JP7265127B2 - Machine learning program, machine learning method and machine learning system - Google Patents

Description

The present invention relates to a machine learning program, a machine learning method and a machine learning system.

For example, a business operator that provides services to users (hereinafter simply referred to as a business operator) operates a business system that performs processes necessary for providing services. Specifically, for example, a business operator builds a business system that extracts sensor data indicating the occurrence of an abnormal state (hereinafter also referred to as abnormal data) from sensor data output from a large number of sensors.

In such a business system, for example, when new sensor data output from a sensor is received, the received new sensor data is input to a supervised learning model that has learned a large number of sensor data in advance. The business system then estimates whether or not the new sensor data is abnormal data based on the data output from the supervised learning model.

Here, when generating a learning model as described above, the operator needs to prepare in advance a large number of abnormal data to be used as teacher data. Therefore, for example, when the output frequency of abnormal data from each sensor is low, the operator cannot prepare the necessary number of abnormal data in advance, and cannot generate a supervised learning model with high recognition accuracy. There is

Therefore, the business system generates a deep unsupervised learning model by, for example, learning only sensor data indicating a normal state (hereinafter also referred to as normal data). Then, for example, when receiving new sensor data, the business system restores the received sensor data by inputting the received sensor data into the deep unsupervised learning model. After that, the business system estimates the abnormal state of the received sensor data by comparing the received sensor data and the restored sensor data (hereinafter also referred to as restored sensor data).

As a result, even if the operator cannot prepare a sufficient number of abnormal data, it is possible to estimate the abnormal state of new sensor data (see, for example, Non-Patent Document 1). .

O. Lyudchik, “Outlier detection using autoencoders”, 2016

However, the above method extracts new sensor data as abnormal data when the features of new sensor data stochastically deviate from the features of learned normal data. Therefore, in the above method, even sensor data that should not be extracted as abnormal data may be extracted as abnormal data.

Therefore, in one aspect, the present invention provides a machine learning program that enables accurate extraction of abnormal data to be extracted even when the number of abnormal data that can be used as teacher data is small. An object is to provide a learning method and a machine learning system.

In one aspect of the embodiment, the first model is generated by performing deep unsupervised learning of a plurality of sensor data output from a plurality of sensors, and for each of the plurality of sensor data, each sensor data and each From the result of comparison with the restored sensor data output from the first model along with the input of the sensor data, a feature amount corresponding to each sensor data is extracted, and for each of the extracted feature amounts, each feature amount and each A computer is caused to execute a process of generating a second model by performing supervised learning of data in which sensor data corresponding to the feature quantity is associated with state information indicating whether or not it corresponds to an abnormal state.

According to one aspect, even when the number of abnormal data that can be used as teacher data is small, it is possible to accurately extract abnormal data to be extracted.

FIG. 1 is a diagram illustrating the configuration of an information processing system 10. As shown in FIG. FIG. 2 is a diagram for explaining the hardware configuration of the information processing device 1. As shown in FIG. FIG. 3 is a functional block diagram of the information processing apparatus 1. As shown in FIG. FIG. 4 is a flowchart for explaining an outline of machine learning processing according to the first embodiment. FIG. 5 is a flowchart illustrating details of machine learning processing according to the first embodiment. FIG. 6 is a flowchart illustrating details of machine learning processing according to the first embodiment. FIG. 7 is a flowchart illustrating details of machine learning processing according to the first embodiment. FIG. 8 is a flowchart illustrating details of machine learning processing according to the first embodiment. FIG. 9 is a flowchart illustrating details of machine learning processing according to the first embodiment. FIG. 10 is a flowchart illustrating details of machine learning processing according to the first embodiment. FIG. 11 is a flowchart illustrating details of machine learning processing according to the first embodiment. FIG. 12 is a diagram illustrating details of machine learning processing according to the first embodiment. FIG. 13 is a diagram illustrating details of machine learning processing according to the first embodiment. FIG. 14 is a diagram explaining a specific example of the sensor data 131. As shown in FIG. FIG. 15 is a diagram illustrating a specific example of the restored sensor data 131a. FIG. 16 is a diagram illustrating a specific example of the difference sensor data 131b. FIG. 17 is a diagram illustrating a specific example of the difference sensor data 131b. FIG. 18 is a diagram illustrating a specific example of the process of S42. FIG. 19 is a diagram illustrating a specific example of the process of S43. FIG. 20 is a flow chart for explaining machine learning processing and abnormality estimation processing in the second embodiment. FIG. 21 is a flow chart for explaining machine learning processing and abnormality estimation processing in the second embodiment. FIG. 22 is a flowchart for explaining machine learning processing and abnormality estimation processing in the second embodiment. FIG. 23 is a flowchart for explaining machine learning processing and abnormality estimation processing in the second embodiment. FIG. 24 is a flowchart for explaining machine learning processing and abnormality estimation processing in the second embodiment. FIG. 25 is a diagram for explaining machine learning processing in the second embodiment. FIG. 26 is a diagram for explaining machine learning processing in the second embodiment.

[Configuration of information processing system]
First, the configuration of the information processing system 10 will be described. FIG. 1 is a diagram illustrating the configuration of an information processing system 10. As shown in FIG.

As shown in FIG. 1, an information processing system 10 includes, for example, an information processing device 1 comprising one or more physical machines, and a plurality of sensors 5 that transmit detected sensor data to the information processing device 1. .

The sensor 5 is, for example, a wearable sensor worn by a worker working in a factory, and is connected to the information processing device 1 via the network NW. When the sensor 5 detects, for example, acceleration data, sound pressure data, atmospheric pressure data, illuminance data, or the like, the sensor 5 transmits these data to the information processing apparatus 1 as sensor data. Specifically, the sensor 5 transmits each sensor data to the information processing device 1 by performing wireless communication such as Bluetooth (registered trademark) or Wifi.

The information processing apparatus 1 generates a first model, which is a deep unsupervised learning model, by performing deep unsupervised learning on a plurality of sensor data output from the plurality of sensors 5, for example. Then, the information processing device 1 compares each sensor data with the sensor data output from the first model (hereinafter also referred to as restored sensor data) accompanying the input of each sensor data for each of the plurality of sensor data. , a feature quantity corresponding to each sensor data is extracted. Specifically, the information processing apparatus 1 extracts the feature amount by a method capable of incorporating necessary domain knowledge from the comparison result between the sensor data and the restored sensor data.

After that, the information processing apparatus 1 supervises data in which each feature amount is associated with state information indicating whether or not sensor data corresponding to each feature amount corresponds to an abnormal state for each extracted feature amount. A second model, which is a supervised learning model, is generated by learning.

As a result, the information processing apparatus 1 uses a learning model generated from teacher data including feature quantities in which domain knowledge is incorporated, thereby preventing sensor data that should not be extracted as abnormal data from being extracted as abnormal data. suppression becomes possible. Therefore, the information processing apparatus 1 can accurately determine whether the sensor data output by the sensor 5 is abnormal data.

[Hardware configuration of information processing system]
Next, the hardware configuration of the information processing system 10 will be described. FIG. 2 is a diagram for explaining the hardware configuration of the information processing device 1. As shown in FIG.

The information processing apparatus 1 has a CPU 101 as a processor, a memory 102, an external interface (I/O unit) 103, and a storage medium 104, as shown in FIG. Each unit is connected to each other via a bus 105 .

The storage medium 104 has, for example, a program storage area (not shown) that stores a program 110 for performing processing for generating each learning model (hereinafter also referred to as machine learning processing). The storage medium 104 also has a storage unit 130 (hereinafter also referred to as an information storage area 130) that stores information used when performing machine learning processing, for example. Note that the storage medium 104 may be, for example, an HDD (Hard Disk Drive) or an SSD (Sokid State Drive).

The CPU 101 executes a program 110 loaded from the storage medium 104 to the memory 102 to perform machine learning processing.

Also, the external interface 103 communicates with a plurality of sensors 5, for example.

[Functions of information processing system]
Next, functions of the information processing system 10 will be described. FIG. 3 is a functional block diagram of the information processing apparatus 1. As shown in FIG.

As shown in FIG. 3, the information processing apparatus 1 includes an information reception unit 111, an information management unit 112, a second 1 model generation unit 113, information restoration unit 114, first feature extraction unit 115, second feature extraction unit 116, feature combination unit 117, second model generation unit 118, data selection unit 119, model Various functions including a selection unit 120, an output information generation unit 121, and an information output unit 122 are implemented.

For example, as shown in FIG. 3, the information processing apparatus 1 includes sensor data 131, teacher data 132, first model 133, second model 134, first weight information 135, and second weight information. 136 , output information 137 , and model selection information 138 are stored in the information storage area 130 .

First, each function when generating the first model 133 and the second model 134 will be described.

The information reception unit 111 receives sensor data 131 transmitted from the sensor 5, for example. The sensor data 131 here is, for example, the sensor data 131 used for generating the first model 133 and the second model 134 .

The information management unit 112 stores, for example, the sensor data 131 received by the information reception unit 111 in the information storage area 130 .

The information management unit 112 also associates, for example, the sensor data 131 stored in the information storage area 130 with state information (hereinafter also referred to as a label) indicating whether each sensor data 131 corresponds to an abnormal state. The attached teacher data 132 is generated. The information management unit 112 then stores the generated teacher data 132 in the information storage area 130, for example. The label corresponding to each sensor data 131 is, for example, a label indicating that each sensor data 131 is abnormal data or a label indicating that each sensor data 131 is normal data. Also, the label corresponding to each sensor data 131 is determined in advance by an operator who refers to the contents of each sensor data 131, for example.

The first model generation unit 113 generates the first model 133 by performing deep unsupervised learning with the sensor data 131 included in the teacher data 132 stored in the information storage area 130 as input. Specifically, the first model generating unit 113 generates, for example, the sensor data 131 associated with the label indicating normal data among the sensor data 131 included in the teacher data 132 stored in the information storage area 130. The first model 133 is generated by performing deep unsupervised learning with . Then, the information management unit 112 stores, in the information storage area 130, first weight information 135 corresponding to the first model 133 generated by the first model generation unit 113, for example.

The information restoration unit 114 inputs each of the sensor data 131 included in the teacher data 132 stored in the information storage area 130 to the first model 133 generated by the first model generation unit 113 . Then, the information restoring unit 114 acquires sensor data 131a (hereinafter also referred to as restored sensor data 131a) output from the first model 133 in accordance with the input of each sensor data 131 .

For each sensor data 131 included in the teacher data 132 stored in the information storage area 130, the first feature extraction unit 115 compares each sensor data 131 with the restored sensor data 131a corresponding to each sensor data 131. , the first feature quantity corresponding to each sensor data 131 is extracted.

For each sensor data 131 included in the teacher data 132 stored in the information storage area 130, the second feature extraction unit 116 extracts a second feature amount (hereinafter referred to as another feature (also called quantity).

The feature combining unit 117 combines the first feature amount extracted by the first feature extraction unit 115 and the A third feature amount (hereinafter also referred to as a combined feature amount) is generated by combining the second feature amount.

The second model generation unit 118 receives as input data in which each feature amount and the label of the sensor data 131 corresponding to each feature amount are associated with each third feature amount generated by the feature combining unit 117. A second model 134 is generated by learning. Then, the information management unit 112 stores, in the information storage area 130, second weight information 136 corresponding to the second model 134 generated by the second model generation unit 118, for example.

Note that the second model generation unit 118 generates, for example, data in which each feature amount and the label of the sensor data 131 corresponding to each feature amount are associated with each first feature amount extracted by the first feature extraction unit 115. The second model 134 may be generated by performing supervised learning with input.

When generating a plurality of first models 133 , the data selection unit 119 selects the teacher data 132 to be used for generating each of the first models 133 from the teacher data 132 stored in the information storage area 130 .

When the plurality of first models 133 are generated, the model selection unit 120 selects the first model 133 that satisfies a predetermined condition from the plurality of generated first models 133 . Specifically, the model selection unit 120 selects, for example, the first model 133 that can accurately determine whether each sensor data 131 is abnormal data. Then, the information management unit 112 stores, in the information storage area 130, model selection information 138, which is information indicating the first model 133 selected by the model selection unit 120, for example.

Next, each function after the generation of the first model 133 and the second model 134 (each function when using the first model 133 and the second model 134) will be described.

The information reception unit 111 receives new sensor data 131 transmitted from the sensor 5, for example. The new sensor data 131 here is, for example, sensor data 131 for which it is determined whether or not it is abnormal data.

The information restoring unit 114 inputs the new sensor data 131 received by the information receiving unit 111 to the first model 133 generated by the first model generating unit 113 . Then, the information restoration unit 114 acquires new restored sensor data 131 a output from the first model 133 in response to the input of the new sensor data 131 .

The first feature extracting unit 115 compares the new sensor data 131 received by the information receiving unit 111 and the new restored sensor data 131a obtained by the information restoring unit 114, and from the result of comparison, the first feature corresponding to the new sensor data 131 1. Extract feature values.

The second feature extraction unit 116 extracts a second feature amount corresponding to the new sensor data 131 from the new sensor data 131 received by the information receiving unit 111 .

The feature combining unit 117 combines the first feature amount extracted by the first feature extraction unit 115 and the second feature amount extracted by the second feature extraction unit 116 to generate a third feature amount.

The output information generator 121 inputs the third feature amount generated by the feature combiner 117 to the second model 134 generated by the second model generator 118 . Then, the output information generation unit 121 acquires the score output from the second model 134 along with the input of the third feature amount.

Note that the output information generation unit 121 obtains a score by, for example, inputting the first feature amount extracted by the first feature extraction unit 115 to the second model 134 generated by the second model generation unit 118. There may be.

The information output unit 122 outputs the score acquired by the output information generation unit 121 as output information 137 . Specifically, the information output unit 122 outputs, for example, the output information 137 to the operator's operation terminal (not shown).

[Outline of the first embodiment]
Next, an outline of the first embodiment will be described. FIG. 4 is a flowchart for explaining an outline of machine learning processing according to the first embodiment.

As shown in FIG. 4, the information processing apparatus 1 waits until the machine learning execution timing (NO in S1). The machine learning execution timing may be, for example, the timing at which the operator inputs information to the information processing apparatus 1 to start the machine learning process.

Then, when it is time to execute machine learning (YES in S1), the information processing device 1 performs deep unsupervised learning on the plurality of sensor data 131 output from the plurality of sensors 5, thereby obtaining the first model 133. Generate (S2).

Subsequently, the information processing apparatus 1 generates the restored sensor data 131a output from the first model 133 generated in the process of S2 according to the input of each sensor data 131 and the sensor data 131 for each of the plurality of sensor data 131. From the comparison result, a feature quantity (for example, a first feature quantity) corresponding to each sensor data 131 is extracted (S3).

After that, the information processing apparatus 1 stores each feature amount and state information (label) indicating whether or not the sensor data corresponding to each feature amount corresponds to an abnormal state for each feature amount extracted in the process of S3. A second model 134 is generated by performing supervised learning of the associated data (S4).

That is, the information processing apparatus 1 extracts feature amounts by a method that can incorporate necessary domain knowledge from the comparison result between the sensor data 131 and the restored sensor data 131a.

As a result, the information processing apparatus 1 extracts the sensor data 131 that should not be extracted as abnormal data as abnormal data by using the learning model generated from the teacher data 132 including the feature amount in which the domain knowledge is incorporated. can be suppressed. Therefore, the information processing apparatus 1 can accurately determine whether the sensor data 131 output by the sensor 5 is abnormal data.

[Details of the first embodiment]
Next, details of the first embodiment will be described. 5 to 11 are flowcharts for explaining the details of machine learning processing according to the first embodiment. 12 to 19 are diagrams for explaining the details of the machine learning process according to the first embodiment.

[Data accumulation processing]
First, processing for accumulating sensor data 131 (hereinafter also referred to as data accumulation processing) will be described. FIG. 5 is a flowchart for explaining data accumulation processing.

As shown in FIGS. 5 and 12, the information reception unit 111 of the information processing device 1 waits until it receives the sensor data 131 from one of the sensors 5 (NO in S11).

When the sensor data 131 is received (YES in S11), the information management unit 112 of the information processing device 1 stores the received sensor data 131 in the information storage area . A specific example of the sensor data 131 will be described below.

[Specific example of sensor data]
FIG. 14 is a diagram explaining a specific example of the sensor data 131. As shown in FIG. The sensor data 131 shown in FIG. 14 is waveform data representing the transition of the X-axis acceleration detected by the sensor 5 in a predetermined time period.

Specifically, of the waveform data in the sensor data 131 shown in FIG. 14, the data corresponding to the first half of the time period indicates that the X-axis acceleration changed over time. On the other hand, among the waveform data in the sensor data 131 shown in FIG. 14, the data corresponding to the latter time period indicate that the X-axis acceleration did not change.

[Data generation process]
Next, processing for generating teacher data 132 from sensor data 131 (hereinafter also referred to as data generation processing) will be described. FIG. 6 is a flowchart for explaining data generation processing.

As shown in FIG. 6, the information reception unit 111 waits until it receives a label corresponding to the sensor data 131 stored in the information storage area 130 (NO in S21).

Specifically, the information reception unit 111 waits until it receives a label corresponding to each of the one or more sensor data 131 stored in the information storage area 130, for example. Note that the label corresponding to each sensor data 131 may be input to the information processing apparatus 1 by the operator, for example.

When the label is accepted (YES in S21), the information management unit 112 associates the sensor data 131 stored in the information storage area 130 with the label accepted in the process of S21, as shown in FIG. to generate teacher data 132 (S22).

Specifically, the information management unit 112 extracts at least part of the sensor data 131 stored in the information storage area 130 . Then, the information management unit 112 generates teacher data 132 by associating each sensor data 131 and a label corresponding to each sensor data 131 for each extracted sensor data 131 .

After that, the information management unit 112 stores the teacher data 132 generated in the process of S22 in the information storage area 130 (S23).

[Details of machine learning processing]
Next, the details of the machine learning process will be described. 7 to 9 are flowcharts explaining the details of the machine learning process.

As shown in FIG. 7, the information reception unit 111 waits until the machine learning execution timing (NO in S31).

Then, when the machine learning execution timing has come (YES in S31), the information management unit 112 selects the teacher data 132 including the sensor data 131 corresponding to the normal data from the teacher data 132 stored in the information storage area 130. Specify (S32). That is, in this case, the information management unit 112 identifies the teacher data 132 including labels corresponding to normal data.

Subsequently, as shown in FIG. 12, the first model generation unit 113 of the information processing apparatus 1 performs deep unsupervised learning using as input the sensor data 131 included in the teacher data 132 specified in the process of S32. , to generate the first model 133 (S33).

Specifically, the first model generation unit 113 generates the first model 133, which is a deep unsupervised learning model, by using, for example, an autoencoder method, a GAN (Generative Adversarial Network) method, or the like. The first model generator 113 then outputs first weight information 135 corresponding to the generated first model 133, for example.

Next, as shown in FIG. 12, the information restoration unit 114 of the information processing device 1 restores each of the sensor data 131 included in the teacher data 132 stored in the information storage area 130 to the first data generated in the process of S33. The restored sensor data 131a is obtained by inputting it into the model 133 (S34). A specific example of the restored sensor data 131a will be described below.

[Specific example of restored sensor data]
FIG. 15 is a diagram illustrating a specific example of the restored sensor data 131a. Restored sensor data 131a shown in FIG. 15 is waveform data output when the sensor data 131 described in FIG. 14 is input to the first model 133. FIG.

Specifically, the waveform data in the sensor data 131 shown in FIG. 15 indicates that the X-axis acceleration constantly changed over time.

Returning to FIG. 8, the first feature extraction unit 115 of the information processing device 1, as shown in FIG. A first feature amount corresponding to each sensor data 131 is extracted from the comparison result with the restored sensor data 131a acquired in the process of S34 (S41). Details of the process of S41 will be described below.

[Details of processing in S41]
FIG. 9 is a flowchart for explaining the details of the process of S41.

For each sensor data 131 included in the teacher data 132 stored in the information storage area 130, the first feature extraction unit 115 calculates the difference between each sensor data 131 and the restored sensor data 131a corresponding to each sensor data 131. The sensor data 131b is generated (S201). A specific example of the difference sensor data 131b will be described below.

[Specific example of differential sensor data]
16 and 17 are diagrams illustrating specific examples of the difference sensor data 131b. The difference sensor data 131b shown in FIG. 16 is waveform data showing transition of the difference in X-axis acceleration at each time between the sensor data 131 explained in FIG. 14 and the restored sensor data 131a explained in FIG.

Specifically, the waveform data in the sensor data 131 shown in FIG. 14 indicates that the X-axis acceleration changed only in the first half of the time period. Also, the waveform data in the restored sensor data 131a shown in FIG. 15 is waveform data corresponding to the case where the X-axis acceleration changes in all time zones.

Therefore, in this case, as shown in FIG. 16, first feature extraction section 115 generates waveform data indicating that the change in X-axis acceleration in the latter half of the time period is greater than the change in X-axis acceleration in the first half of the time period. It is generated as difference sensor data 131b.

Returning to FIG. 9, for each sensor data 131 included in the teacher data 132 stored in the information storage area 130, the first feature extraction unit 115 performs A first feature quantity corresponding to each sensor data 131 is extracted from each of the difference sensor data 131b (S202).

Specifically, the first feature extraction unit 115 extracts, for example, the waveform data corresponding to time t1 (hereinafter also referred to as first waveform data) from the waveform data corresponding to the difference sensor data 131b shown in FIG. The first feature amount is extracted from each of waveform data corresponding to (hereinafter also referred to as second waveform data).

More specifically, the first feature extraction unit 115 extracts, for example, the average amplitude in the first waveform data (hereinafter also referred to as the first average amplitude), the average amplitude in the second waveform data (hereinafter also referred to as the second average amplitude) ) and are calculated respectively. Then, the first feature extraction unit 115 extracts, for example, a value obtained by subtracting the second average amplitude from the first average amplitude and a value obtained by subtracting the first average amplitude from the second average amplitude as first feature amounts.

That is, in the differential sensor data 131b shown in FIG. 16, the second average amplitude is larger than the first average amplitude. Therefore, for example, when the sensor 5 that outputs the sensor data 131 described in FIG. 13 is a wearable sensor worn by a factory worker, the first feature extracted from the difference sensor data 131b shown in FIG. 16 is , for example, the feature amount reflects information that the worker's motion suddenly stopped.

On the other hand, in the differential sensor data 131b shown in FIG. 17, the first average amplitude is larger than the second average amplitude. Therefore, the first feature amount extracted from the difference sensor data 131b shown in FIG. 17 is, for example, a feature amount reflecting information indicating that the worker's action suddenly started.

Here, when the worker suddenly stops working, it can generally be determined that there is a possibility that some kind of abnormal situation has occurred in the worker. Therefore, as will be described later, the information processing apparatus 1 receives, as input, data in which the first feature amount extracted from the difference sensor data 131b shown in FIG. 16 is associated with labels corresponding to abnormal data. study with.

On the other hand, it can generally be determined that a worker's sudden start of movement is a normal behavior of the worker. Therefore, as will be described later, the information processing apparatus 1 receives, as input, data in which the first feature amount extracted from the difference sensor data 131b shown in FIG. 17 is associated with labels corresponding to normal data. study with.

As a result, the information processing apparatus 1 can generate a learning model for extracting abnormal data to be extracted even when the number of abnormal data that can be used as the teacher data 132 is not sufficient. In addition, the information processing apparatus 1 performs supervised learning using data in which necessary domain knowledge is incorporated as an input, so that the extraction target is higher than the deep unsupervised learning model generated only from the teacher data 132 corresponding to normal data. It is possible to generate a learning model that can accurately extract abnormal data.

Note that the first feature extraction unit 115 may extract the first feature amount from the plurality of sensor data 131 output from the plurality of sensors 5 of different types in the process of S41.

Specifically, the first feature extraction unit 115 extracts, for example, the average amplitude in the waveform data corresponding to the differential sensor data 131b shown in FIG. A product of the maximum value and a value obtained by multiplying the sigmoid function may be extracted as the first feature amount.

As a result, the first feature extraction unit 115, for example, not only provides information indicating whether or not the worker's motion suddenly stopped, but also indicates whether or not there was a loud sound when the worker's motion suddenly stopped. Information can be reflected in the first feature amount. Therefore, in this case, the first feature extracting unit 115 can extract the first feature quantity in a state in which it is possible to distinguish between, for example, the case where the worker squats down and the case where the worker falls down.

Returning to FIG. 8, the second feature extraction unit 116 of the information processing device 1, as shown in FIG. A corresponding second feature amount is extracted (S42).

Specifically, the second feature extraction unit 116 may extract, for example, an ECDF (Empirical Cumulative Distribution Feature) feature amount as the second feature amount in the process of S42. In addition, the second feature extraction unit 116 may use, for example, a feature extraction algorithm described manually, or a deep neural network such as CNN (Convolutional Neural Network) or LSTM (Long Short-Term Memory). A network model may be used. A specific example of the processing of S42 will be described below.

[Specific example of processing in S42]
FIG. 18 is a diagram illustrating a specific example of the process of S42. Specifically, FIG. 18 is a diagram for explaining a case of extracting an ECDF feature quantity, and is a diagram showing the cumulative distribution of the X-axis acceleration values included in the sensor data 131 explained with reference to FIG. 13 .

In the process of S42, the second feature extraction unit 116 refers to, for example, the diagram shown in FIG. 18, and the X-axis acceleration sensor value (X-axis ) is extracted as the second feature amount. Specifically, for example, as shown in FIG. 18, the second feature extraction unit 116 extracts each value of the X-axis acceleration sensor corresponding to a plurality of cumulative distribution values (for example, a plurality of predetermined values). , is extracted as the second feature amount.

Returning to FIG. 8, the feature combining unit 117 of the information processing apparatus 1, as shown in FIG. A third feature amount is generated by combining the first feature amount and the second feature amount extracted in the process of S42 (S43). A specific example of the processing of S43 will be described below.

[Specific example of processing in S43]
FIG. 19 is a diagram illustrating a specific example of the process of S43. Specifically, FIG. 19A is a diagram illustrating a specific example of the first feature amount extracted in the process of S41, and FIG. 19B is a diagram illustrating a specific example of the second feature amount extracted in the process of S42. FIG. 19C is a diagram for explaining an example, and FIG. 19C is a diagram for explaining a specific example of the third feature amount generated in the process of S43.

The first feature quantity shown in FIG. 19(A) is a 1×2 type matrix containing “0.01” and “−0.03” as elements, and the second feature quantity shown in FIG. 19(B) is , is a 1×5 type matrix containing “5.0”, “1.0”, “0.0”, “0.0” and “−1.0” as elements.

Therefore, in the process of S43, the feature combining unit 117 combines, for example, the matrix shown in FIG. 19A and the matrix shown in FIG. "0.01" and "-0.03", and "5.0", "1.0", "0.0", "0.0" and " −1.0” as elements is generated as the third feature amount.

Returning to FIG. 8, the second model generation unit 118 of the information processing apparatus 1 generates each feature amount and the sensor corresponding to each feature amount for each third feature amount generated in the process of S43, as shown in FIG. A second model 134 is generated by performing supervised learning of the data associated with the label of the data 131 (S44).

Specifically, the second model generator 118 generates the second model 134, which is a supervised learning model, by using, for example, a support vector machine (SVM), k-NN method, deep neural network, or the like. Then, the second model generation unit 118 outputs second weight information 136 that is the feature amount weight of the generated second model 134 .

Also, the second model generation unit 118 may output the second weight information 136 by using, for example, an ensemble learning method having a plurality of machine learning models (weak learners) therein. In this case, the second model generation unit 118 can automatically separate and learn each pattern even when there are a plurality of patterns of abnormal conditions, and can accurately detect abnormal conditions. It is possible to generate a second model 134 that can

Furthermore, the second model generation unit 118 can use, for example, the Random Forest method, the boosting method, the stacking method, or the like as the ensemble learning method. Also, the second model generation unit 118 can use, for example, the AdaBoost method, the LPBoost method, the Gradient Boosting method, the XGBoost method, the LightGBM method, or the CatBoost method as the boosting method.

Note that the feature amount weight of the supervised learning model here is a concept that includes, for example, weights for individual weak learners used in the boosting method, weights for neurons in each layer in a neural network, and the like.

Returning to FIG. 8, the information management unit 112 stores the first weight information 135 corresponding to the first model 133 generated in the process of S33 and the second weight information 136 corresponding to the second model 134 generated in the process of S44. is stored in the information storage area 130 (S45).

[Details of anomaly extraction processing]
Next, the process of determining whether or not the sensor data 131 output from the sensor 5 is abnormal data (hereinafter referred to as abnormality extraction process) will be described. 10 and 11 are flowcharts for explaining anomaly extraction processing.

As shown in FIG. 10, the information receiving unit 111 waits until it receives the sensor data 131 (new sensor data 131) output from the sensor 5 (NO in S51).

Then, when the sensor data 131 is received (YES in S51), the information restoration unit 114 converts the sensor data 131 received in the process of S51 into the first model 133 generated in the process of S33, as shown in FIG. By inputting, the restored sensor data 131a is acquired (S52).

Subsequently, as shown in FIG. 13, the first feature extraction unit 115 compares the sensor data 131 received in the process of S51 with the restored sensor data 131a acquired in the process of S52, and the A first feature quantity corresponding to the sensor data 131 is extracted (S53).

Further, as shown in FIG. 13, the second feature extraction unit 116 extracts a second feature amount corresponding to the sensor data 131 received in the process of S51 from the sensor data 131 received in the process of S51 (S54). .

Then, as shown in FIG. 13, the feature combining unit 117 generates a third feature amount by combining the first feature amount extracted in the process of S53 and the second feature amount extracted in the process of S54 (S55 ).

After that, as shown in FIGS. 11 and 13, the output information generation unit 121 of the information processing device 1 inputs the third feature amount generated in the process of S55 to the second model 134 generated in the process of S43. (S61).

Then, as shown in FIG. 13, the information output unit 122 of the information processing device 1 outputs the output information 137 including the estimation result corresponding to the score acquired in the process of S61 (S62).

Specifically, for example, when the score acquired in the process of S61 is greater than a predetermined threshold value, the information output unit 122 estimates that the sensor data 131 received in the process of S31 is normal data. For example, when the score acquired in the process of S61 is smaller than a predetermined threshold value, the information output unit 122 estimates that the sensor data 131 received in the process of S31 is abnormal data. Then, the information output unit 122 outputs, for example, the output information 137 including the estimation result corresponding to the score acquired in the process of S61 to the operator's operation terminal (not shown).

In this manner, the information processing apparatus 1 according to the present embodiment generates the first model 133 by performing deep unsupervised learning on the plurality of sensor data 131 output from the plurality of sensors 5 . Then, the information processing apparatus 1 compares each of the sensor data 131 with the restored sensor data 131a output from the first model 133 in accordance with the input of each sensor data 131, for each of the plurality of sensor data 131. A feature quantity corresponding to the sensor data 131 is extracted. After that, the information processing apparatus 1 teaches data in which each feature amount is associated with state information indicating whether or not the sensor data 131 corresponding to each feature amount corresponds to an abnormal state for each extracted feature amount. A second model 134 is generated by performing attached learning.

That is, the information processing apparatus 1 extracts feature amounts by a method that can incorporate necessary domain knowledge from the comparison result between the sensor data 131 and the restored sensor data 131a.

As a result, the information processing apparatus 1 can generate a learning model (a supervised learning model) from the teacher data 132 including feature quantities in which domain knowledge is incorporated. By using the generated learning model, the information processing apparatus 1 can prevent the sensor data 131 that should not be extracted as abnormal data from being extracted as abnormal data. Therefore, the information processing apparatus 1 can accurately determine whether the sensor data 131 output by the sensor 5 is abnormal data.

[Second embodiment]
Next, a second embodiment will be described. 20 to 24 are flowcharts for explaining machine learning processing and abnormality estimation processing in the second embodiment. 25 and 26 are diagrams for explaining machine learning processing in the second embodiment.

Note that the data accumulation processing and data generation processing in the second embodiment have the same contents as the data accumulation processing and data generation processing in the first embodiment, so the description thereof will be omitted.

[Machine learning processing in the second embodiment]
First, machine learning processing in the second embodiment will be described. 20 to 22 are flowcharts for explaining machine learning processing in the second embodiment.

As shown in FIG. 20, the information reception unit 111 waits until the machine learning execution timing (NO in S71).

Then, when the machine learning execution timing has come (YES in S71), the information management unit 112 selects the teacher data 132 including the sensor data 131 corresponding to the normal data from the teacher data 132 stored in the information storage area 130. Specify (S72).

Subsequently, as shown in FIG. 25, the data selection unit 119 of the information processing device 1 selects each of the plurality of groups of teacher data 132 from the teacher data 132 specified in the process of S72 (S73).

Specifically, the data selection unit 119 selects a plurality of groups of teacher data 132 such that, for example, at least part of the teacher data 132 included in each group is different from the teacher data 132 included in other groups. . Also, the data selection unit 119 selects a plurality of groups of teacher data 132 such that, for example, each group contains a different number of teacher data 132 .

Then, as shown in FIG. 25, the first model generation unit 113 identifies one group among the plurality of groups of teacher data 132 selected in the process of S73 (S74).

Subsequently, as shown in FIG. 25, the first model generation unit 113 performs deep unsupervised learning using as input the sensor data 131 included in the teacher data 132 of the group identified in the process of S74, thereby generating the first A model 133 is generated (S75).

Next, as shown in FIGS. 21 and 25, the information restoration unit 114 converts each of the sensor data 131 included in the teacher data 132 of the group identified in the process of S74 into the first model 133 generated in the process of S75. to obtain the restored sensor data 131a (S81).

Then, as shown in FIG. 25, the first feature extraction unit 115 extracts each sensor data 131 and the restored sensor data 131a acquired in the process of S81 for each sensor data 131 included in the teacher data 132 specified in the process of S74. , the first feature amount corresponding to each sensor data 131 is extracted (S82).

25, for each sensor data 131 included in the teacher data 132 of the group specified in the process of S74, the second feature extraction unit 116 extracts the second feature corresponding to each sensor data 131 from each sensor data 131. 2 Extract feature values (S83).

Furthermore, as shown in FIG. 25, the feature combining unit 117 combines the first feature amount extracted in the process of S82 with the A third feature amount is generated by combining the second feature amount extracted in step S84.

After that, as shown in FIG. 25, the second model generation unit 118 associates each feature amount with the label of the sensor data 131 corresponding to each feature amount for each third feature amount generated in the process of S84. A second model 134 is generated by performing supervised learning of the obtained data (S85).

Note that in the process of S81, the information restoration unit 114 restores each of the sensor data 131 included in the teacher data 132 stored in the information storage area 130 in the process of S75, as in the first embodiment. The restored sensor data 131 a may be acquired by inputting the generated first model 133 .

In this case, the first feature extraction unit 115 extracts the first feature amount and the second feature amount for each sensor data 131 included in the teacher data 132 stored in the information storage area 130 in the processing from S82 to S84. Extraction may be performed, and the third feature amount may be generated for each sensor data 131 included in the teacher data 132 stored in the information storage area 130 .

Returning to FIG. 21, the model selection unit 120 of the information processing device 1 determines whether or not all groups have been identified in the process of S74 (S86).

As a result, when it is determined that all the groups have not been specified (NO in S86), the first model generation unit 113 executes the processes after S74 again.

On the other hand, if it is determined that all groups have been specified (YES in S86), the model selection unit 120 selects a predetermined (S91).

Specifically, the model selection unit 120 selects, for example, the second model 134 with the highest recognition performance among the second models 134 generated in the process of S85. The recognition performance here is calculated not only from the accuracy rate of recognition in the objective function of supervised learning (e.g. precision, recall, F1 score or AUC value, etc.), but also from the regularization term. including cases where

Note that the model selection unit 120 selects the second model 134 by, for example, specifying the maximum value of selectable deep unsupervised learning models and using a discrete optimization algorithm such as the greedy method. good. In addition, the model selection unit 120, for example, designs an objective function that gives a penalty as a regularization term to the number of selected deep unsupervised learning models, and uses a continuous optimization algorithm such as the Group Lasso method, Selection of the second model 134 may be performed.

Furthermore, the model selection unit 120 may select a plurality of second models 134 that maximize the recognition performance in the process of S91. In the process of S91, the model selection unit 120 may select a learning model that combines the plurality of second models 134 generated in the process of S85.

Returning to FIG. 22, the model selection unit 120 identifies the first model 133 corresponding to the second model 134 selected in the process of S91 among the first models 133 generated in the process of S75, as shown in FIG. (S92).

After that, the information management unit 112 stores the first weight information 135 corresponding to the first model 133 specified in the process of S92, the second weight information 136 corresponding to the second model 134 selected in the process of S91, and the Model selection information 138 indicating the first model 133 specified in the process is stored in the information storage area 130 (S93).

That is, when the contents of the teacher data 132 used to generate the first model 133 and the second model 134 are biased, the information processing apparatus 1 selects a learning model that can accurately extract abnormal data to be extracted. may not be generated.

Specifically, for example, the rate of occurrence of the sensor data 131 output while the worker is working in the factory and the sensor data 131 output while the worker is working in the office is If there is a large difference depending on the time zone, the teacher data 132 used to generate the learning model may include one sensor data 131 more than necessary. Therefore, the information processing apparatus 1 may not be able to generate a learning model capable of accurately extracting abnormal data to be extracted due to occurrence of over-learning or the like.

Therefore, in the machine learning process according to the second embodiment, a plurality of first models 133 are generated by using teacher data 132 different from each other. Then, the information processing apparatus 1 selects the first model 133 determined to be capable of accurately extracting the abnormal data to be extracted.

As a result, the information processing device 1 can generate a learning model that can more accurately extract abnormal data to be extracted.

Note that the model selection unit 120 may perform the processes after S91 on the second model 134 generated up to that point before all groups are identified in the process of S74. Then, the model selection unit 120 may repeat the processes after S91, for example, each time a predetermined number of second models 134 are generated.

This enables the model selection unit 120 to quickly select the second model 134 in the process of S91 and specify the first model 133 in the process of S92.

In addition, in the process of S73, the data selection unit 119 not only changes the number and contents of the teacher data 132 used in deep unsupervised learning for each first model 133 to be generated, but also The hyperparameters used may be changed, or the network architecture of the neural network used in deep unsupervised learning may be changed.

Specifically, the data selection unit 119 may change the hyperparameters by, for example, changing the optimization algorithm used in the stochastic gradient descent method, changing the learning rate, or changing the momentum value. The data selection unit 119 may change the network architecture of the neural network, for example, by changing the sample size in the input layer or the convolution size in the CNN layer.

Furthermore, in the process of S73, the data selection unit 119 may change the number and contents of the teacher data 132 used in deep unsupervised learning, for example, by selecting the teacher data 132 stored in the information storage area 130 a plurality of times. A plurality of groups of teacher data 132 may be selected by sampling . As the sampling here, for example, random sampling may be performed, or sampling by the Bagging method may be performed. Then, the data selection unit 119 may select a plurality of groups of teacher data 132 and then assign a weight to each of the selected teacher data 132 .

[Abnormality extraction processing in the second embodiment]
Next, anomaly extraction processing in the second embodiment will be described. 23 and 24 are flowcharts for explaining anomaly extraction processing according to the second embodiment.

As shown in FIG. 23, the information receiving unit 111 waits until it receives the sensor data 131 (new sensor data 131) output from the sensor 5 (NO in S101).

Then, when the sensor data 131 is accepted (YES in S101), the information restoration unit 114 converts the sensor data 131 accepted in the process of S101 to the first model 133 specified in the process of S92, as shown in FIG. By inputting, the restored sensor data 131a is acquired (S102).

Subsequently, as shown in FIG. 26, the first feature extraction unit 115 compares the sensor data 131 received in the process of S101 with the restored sensor data 131a acquired in the process of S102, and determines the A first feature quantity corresponding to the sensor data 131 is extracted (S103).

Further, as shown in FIG. 26, the second feature extraction unit 116 extracts a second feature quantity corresponding to the sensor data 131 received in the process of S101 from the sensor data 131 received in the process of S101 (S104). .

Then, as shown in FIG. 26, the feature combining unit 117 generates a third feature amount by combining the first feature amount extracted in the process of S103 and the second feature amount extracted in the process of S104 (S105 ).

After that, as shown in FIGS. 24 and 26, the output information generation unit 121 obtains a score by inputting the third feature amount generated in the process of S105 to the second model 134 selected in the process of S91. (S111).

Then, as shown in FIG. 26, the information output unit 122 outputs output information 137 including the estimation result corresponding to the score acquired in the process of S111 (S112).

As a result, the information processing device 1 can more accurately extract the abnormal data to be extracted.

In the above example, the sensor 5 is a wearable sensor worn by a worker working in a factory, but the sensor 5 is a sensor attached to each device in the factory, The image information output by each sensor may be used to monitor the operation status of each device. Further, the sensor 5 is, for example, a sensor installed inside a store, and may be used for detecting suspicious persons, counterfeit brand products, etc. inside the store.

Furthermore, in the above example, a case where a plurality of sensors 5 exist has been described, but the information processing system 10 may have only a single sensor 5 . Further, in the above example, the case where there are a plurality of sensor data 131 used for learning has been described, but the information processing apparatus 1 uses only a single sensor data 131 as the sensor data 131 used for learning. may

The above embodiments are summarized as follows.

(Appendix 1)
Generate a first model by performing deep unsupervised learning of multiple sensor data output from multiple sensors,
For each of the plurality of sensor data, extracting a feature amount corresponding to each sensor data from a comparison result between each sensor data and the restored sensor data output from the first model in accordance with the input of each sensor data,
For each of the extracted feature amounts, performing supervised learning of data in which each feature amount is associated with state information indicating whether or not sensor data corresponding to each feature amount corresponds to an abnormal state, 2 generate a model,
A machine learning program characterized by causing a computer to execute processing.

(Appendix 2)
In Appendix 1,
In the process of generating the first model, among the plurality of sensor data, deep unsupervised learning of sensor data associated with the state information indicating a normal state is performed.
A machine learning program characterized by:

(Appendix 3)
In Appendix 1,
In the process of extracting the feature amount,
generating differential sensor data indicating a difference between each sensor data and the restored sensor data corresponding to each sensor data for each of the plurality of sensor data;
extracting the feature quantity from each of the generated differential sensor data;
A machine learning program characterized by:

(Appendix 4)
In Appendix 3,
Each of the plurality of sensor data is information indicating transition in a predetermined time period of information output from one of the plurality of sensors,
In the process of extracting the feature amount, the feature amount is extracted from each of the difference sensor data for each of a plurality of time periods included in the predetermined time period.
A machine learning program characterized by:

(Appendix 5)
In Appendix 1,
In the process of generating the second model,
extracting other feature quantities from each of the plurality of sensor data;
generating a combined feature amount by combining the feature amount and the other feature amount for each of the plurality of sensor data;
For each of the generated combined feature amounts, perform supervised learning of information in which each combined feature amount is associated with state information indicating whether or not sensor data corresponding to each combined feature amount corresponds to an abnormal state;
A machine learning program characterized by:

(Appendix 6)
In Supplementary Note 1, further,
When new sensor data is output from any one of the plurality of sensors, the new sensor data and the sensor data output from the first model in accordance with the input of the new sensor data. Extract new features from the results of comparison with
Acquiring a score output from the second model in accordance with the input of the new feature amount,
outputting the obtained score;
A machine learning program characterized by causing a computer to execute processing.

(Appendix 7)
In Appendix 6,
In the process of generating the first model,
generating a plurality of groups each including at least a portion of the plurality of sensor data;
generating the first model for each of the plurality of groups;
In the process of extracting the feature amount, the feature amount is extracted for each of the plurality of generated groups by using sensor data included in each group,
In the process of generating the second model, for each of the generated groups, the second model is generated by using the feature amount corresponding to each group;
identifying the first model used to generate the second model that satisfies a predetermined condition among the generated second models;
outputting information indicating the identified first model;
A machine learning program characterized by causing a computer to execute processing.

(Appendix 8)
In Appendix 7,
Sensor data included in each of the plurality of groups is at least partially different from sensor data included in other groups of the plurality of groups,
A machine learning program characterized by:

(Appendix 9)
In Appendix 6,
In the process of generating the first model,
generating a plurality of groups each including at least a portion of the plurality of sensor data;
generating the first model for each of the plurality of groups;
In the process of extracting the feature amount, extracting the feature amount by using each model for each of the generated first models,
In the process of generating the second model, for each of the generated first models, the second model is generated by using the feature amount corresponding to each model;
identifying the first model used to generate the second model that satisfies a predetermined condition among the generated second models;
outputting information indicating the identified first model;
A machine learning program characterized by causing a computer to execute processing.

(Appendix 10)
Generate a first model by performing deep unsupervised learning of multiple sensor data output from multiple sensors,
For each of the plurality of sensor data, extracting a feature amount corresponding to each sensor data from a comparison result between each sensor data and the restored sensor data output from the first model in accordance with the input of each sensor data,
For each of the extracted feature amounts, performing supervised learning of data in which each feature amount is associated with state information indicating whether or not sensor data corresponding to each feature amount corresponds to an abnormal state, 2 generate a model,
A machine learning method characterized by having a computer execute processing.

(Appendix 11)
In Appendix 10,
In the process of generating the first model, among the plurality of sensor data, deep unsupervised learning of sensor data associated with the state information indicating a normal state is performed.
A machine learning method characterized by:

(Appendix 12)
In Appendix 10,
In the process of extracting the feature amount,
generating differential sensor data indicating a difference between each sensor data and the restored sensor data corresponding to each sensor data for each of the plurality of sensor data;
extracting the feature quantity from each of the generated differential sensor data;
A machine learning method characterized by:

(Appendix 13)
In Supplementary Note 10, further,
When new sensor data is output from any one of the plurality of sensors, the new sensor data and the sensor data output from the first model in accordance with the input of the new sensor data. Extract new features from the results of comparison with
Acquiring a score output from the second model in accordance with the input of the new feature amount,
outputting the obtained score;
A machine learning method characterized by having a computer execute processing.

(Appendix 14)
In a machine learning system having a plurality of sensors and an information processing device connected to the plurality of sensors,
each of the plurality of sensors transmits detected sensor data to the information processing device;
The information processing device is
generating a first model by performing deep unsupervised learning of a plurality of sensor data transmitted from the plurality of sensors;
For each of the plurality of sensor data, extracting a feature amount corresponding to each sensor data from a comparison result between each sensor data and the restored sensor data output from the first model in accordance with the input of each sensor data,
For each of the extracted feature amounts, performing supervised learning of data in which each feature amount is associated with state information indicating whether or not sensor data corresponding to each feature amount corresponds to an abnormal state, 2 generate a model,
A machine learning system characterized by:

(Appendix 15)
In Appendix 14,
The information processing device performs deep unsupervised learning of sensor data associated with the state information indicating a normal state among the plurality of sensor data,
A machine learning system characterized by:

(Appendix 16)
In Appendix 14,
The information processing device is
generating differential sensor data indicating a difference between each sensor data and the restored sensor data corresponding to each sensor data for each of the plurality of sensor data;
extracting the feature quantity from each of the generated differential sensor data;
A machine learning system characterized by:

(Appendix 17)
In Supplementary Note 14, further,
The information processing device is
When new sensor data is output from any one of the plurality of sensors, the new sensor data and the sensor data output from the first model in accordance with the input of the new sensor data. Extract new features from the results of comparison with
Acquiring a score output from the second model in accordance with the input of the new feature amount,
outputting the obtained score;
A machine learning system characterized by:

1: Information processing device 5: Operation terminal NW: Network 10: Information processing system

Claims (10)

Generate a first model by performing deep unsupervised learning of multiple sensor data output from multiple sensors,
For each of the plurality of sensor data, extracting a feature amount corresponding to each sensor data from a comparison result between each sensor data and the restored sensor data output from the first model in accordance with the input of each sensor data,
extracting other feature quantities from each of the plurality of sensor data;
generating a combined feature amount by combining the feature amount and the other feature amount for each of the plurality of sensor data;
Performing supervised learning of data in which each combined feature amount and state information indicating whether or not sensor data corresponding to each combined feature amount corresponds to an abnormal state for each of the generated combined feature amounts. Generate a second model by
A machine learning program characterized by causing a computer to execute processing. In claim 1,
In the process of generating the first model, among the plurality of sensor data, deep unsupervised learning of sensor data associated with the state information indicating a normal state is performed.
A machine learning program characterized by: In claim 1,
In the process of extracting the feature amount,
generating differential sensor data indicating a difference between each sensor data and the restored sensor data corresponding to each sensor data for each of the plurality of sensor data;
extracting the feature quantity from each of the generated differential sensor data;
A machine learning program characterized by: In claim 3,
Each of the plurality of sensor data is information indicating transition in a predetermined time period of information output from one of the plurality of sensors,
In the process of extracting the feature amount, the feature amount is extracted from each of the difference sensor data for each of a plurality of time periods included in the predetermined time period.
A machine learning program characterized by: In claim 1, further:
When new sensor data is output from any one of the plurality of sensors, the new sensor data and the sensor data output from the first model in accordance with the input of the new sensor data. Extract new features from the results of comparison with
Acquiring a score output from the second model in accordance with the input of the new feature amount,
outputting the obtained score;
A machine learning program characterized by causing a computer to execute processing. In claim 5 ,
In the process of generating the first model,
generating a plurality of groups each including at least a portion of the plurality of sensor data;
generating the first model for each of the plurality of groups;
In the process of extracting the feature amount, the feature amount is extracted for each of the plurality of generated groups by using sensor data included in each group,
In the process of generating the second model, for each of the generated groups, the second model is generated by using the feature amount corresponding to each group;
identifying the first model used to generate the second model that satisfies a predetermined condition among the generated second models;
outputting information indicating the identified first model;
A machine learning program characterized by causing a computer to execute processing. In claim 6 ,
Sensor data included in each of the plurality of groups is at least partially different from sensor data included in other groups of the plurality of groups,
A machine learning program characterized by: In claim 5 ,
In the process of generating the first model,
generating a plurality of groups each including at least a portion of the plurality of sensor data;
generating the first model for each of the plurality of groups;
In the process of extracting the feature amount, extracting the feature amount by using each model for each of the generated first models,
In the process of generating the second model, for each of the generated first models, the second model is generated by using the feature amount corresponding to each model;
identifying the first model used to generate the second model that satisfies a predetermined condition among the generated second models;
outputting information indicating the identified first model;
A machine learning program characterized by causing a computer to execute processing. Generate a first model by performing deep unsupervised learning of multiple sensor data output from multiple sensors,
For each of the plurality of sensor data, extracting a feature amount corresponding to each sensor data from a comparison result between each sensor data and the restored sensor data output from the first model in accordance with the input of each sensor data,
extracting other feature quantities from each of the plurality of sensor data;
generating a combined feature amount by combining the feature amount and the other feature amount for each of the plurality of sensor data;
Performing supervised learning of data in which each combined feature amount and state information indicating whether or not sensor data corresponding to each combined feature amount corresponds to an abnormal state for each of the generated combined feature amounts. Generate a second model by
A machine learning method characterized by having a computer execute processing. In a machine learning system having a plurality of sensors and an information processing device connected to the plurality of sensors,
each of the plurality of sensors transmits detected sensor data to the information processing device;
The information processing device is
generating a first model by performing deep unsupervised learning of a plurality of sensor data transmitted from the plurality of sensors;
For each of the plurality of sensor data, extracting a feature amount corresponding to each sensor data from a comparison result between each sensor data and the restored sensor data output from the first model in accordance with the input of each sensor data,
extracting other feature quantities from each of the plurality of sensor data;
generating a combined feature amount by combining the feature amount and the other feature amount for each of the plurality of sensor data;
Performing supervised learning of data in which each combined feature amount and state information indicating whether or not sensor data corresponding to each combined feature amount corresponds to an abnormal state for each of the generated combined feature amounts. Generate a second model by
A machine learning system characterized by:

Images