 3. The result is information related to the clustering performed by the second learning means 4.

The evaluation result storage means 3 stores the evaluation results of the learning performed by the first learning means 2, and also stores information about the clustering performed by the second learning means 4. For example, the evaluation result storage means 3 stores the evaluation results for the input data closest to the center of the cluster as the evaluation result corresponding to the cluster. The evaluation result storage means 3 may store the average value of multiple evaluation results close to the center of the cluster as the evaluation result corresponding to the cluster, may store the worst value of all evaluation results of the input data belonging to the cluster, or may store the average value and variance of all evaluation results of the input data belonging to the cluster. The input data being closest to the center of the cluster or close to the cluster center is determined from the distance between the input data and the cluster center. If the k-means method is used after extracting the features, the determination may be made from the distance between the features of the cluster center and the features of the input. In this case, the distance between the input data and the cluster center is the square root of the sum of the squares of the differences between the elements of the feature vector of the input data and the feature vector of the cluster center when the features are expressed as vectors. If the time series data is used as is, the determination may be made using shape-based distance (SBD), DTW (Dynamic Time Warping), or the like.

Figure 3 is a diagram showing the configuration of the reliability evaluation device 1 according to the first embodiment at the time of inference. The first learning device 21 receives input data at the time of inference and outputs an inference result. The reliability evaluation device 1 further has a neighborhood group extraction means 5 that receives the input data at the time of inference. In other words, the input data at the time of inference is input to the first learning device 21 and the neighborhood group extraction means 5, which is the learning result of the second learning means 4.

The neighborhood group extraction means 5 has the clustering means 41 that the second learning means 4 has. That is, the clustering means 41 is included in the second learning means 4 and the neighborhood group extraction means 5. The neighborhood group extraction means 5 performs clustering on the input data at the time of inference based on the results of clustering performed by the second learning means 4 to extract multiple clusters. For example, the neighborhood group extraction means 5 extracts multiple clusters having centers in the vicinity of the results of clustering on the input data. The proximity is determined from the distance between the input data and the cluster center. If the k-means method is used after extracting the features, the determination can be made from the distance between the features of the cluster center and the features of the input. If the time series data is used as is, the determination can be made using shape-based distance (SBD), DTW (Dynamic Time Warping), etc.

The reliability evaluation device 1 further includes an evaluation value estimation means 6 that reads out the evaluation results for each of the multiple clusters extracted by the neighborhood group extraction means 5 from the evaluation result storage means 3. The reliability evaluation device 1 further includes an evaluation value determination means 7. The evaluation value estimation means 6 outputs the evaluation results for each of the multiple clusters read out from the evaluation result storage means 3 and information on the multiple clusters extracted by the neighborhood group extraction means 5 to the evaluation value determination means 7.

The evaluation value determination means 7 determines a reliability evaluation value for evaluating the reliability of the learning performed by the first learning means 2 based on the evaluation results stored by the evaluation result storage means 3 and information on the multiple clusters extracted by the neighborhood group extraction means 5. For example, the evaluation value determination means 7 calculates a weighted sum of the evaluation results according to the distance to the center of each of the multiple clusters, determines the calculated weighted sum of the evaluation results as the reliability evaluation value, and outputs the determined reliability evaluation value. The above distance is Euclidean distance or Mahalanobis distance. The evaluation value determination means 7 may adopt the evaluation value of the cluster closest to the center of the cluster and determine the adopted evaluation value as the reliability evaluation value. The evaluation value determination means 7 may adopt the worst value among the evaluation results of the multiple extracted clusters and determine the adopted evaluation value as the reliability evaluation value.

When the evaluation results read by the evaluation value estimation means 6 are the mean values and variances of multiple clusters, the evaluation value determination means 7 may output the average values of the mean values and variances as the reliability evaluation value, or may output the worst values of the mean values and variances. For a cluster whose distance to the nearest cluster is within a predetermined threshold, the evaluation value determination means 7 may adopt the evaluation value of that cluster, and for a cluster whose distance to the nearest cluster exceeds the threshold, the evaluation value determination means 7 may output the result of increasing the evaluation value of the cluster according to the distance, that is, the result of correcting the evaluation value in the direction that worsens the evaluation value.

In the first embodiment, an evaluation value is used in which the larger the value, the lower the reliability, but an evaluation value in which the smaller the value, the lower the reliability may be used. When an evaluation value in which the smaller the value, the lower the reliability, is used, the evaluation value determination means 7 corrects the evaluation value in a direction that decreases the reliability evaluation value when correcting the evaluation value in a worsening direction.

As described above, the reliability evaluation device 1 according to the first embodiment evaluates the reliability of the inference result according to the domain to which the input data at the time of inference belongs. The inference result is the output of the first learning means 2. More specifically, the inference result is the inference result when the learning result performed by the first learning means 2 is used. Therefore, the reliability evaluation device 1 can evaluate the reliability of the inference result with a relatively high degree of accuracy. In other words, the reliability evaluation device 1 can appropriately evaluate the reliability of the inference when the learning result performed by the first learning means 2 is used.

Embodiment 2.
4 is a diagram showing the configuration of the reliability evaluation device 1A according to the second embodiment at the time of inference. The configuration of the reliability evaluation device 1A differs from the configuration of the reliability evaluation device 1 according to the first embodiment only at the time of inference. In the second embodiment, the configuration and operation at the time of inference that differ from the first embodiment will be mainly described. The reliability evaluation device 1A has all the components that the reliability evaluation device 1 has. The reliability evaluation device 1A further has a neighborhood direction determination means 8 that determines whether the input data at the time of inference is data located in the middle of the input data at the time of learning, or is data in a region outside the range of the input data at the time of learning.

During inference, input data is input to the first learning device 21 and the neighborhood group extraction means 5. The neighborhood group extraction means 5 extracts multiple clusters having centers in the vicinity of the clustering results for the input data. The evaluation value estimation means 6 reads out the evaluation results for each of the multiple clusters extracted by the neighborhood group extraction means 5 from the evaluation result storage means 3, and outputs the read evaluation results and information on the extracted multiple clusters to the evaluation value determination means 7. The evaluation value determination means 7 calculates a weighted sum of the evaluation results according to the distance to the center of each of the multiple clusters, and determines and outputs the calculated weighted sum of the evaluation results as a reliability evaluation value. The above distance is Euclidean distance or Mahalanobis distance.

The neighborhood direction determination means 8 determines whether the input data at the time of inference is data located in the middle of the input data at the time of learning, or is data in a region outside the range of the input data at the time of learning, and outputs the determination result to the evaluation value determination means 7. For example, the neighborhood direction determination means 8 converts the input data into features, calculates the difference between the value of each element of the feature obtained by the conversion and the value of each element of the feature at the center of the cluster determined to be nearby, and determines that the input data at the time of inference is data in a region outside the range of the input data at the time of learning if the signs of the differences are all the same.

The neighborhood direction determination means 8 may derive one large cluster including all clusters analyzed during learning, and determine whether the input data at the time of inference belongs to the one large cluster, thereby determining whether the input data at the time of inference is data located in the middle of the input data at the time of learning, or is data in an area outside the range of the input data at the time of learning. Specifically, when the neighborhood direction determination means 8 determines that the input data at the time of inference does not belong to the one large cluster, it may determine that the input data at the time of inference is data in an area outside the range of the input data at the time of learning.

When the neighborhood direction determination means 8 determines that the input data at the time of inference is data in an area outside the range of the input data at the time of learning, the evaluation value determination means 7 determines and outputs the result of increasing the evaluation value according to the distance between the input data and the nearest cluster of the input data as a reliability evaluation value. In other words, when the neighborhood direction determination means 8 determines that the input data at the time of inference is data in an area outside the range of the input data at the time of learning, the evaluation value determination means 7 determines a reliability evaluation value that evaluates the reliability lower than when the input data at the time of inference is data located in the middle of the input data at the time of learning. In other words, the evaluation value determination means 7 corrects the evaluation value in a direction that worsens the evaluation value. Furthermore, the evaluation value determination means 7 corrects the evaluation value so as to evaluate the reliability lower.

When the neighborhood direction determination means 8 determines that the input data at the time of inference is not data in a region outside the range of the input data at the time of learning, the evaluation value determination means 7 determines the evaluation value as the reliability evaluation value and outputs it without modifying the evaluation value.

As described above, the reliability evaluation device 1A according to the second embodiment can evaluate the reliability of the inference result with a relatively high degree of accuracy by evaluating the reliability as low when the input data at the time of inference is in a region outside the range of the input data at the time of learning.

Embodiment 3.
FIG. 5 is a diagram showing the configuration of the reliability evaluation device 1B according to the third embodiment at the time of inference. The configuration of the reliability evaluation device 1B differs from the configuration of the reliability evaluation device 1A according to the second embodiment only at the time of inference. In the third embodiment, the configuration and operation at the time of inference that differ from the second embodiment will be mainly described. The reliability evaluation device 1B has all the components that the reliability evaluation device 1A has. The reliability evaluation device 1B further has a characteristic variation determination means 9 that determines whether or not there is a characteristic variation in the controlled object. For example, the controlled object is an industrial robot, a machine tool, or an industrial machine. The evaluation value determination means 7 may modify the reliability evaluation value based on the determination result obtained by the characteristic variation determination means 9.

As described above, the characteristic variation determination means 9 determines whether or not there is a characteristic variation in the controlled object. For example, the characteristic variation determination means 9 determines whether or not there is a characteristic variation in the controlled object based on the friction torque seen from the motor of each axis that drives the controlled object, or based on the resonant frequency of the controlled object. The characteristic variation determination means 9 can determine whether or not there is a characteristic variation based on input data, and can also determine whether or not there is a characteristic variation based on environmental variation information.

When determining whether or not there is a characteristic variation based on input data, the characteristic variation determination means 9 determines whether or not there is a characteristic variation based on time series data included in the input data. For example, the characteristic variation determination means 9 identifies the friction coefficient of each axis that drives the controlled object based on the position and current of the motor of each axis that drives the controlled object. If the identified friction coefficient is outside a predetermined range, the characteristic variation determination means 9 determines that there is a characteristic variation, and outputs the determination result and the difference between the identified friction coefficient and the predetermined range to the evaluation value determination means 7. The range is defined by a threshold value.

If the characteristic variation determination means 9 determines that there is no characteristic variation, the evaluation value determination means 7 outputs the evaluation value without modifying it. If the characteristic variation determination means 9 determines that there is a characteristic variation, the evaluation value determination means 7 determines a reliability evaluation value that evaluates the reliability lower than when there is no characteristic variation. For example, the evaluation value determination means 7 determines the reliability evaluation value as a result of increasing the evaluation value according to the difference between the friction coefficient identified by the characteristic variation determination means 9 and a predetermined range. In other words, if it is determined that there is a characteristic variation, the evaluation value determination means 7 determines and outputs the result of modifying the evaluation value in a direction that worsens the evaluation value as the reliability evaluation value.

As described above, the characteristic variation determination means 9 can also determine the presence or absence of characteristic variation based on environmental variation information. For example, the environmental variation information is information indicating temperature variation measured by a temperature sensor attached to the controlled object. It is relatively common for the temperature sensor to be attached to the encoder of the motor that drives the controlled object, and in this case, the characteristic variation determination means 9 uses information indicating temperature variation measured by the temperature sensor as the environmental variation information.

For example, a temperature range is set in advance, and the characteristic variation determination means 9 determines that there is a characteristic variation when the temperature measured by the temperature sensor changes from inside to outside the range, and outputs the difference between the temperature measured by the temperature sensor and the range to the evaluation value determination means 7. The range is defined by a threshold value. If the characteristic variation determination means 9 determines that there is no characteristic variation, the evaluation value determination means 7 outputs the evaluation value without modifying it.

When the characteristic variation determination means 9 determines that there is a characteristic variation, the evaluation value determination means 7 determines a reliability evaluation value that evaluates the reliability lower than when there is no characteristic variation. For example, the evaluation value determination means 7 outputs the result of increasing the evaluation value according to the difference between the friction coefficient identified by the characteristic variation determination means 9 and a predetermined range as the reliability evaluation value. In other words, when it is determined that there is a characteristic variation, the evaluation value determination means 7 corrects the evaluation value in a direction that worsens the evaluation value.

As described above, the reliability evaluation device 1B according to the third embodiment can correct the reliability evaluation value depending on whether or not there is a change in the characteristics of the controlled object, and therefore can evaluate the reliability of the inference result with a relatively high degree of accuracy.

Embodiment 4.
6 is a diagram showing the configuration of the reliability evaluation device 1C according to the fourth embodiment at the time of inference. The configuration of the reliability evaluation device 1C differs from the configuration of the reliability evaluation device 1B according to the third embodiment only at the time of inference. In the fourth embodiment, the configuration and operation at the time of inference that differ from the third embodiment will be mainly described. The reliability evaluation device 1C has all the components that the reliability evaluation device 1B has. The reliability evaluation device 1C further has a filter means 10 and a disturbance intrusion determination means 11. The evaluation value determination means 7 may modify the evaluation value based on the determination result obtained by the disturbance intrusion determination means 11.

For example, the filter means 10 is a high-pass filter or a band-pass filter. For example, when a disturbance with a relatively high frequency may act on the input data, a high-pass filter is used as the filter means 10. An example of a disturbance is noise. When it is expected that the drive mechanism of the controlled object will deteriorate and the disturbance according to the rotation speed of the drive mechanism will become large, a band-pass filter is used as the filter means 10. For example, the deterioration of the drive mechanism is deterioration of the bearings built into the drive mechanism.

The disturbance detection means 11 determines whether a disturbance equal to or greater than a predetermined threshold has acted. Specifically, the disturbance detection means 11 determines whether a disturbance equal to or greater than a reference value has been mixed into the input data based on the time series data output from the filter means 10. For example, the disturbance detection means 11 calculates the absolute value of the output from the filter means 10, passes the calculated absolute value through a moving average filter, compares the absolute value after passing through the moving average filter with a pre-stored threshold, and determines that a disturbance equal to or greater than a reference value has been mixed into the input data if it is determined that the absolute value after passing through the moving average filter is equal to or greater than the threshold.

When the disturbance detection means 11 determines that a disturbance equal to or greater than the reference value has been detected, it outputs the detection result and the amount of excess from the threshold to the evaluation value determination means 7. When the disturbance detection means 11 determines that a disturbance equal to or greater than the reference value has been detected, the evaluation value determination means 7 corrects the evaluation value in a direction that increases the evaluation value according to the amount of excess from the threshold. In other words, the evaluation value determination means 7 corrects the evaluation value in a direction that deteriorates the evaluation value. The evaluation value determination means 7 outputs the corrected evaluation value as the reliability evaluation value. In this way, when the disturbance detection means 11 determines that a disturbance equal to or greater than the threshold has acted, the evaluation value determination means 7 determines a reliability evaluation value that evaluates reliability lower than when no disturbance equal to or greater than the threshold has acted.

As described above, the reliability evaluation device 1C according to the fourth embodiment can revise the reliability evaluation value in consideration of a disturbance equal to or greater than a threshold value that acts during inference. Therefore, even if a disturbance equal to or greater than a threshold value acts during inference, the reliability of the inference result can be evaluated with relatively high accuracy.

Embodiment 5.
7 is a diagram showing a configuration during learning of a reliability evaluation device 1D according to the fifth embodiment. The reliability evaluation device 1D has the first learning means 2, evaluation result storage means 3, and second learning means 4 that the reliability evaluation device 1 according to the first embodiment has. Data input to the second learning means 4 of the fifth embodiment is different from the data input to the second learning means 4 of the first embodiment. In the fifth embodiment, the configuration and operation during learning that differ from those of the first embodiment will be mainly described.

In the first embodiment, the input of the second learning means 4 is the same as the input of the first learning means 2, that is, the input data itself is used as information related to the input data, but in the fifth embodiment, the information related to the input data that is the input of the second learning means 4 is operation condition data. The operation condition data is data related to the input data. As in the first embodiment, in the fifth embodiment, the input data is, for example, data indicating the position, speed, acceleration, or current of the motor of each axis that drives the control target, or the output of a sensor attached to the control target. The operation condition data is operation command data when collecting the input data. The operation command data may be time-series data of the motor position command, speed command, or acceleration command of each axis that drives the control target, or may be data indicating the maximum value of the acceleration or speed in the motor position command of each axis that drives the control target.

The second learning means 4 of the fifth embodiment performs clustering on the operating condition data. When time-series data of a position command, a speed command, or an acceleration command is input as the operating condition data, the second learning means 4 of the fifth embodiment may use the k-means method after extracting features such as the maximum or average value of the time-series data by the clustering means 41, as in the second learning means 4 of the first embodiment.

The second learning means 4 of the fifth embodiment may perform clustering using, for example, the kSHAPE method without extracting features. When data indicating the maximum value of acceleration or speed in the motor position command is input to the second learning means 4, the features are input directly, so the second learning means 4 uses the k-means method. The clustering results obtained by the second learning means 4 are output to the evaluation result storage means 3 together with the evaluation results of the learning performed by the first learning means 2. The clustering results obtained by the second learning means 4 are information related to the clustering performed by the second learning means 4.

The evaluation result storage means 3 stores the evaluation result of the learning performed by the first learning means 2 and information about the clustering performed by the second learning means 4. The evaluation result storage means 3 stores the evaluation result for the input data closest to the center of the cluster as the evaluation result corresponding to the cluster. The evaluation result storage means 3 may store the average value of multiple evaluation results close to the center of the cluster as the evaluation result corresponding to the cluster, may store the worst value of all evaluation results of the input data belonging to the cluster, or may store the average value and variance of all evaluation results of the input data belonging to the cluster. The input data being closest to the center of the cluster or close to the cluster center is determined from the distance between the input data and the cluster center. If the k-means method is used after extracting the features, the determination may be made from the distance between the features of the cluster center and the features of the input. In this case, the distance between the input data and the cluster center is the square root of the sum of the squares of the differences between the elements of the feature vector of the input data and the feature vector of the cluster center when the features are expressed as vectors. If the time series data is used as is, the determination may be made using shape-based distance (SBD), DTW (Dynamic Time Warping), or the like.

Figure 8 is a diagram showing the configuration of the reliability evaluation device 1D according to the fifth embodiment at the time of inference. The reliability evaluation device 1D has the neighborhood group extraction means 5, evaluation value estimation means 6, and evaluation value determination means 7 of the reliability evaluation device 1 according to the first embodiment. The input data at the time of inference input to the neighborhood group extraction means 5 of the fifth embodiment is operating condition data. For example, the operating condition data at the time of inference is time-series data of the motor position command, speed command, or acceleration command of each axis that drives the control object, or data indicating the maximum value of the acceleration or speed in the motor position command of each axis that drives the control object.

The neighborhood group extraction means 5 extracts multiple clusters by clustering the input data at the time of inference based on the clustering results performed by the second learning means 4. For example, the neighborhood group extraction means 5 extracts multiple clusters whose centers are in the vicinity of the clustering results for the operating condition data. The neighborhood is determined from the distance between the input data and the cluster center. If the k-means method is used after the feature amount is extracted, the determination can be made from the distance between the feature amount of the cluster center and the feature amount of the input. If the time series data is used as is, the determination can be made using shape-based distance (SBD), DTW (Dynamic Time Warping), or the like. The evaluation value estimation means 6 reads out the evaluation results for each of the multiple clusters extracted by the neighborhood group extraction means 5 from the evaluation result storage means 3, and outputs the read out evaluation results and information on the multiple extracted clusters to the evaluation value determination means 7.

The evaluation value determination means 7 determines a reliability evaluation value that evaluates the reliability of the learning performed by the first learning means 2 based on the evaluation results stored by the evaluation result storage means 3 and the information on the multiple clusters extracted by the neighborhood group extraction means 5.

As described above, the reliability evaluation device 1D according to the fifth embodiment determines the reliability evaluation value using the operating condition data, and therefore can evaluate the reliability of the inference result with a relatively small amount of calculation.

Embodiment 6.
FIG. 9 is a diagram showing a configuration of a reliability evaluation device 1E according to the sixth embodiment during learning. The reliability evaluation device 1E has the first learning means 2 and the evaluation result storage means 3 of the reliability evaluation device 1 according to the first embodiment. The reliability evaluation device 1E has a second learning means 4E instead of the second learning means 4 of the reliability evaluation device 1. The second learning means 4E has a class classification means 42 that performs class classification on information related to input data. In this embodiment, the input data itself is used as information related to the input data. In addition, not only the input data but also the class correct data of the input data is input to the second learning means 4E. The evaluation result storage means 3 is input with the class correct data input to the second learning means 4E, not with the output of the second learning means 4E. In the sixth embodiment, the configuration and operation during learning that are different from those of the first embodiment will be mainly described.

For example, the classifying means 42 uses a support vector machine or a neural network. Examples of neural networks are a feedforward neural network or a recurrent neural network. When the classifying means 42 uses a neural network, when there are N types of classes, the number of nodes in the output layer of the neural network is set to N, and only one node outputs 1, while the remaining nodes output 0. N is an integer equal to or greater than 2. The classifying means 42 performs learning by assigning which node's output should be set to 1 for each correct class. For example, the classifying means 42 sets the output of the first node to 1 for the first class, and the output of the third node to 1 for the third class.

The evaluation result storage means 3 receives the evaluation results of the learning performed by the first learning means 2 and the class correct answer data. The evaluation result storage means 3 stores the evaluation results for each class indicated by the class correct answer data. For example, the evaluation results stored in the evaluation result storage means 3 are the average value of the evaluation results for each class, the worst value of the evaluation results for each class, or the average value and variance of the evaluation results for each class.

The configuration of the reliability evaluation device 1E according to the sixth embodiment at the time of inference is the same as the configuration of the reliability evaluation device 1 according to the first embodiment at the time of inference. However, while in the first embodiment the neighborhood group extraction means 5 has the clustering means 41 of the second learning means 4, in the sixth embodiment the neighborhood group extraction means 5 has the class classification means 42 of the second learning means 4E.

In the sixth embodiment, the neighborhood group extraction means 5 extracts a plurality of classes by classifying the input data at the time of inference based on the result of the class classification performed by the second learning means 4E. Specifically, when the input data is input, the neighborhood group extraction means 5 performs the inference learned by the class classification means 42 and outputs the class closest to the inference result and the distance from the class. In the sixth embodiment, the closest class is the class corresponding to the node closest to 1 among the outputs of each node, and the distance from the class is the difference between the value of the node closest to 1 and 1. For example, if the number of nodes in the output layer is 5 and the outputs of each node are 0, 0.9, 0.1, 0.3, and 0.7, respectively, the closest class is the class to which the output of the second node is assigned to 1, and the distance from the class is 0.1. The neighborhood group extraction means 5 may output not only the distance of the class closest to the inference result, but also the class second or third closest to the inference result and the distance to the class.

The evaluation value estimation means 6 reads out the evaluation results for each of the multiple classes extracted by the neighborhood group extraction means 5 from the evaluation result storage means 3, and outputs the read out evaluation results and information on the multiple extracted classes to the evaluation value determination means 7. When multiple classes are extracted by the neighborhood group extraction means 5, the evaluation value determination means 7 calculates a weighted sum of the evaluation results according to the distance to the class, and determines and outputs the weighted sum as a reliability evaluation value.

The evaluation value determination means 7 may adopt the evaluation value of the closest class, which is the class corresponding to the node closest to 1 among the outputs of each node, or may adopt the worst value among the extracted classes. When the read evaluation results are the average value and variance for each class, the evaluation value determination means 7 may output the average value of the average value and variance, or the worst value of the average value and variance, as the reliability evaluation value. When the distance to the closest class is within a predetermined threshold, the evaluation value determination means 7 may adopt the evaluation value of the class, and when the distance to the closest class exceeds the predetermined threshold, the evaluation value determination means 7 may output a result in which the evaluation value of the class is increased according to the distance exceeding the threshold. In other words, when the distance to the closest class exceeds the threshold, the evaluation value determination means 7 may output a result in which the evaluation value is corrected in a direction that worsens the evaluation value.

As described above, the reliability evaluation device 1E according to the sixth embodiment evaluates the reliability of the inference result according to the domain to which the input data at the time of inference belongs. The inference result is the output of the first learning means 2. More specifically, the inference result is the result of the learning performed by the first learning means 2. Therefore, the reliability evaluation device 1E can evaluate the reliability of the inference result with a relatively high degree of accuracy. In other words, the reliability evaluation device 1E can appropriately evaluate the reliability of the inference when the learning result performed by the first learning means 2 is used.

Embodiment 7.
10 is a diagram showing a configuration during learning of a reliability evaluation apparatus 1F according to the seventh embodiment. The reliability evaluation apparatus 1F has the evaluation result storage means 3 and the second learning means 4 of the reliability evaluation apparatus 1 according to the first embodiment. The reliability evaluation apparatus 1F has a first learning means 2F instead of the first learning means 2 of the reliability evaluation apparatus 1. In the seventh embodiment, differences from the first embodiment will be mainly described.

The first learning means 2 of the reliability evaluation device 1 according to the first embodiment has a first learning device 21, and learns the relationship between the input data and teacher data. The first learning means 2F of the reliability evaluation device 1F according to the seventh embodiment does not learn the relationship between the input data and teacher data, but has a first learning device 21F that is an optimal parameter search means that receives input data and a target specification and searches for control parameters when the input data is input according to the target specification. That is, the first learning device 21F searches for optimal values of the control parameters of the control target for the input data. Target specifications include shortened operating time, improved trajectory accuracy, and minimal energy consumption.

The first learning means 2F outputs an evaluation result that is not the difference between the teacher data and the output of the first learning device 21F, but an evaluation value of the control result of the control object when the parameters searched by the first learning device 21F are used. For example, the evaluation index of the control result is the settling time, the amount of overshoot, or the maximum tip trajectory error. The first learning means 2F may output the evaluation index of the control result itself as the evaluation result, or may output the difference between the evaluation index and the target value as the evaluation result. In the seventh embodiment, the first learning means 2F outputs the difference between the evaluation index and the target value as the evaluation result. Therefore, the smaller the value of the evaluation result, the higher the control performance can be expected.

For example, the learning method used by the first learner 21F is reinforcement learning, Bayesian optimization, or particle swarm optimization. For example, the control parameters searched for by the first learner 21F are operation command parameters such as the length of the acceleration/deceleration section or the maximum acceleration value, or feedforward control parameters or feedback control parameters of the motor control device of each axis.

The configuration of the reliability evaluation device 1F according to the seventh embodiment at the time of inference is the same as the configuration of the reliability evaluation device 1 according to the first embodiment shown in FIG. 3 at the time of inference. However, in the seventh embodiment, the first learning device 21 in FIG. 3 is replaced with the first learning device 21F. When input data at the time of inference is input to the first learning device 21F, the first learning device 21F searches for control parameters in the same way as during learning, and outputs the control parameters obtained as a result of the search as the inference result.

The reliability evaluation of the control parameters searched by the first learning device 21F is based on the evaluation results based on the control parameters searched during learning. That is, the evaluation result storage means stores values based on the evaluation results for the control parameters searched during learning, rather than values based on the difference from the teacher data during learning. Except for the fact that the reliability evaluation of the control parameters searched by the first learning device 21F is based on the evaluation results based on the control parameters searched during learning, the operation of the reliability evaluation device 1F during inference is the same as the operation of the reliability evaluation device 1 according to embodiment 1 during inference.

As described above, the reliability evaluation device 1F according to the seventh embodiment evaluates the reliability of the inference result based on the control parameters at the time when the input data is input. Therefore, the reliability evaluation device 1F can evaluate the reliability of the inference result with a relatively high degree of accuracy. In other words, the reliability evaluation device 1F can appropriately evaluate the reliability of the learning performed by the first learning means 2F.

FIG. 11 is a diagram showing a processor 91 in the case where at least some of the first learning means 2, second learning means 4, neighborhood group extraction means 5, evaluation value estimation means 6, and evaluation value determination means 7 of the reliability evaluation device 1 according to the first embodiment are realized by the processor 91. In other words, at least some of the functions of the first learning means 2, second learning means 4, neighborhood group extraction means 5, evaluation value estimation means 6, and evaluation value determination means 7 may be realized by the processor 91 executing a program stored in the memory 92. The processor 91 is a CPU (Central Processing Unit), a processing device, an arithmetic device, a microprocessor, or a DSP (Digital Signal Processor). The memory 92 is also shown in FIG. 11.

When at least some of the functions of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 are realized by the processor 91, the at least some of the functions are realized by the processor 91 and software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in the memory 92. The processor 91 realizes at least some of the functions of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 by reading and executing the program stored in the memory 92.

When at least some of the functions of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 are realized by the processor 91, the reliability evaluation device 1 has a memory 92 for storing a program that results in the execution of at least some of the steps executed by the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7. It can also be said that the program stored in the memory 92 causes a computer to execute at least some of the functions of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7.

The memory 92 may be, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable Read Only Memory), or an EEPROM (registered trademark) (Electrically Erasable Programmable Read-Only Memory), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, or a DVD (Digital Versatile Disk).

FIG. 12 is a diagram showing a processing circuit 93 in the case where at least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 of the reliability evaluation device 1 according to the first embodiment is realized by the processing circuit 93. In other words, at least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 may be realized by the processing circuit 93. The processing circuit 93 is dedicated hardware. The processing circuit 93 is, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination of these.

The first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and a part of the evaluation value determination means 7 may be realized by dedicated hardware separate from the remaining parts.

With regard to the multiple functions of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7, some of the multiple functions may be realized by software or firmware, and the remaining multiple functions may be realized by dedicated hardware. In this way, the multiple functions of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 can be realized by hardware, software, firmware, or a combination of these.

At least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, the evaluation value determination means 7, and the neighborhood direction determination means 8 of the reliability evaluation device 1A according to the second embodiment may be realized by a processor. The processor is the same as the processor 91. When at least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, the evaluation value determination means 7, and the neighborhood direction determination means 8 of the reliability evaluation device 1A are realized by a processor, the reliability evaluation device 1A has a memory for storing a program that results in at least a part of the steps executed by the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, the evaluation value determination means 7, and the neighborhood direction determination means 8 being executed. The memory is the same as the memory 92. At least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, the evaluation value determination means 7, and the neighborhood direction determination means 8 of the reliability evaluation device 1A may be realized by a processing circuit. The processing circuit in question is the same as processing circuit 93.

At least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, the evaluation value determination means 7, the neighborhood direction determination means 8, and the characteristic variation determination means 9 of the reliability evaluation device 1B according to the third embodiment may be realized by a processor. The processor is the same as the processor 91. When at least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, the evaluation value determination means 7, the neighborhood direction determination means 8, and the characteristic variation determination means 9 of the reliability evaluation device 1B are realized by a processor, the reliability evaluation device 1B has a memory for storing a program that results in at least a part of the steps executed by the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, the evaluation value determination means 7, the neighborhood direction determination means 8, and the characteristic variation determination means 9 being executed. The memory is the same as the memory 92. At least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, the evaluation value determination means 7, the neighborhood direction determination means 8, and the characteristic variation determination means 9 of the reliability evaluation device 1B may be realized by a processing circuit. The processing circuit is a processing circuit similar to the processing circuit 93.

At least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, the evaluation value determination means 7, the neighborhood direction determination means 8, the characteristic variation determination means 9, the filter means 10, and the disturbance mixing determination means 11 of the reliability evaluation device 1C according to the fourth embodiment may be realized by a processor. The processor is the same as the processor 91. When at least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, the evaluation value determination means 7, the neighborhood direction determination means 8, the characteristic variation determination means 9, the filter means 10, and the disturbance mixing determination means 11 of the reliability evaluation device 1C are realized by a processor, the reliability evaluation device 1C has a memory for storing a program that results in at least a part of the steps executed by the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, the evaluation value determination means 7, the neighborhood direction determination means 8, the characteristic variation determination means 9, the filter means 10, and the disturbance mixing determination means 11 being executed. The memory is the same as the memory 92. At least a part of the first learning means 2, second learning means 4, neighborhood group extraction means 5, evaluation value estimation means 6, evaluation value determination means 7, neighborhood direction determination means 8, characteristic fluctuation determination means 9, filter means 10, and disturbance intrusion determination means 11 of the reliability evaluation device 1C may be realized by a processing circuit. The processing circuit is the same as the processing circuit 93.

At least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 of the reliability evaluation device 1D according to the fifth embodiment may be realized by a processor. The processor is the same as the processor 91. When at least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 of the reliability evaluation device 1D are realized by a processor, the reliability evaluation device 1D has a memory for storing a program that results in at least a part of the steps executed by the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 being executed. The memory is the same as the memory 92. At least a part of the first learning means 2, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 of the reliability evaluation device 1D may be realized by a processing circuit. The processing circuit is the same as the processing circuit 93.

At least a part of the first learning means 2, the second learning means 4E, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 of the reliability evaluation device 1E according to the sixth embodiment may be realized by a processor. The processor is the same as the processor 91. When at least a part of the first learning means 2, the second learning means 4E, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 of the reliability evaluation device 1E are realized by a processor, the reliability evaluation device 1E has a memory for storing a program that results in at least a part of the steps executed by the first learning means 2, the second learning means 4E, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 being executed. The memory is the same as the memory 92. At least a part of the first learning means 2, the second learning means 4E, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 of the reliability evaluation device 1E may be realized by a processing circuit. The processing circuit is the same as the processing circuit 93.

At least a part of the first learning means 2F, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 of the reliability evaluation device 1F according to the seventh embodiment may be realized by a processor. The processor is the same as the processor 91. When at least a part of the first learning means 2F, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 of the reliability evaluation device 1F are realized by a processor, the reliability evaluation device 1F has a memory for storing a program that results in at least a part of the steps executed by the first learning means 2F, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 of the reliability evaluation device 1F. The memory is the same as the memory 92. At least a part of the first learning means 2F, the second learning means 4, the neighborhood group extraction means 5, the evaluation value estimation means 6, and the evaluation value determination means 7 of the reliability evaluation device 1F may be realized by a processing circuit. The processing circuit in question is the same as processing circuit 93.

The configurations shown in the above embodiments are merely examples, and may be combined with other known technologies, or the embodiments may be combined with each other. In addition, parts of the configurations may be omitted or modified without departing from the spirit of the invention.

1, 1A, 1B, 1C, 1D, 1E, 1F reliability evaluation device, 2, 2F first learning means, 3 evaluation result storage means, 4, 4E second learning means, 5 neighborhood group extraction means, 6 evaluation value estimation means, 7 evaluation value determination means, 8 neighborhood direction determination means, 9 characteristic fluctuation determination means, 10 filter means, 11 disturbance inclusion determination means, 21, 21F first learning device, 22 learning result evaluation means, 41 clustering means, 42 classifying means, 91 processor, 92 memory, 93 processing circuit.

Claims (8)

A first learning means for learning by using input data and correct answer data or target specifications as input;
A second learning means for performing clustering or class classification on information regarding the input data;
an evaluation result storage means for storing an evaluation result of the learning performed by the first learning means;
a neighborhood group extraction means for extracting a plurality of clusters or classes by performing clustering or class classification on input data at the time of inference based on a result of the clustering or class classification performed by the second learning means;
and an evaluation value determination means for determining a reliability evaluation value for evaluating the reliability of the inference made by the first learning means, based on the evaluation results stored by the evaluation result storage means and information on the plurality of clusters or classes extracted by the neighborhood group extraction means. a neighborhood direction determining means for determining whether the input data at the time of inference is data located in the middle of the input data at the time of learning, or data in a region outside the range of the input data at the time of learning,
2. The reliability evaluation device according to claim 1, wherein, when the input data at the time of inference is determined by the neighborhood direction determination means to be data in an area outside the range of the input data at the time of learning, the evaluation value determination means determines the reliability evaluation value that evaluates the reliability lower compared to a case in which the input data at the time of inference is data located in the middle of the input data at the time of learning. The control system further includes a characteristic fluctuation determination means for determining whether or not there is a characteristic fluctuation of the controlled object,
3. The reliability evaluation device according to claim 1, wherein, when the characteristic variation determination means determines that the characteristic variation exists, the evaluation value determination means determines the reliability evaluation value that evaluates the reliability lower than when the characteristic variation does not exist. The device further includes a disturbance detection means for detecting whether or not a disturbance equal to or greater than a predetermined threshold value has occurred.
4. The reliability evaluation device according to claim 1, wherein, when the disturbance detection means detects that a disturbance equal to or greater than the threshold has acted, the evaluation value determination means determines the reliability evaluation value that evaluates the reliability lower than when no disturbance equal to or greater than the threshold has acted. The reliability evaluation device according to any one of claims 1 to 4, characterized in that the first learning means inputs input data and correct answer data, and learns the relationship between the input data and the correct answer data. The reliability evaluation device according to any one of claims 1 to 4, characterized in that the input data itself is used as information about the input data to be input to the second learning means. The reliability evaluation device according to any one of claims 1 to 4, characterized in that operating condition data is used as information regarding the input data to be input to the second learning means. The reliability evaluation device according to any one of claims 1 to 4, characterized in that the first learning means has an optimal parameter search means that inputs input data and target specifications and searches for control parameters when the input data is input.

Images