JP7047372B2 - Data identification device and data identification method - Google Patents

Description

The present invention relates to an abnormality detection device, an abnormality detection method, and an abnormality detection program for discriminating whether the data is normal data or abnormal data with respect to data constantly acquired from a production process or data acquired in the past. ..

Machine learning methods for identifying whether or not data contained in a certain data group are normal data are known, and can be broadly divided into two methods: supervised anomaly detection method and unsupervised anomaly detection method. As a supervised abnormality detection method, a method as described in Patent Document 1 is known. However, in supervised anomaly detection, it is necessary to prepare normal data and abnormal data to the same extent in advance in order to obtain sufficient identification accuracy.

However, if most of the products obtained from the production equipment of the product are non-defective and the incidence of defective products is low, most of the data obtained from the production equipment is normal data obtained from good products, and abnormalities obtained from defective products. There may be little data. In such a case, when the supervised anomaly detection method is used, the anomaly data cannot be sufficiently obtained, so that the characteristics of the anomaly data cannot be sufficiently grasped, and the identification accuracy for the anomaly data becomes low. There are challenges.

Since the unsupervised anomaly detection technique is an anomaly detection method that uses normal data as learning data, there is an advantage that the anomaly data can be identified as a state different from the normal data without using the anomaly data for learning. Patent Document 2 proposes an anomaly detection method using an unsupervised anomaly detection technique.

Japanese Unexamined Patent Publication No. 2014-26455 Japanese Unexamined Patent Publication No. 2017-97718

However, when the above unsupervised anomaly detection technology is used in the production process of products, the data acquired may change due to changes in environmental conditions over time, or the data acquired by updating production equipment and peripheral equipment. The above-mentioned unsupervised anomaly detection technology alone cannot cope with this change, and it may not be possible to make an accurate judgment.

The present invention has been made in view of the above points, and an object of the present invention is to accurately detect anomalies using an unsupervised anomaly detection method even under conditions where environmental conditions change over time.

In order to solve the above problems, the present invention is a data identification device that discriminates whether or not the input data is peculiar data in the data group.
A feature amount extraction means for extracting a feature amount from the input data using the first neural network,
A data reconstruction means for generating reconstruction data having the same size as the input data using the second neural network from the feature amount,
A parameter determining means for determining the parameters of the first neural network and the parameters of the second neural network by the error back propagation method from the error between the reconstructed data and the input data.
A data evaluation means for evaluating the magnitude of the error between the reconstructed data and the input data, and
An identification means for discriminating whether or not the input data is singular data by comparing the evaluation value calculated by the evaluation means with a preset threshold value.
A learning mode in which the feature amount extraction means, the data reconstruction means, and the parameter determination means are operated, and a data identification mode in which the feature amount extraction means, the data reconstruction means, the data evaluation means, and the identification means are operated. It is provided with a control means for operating the two modes of, independently or simultaneously.

In the data identification device of the present invention, the data group is derived from an identification target that is continuously conveyed.
The data identification device further includes a data acquisition means for acquiring data related to the identification target and a data storage means for accumulating the data acquired by the data acquisition means.
When the identification target is in the transport state, the control means inputs the data acquired by the data acquisition means to the feature amount extraction means to operate the data identification mode, and the identification target is not in the transport state. At times, it is preferable that the data accumulated by the data storage means is input to the feature amount extraction means when the identification target is in the transport state to operate the learning mode.

In the data identification device of the present invention, it is preferable that the identification target is a conveyed product, a resin or a thread conveyed by a conveying belt.

Further, in order to solve the above problems, the present invention causes the first neural network and the second neural network to learn the data acquired and accumulated for a certain period of time, so that the data currently being acquired is included in the data group. This is a data identification method for identifying whether or not the data is peculiar to the above, and the following steps 1 and 2 are performed.
(Step 1)
Feature 1 is extracted from the data acquired and accumulated for a certain period of time using the first neural network.
Using the feature quantity 1 to the second neural network, reconstruction data having the same size as the accumulated data is generated.
From the error between the reconstructed data and the accumulated data, the parameters of the first neural network and the parameters of the second neural network are determined by the error back propagation method.
(Step 2)
Using the first neural network in which the parameters determined in the step 1 are set, the feature amount 2 is extracted from the data being acquired, and the feature amount 2 is extracted.
Using the second neural network in which the parameters determined in the step 1 are set, reconstruction data having the same size as the data being acquired is generated from the feature amount 2.
Whether or not the data being acquired is singular data is identified from the magnitude of the error between the reconstructed data and the data being acquired.

The data identification method of the present invention is preferably any of the following methods.
-A method in which the step 1 is first performed only once, and then the step 2 is continuously performed.
-In a method in which the step 1 is repeated in a predetermined cycle while the step 2 is continuously performed, the data acquired and accumulated for the fixed period used in the step 1 is the step 1 and the step 1. It is the data obtained in the process of step 2 performed continuously with the process 1 performed immediately before, and is set in the first neural network and the second neural network in the step 2. The parameters are the parameters determined by the step 1 performed most recently in the step 2.
-Data acquired and accumulated for the fixed period used in the step 1 by a method in which the step 1 is repeated at a predetermined cycle and the step 2 is continuously performed while the step 1 is not performed. Is the data obtained in the process of step 2 continuously performed between the step 1 and the step 1 performed immediately before the step 1, and the first neural network and the first neural network in the step 2 and The parameters set in the second neural network are the parameters determined by the most recent step 1 of the step 2, respectively.

According to the present invention, an unsupervised abnormality detection method is used to determine whether the data acquired from a target is normal or abnormal, and the acquired data changes due to changes in environmental conditions over time, or a production device. By providing a function to relearn at an appropriate cycle even in a system that affects the data acquired by updating peripheral devices, it is possible to make a judgment corresponding to the change.

FIG. 1 is a diagram showing a data identification device of the present invention. FIG. 2 is a schematic configuration diagram of the data identification device of the present invention. FIG. 3 is a diagram showing an example of a first neural network used for the feature amount extraction unit 11. FIG. 4 is a diagram showing an example of a second neural network used for the data reconstruction unit 12. FIG. 5 is a diagram showing an example of a neural network in which a first neural network used in the feature amount extraction unit 11 and a second neural network used in the data reconstruction unit 12 are combined. FIG. 6 is an operation flow diagram when the data identification device of the present invention operates in the learning mode. FIG. 7 is an operation flow diagram when the data identification device of the present invention operates in the identification mode. FIG. 8 is a diagram showing one of the methods for carrying out the present invention. FIG. 9 is identification data of each time acquired in one of the methods of carrying out the present invention. FIG. 10 is a diagram showing the time change of the evaluation value output when the data identification method A is used in one of the implementation methods of the present invention. FIG. 11 is a diagram showing the time change of the evaluation value output when the data identification method B is used in one of the implementation methods of the present invention. FIG. 12 is a diagram showing a time change of the evaluation value output when the data identification method C is used in one of the implementation methods of the present invention.

Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings.

<Device configuration>
FIG. 1 shows an example of an embodiment of a data identification device according to the embodiment of the present invention. This data identification device has an arithmetic unit 1, a data acquisition device 2, and a display device 3. The arithmetic unit 1 is a PC (Personal) Co Etc., the program or data is called from the recording device to the memory, and the CPU performs the calculation. The recording device is not limited to the configuration shown in the figure, and an HDD can be connected to the outside of the arithmetic unit 1, a CD, a DVD, a Blu-ray (registered trademark), or the like can be used. The data acquisition device 2 is a sensor typified by a camera, a microphone, or the like, and is electrically connected to the arithmetic unit 1. The display device 3 is a display or the like and is electrically connected to the arithmetic unit 1, and can be configured to display the determination result of the identification target.

FIG. 2 specifically shows an outline of the internal configuration of the data identification device of FIG. 1. The arithmetic unit 1 includes a data acquisition unit 10, a feature amount extraction unit 11, a data reconstruction unit 12, a data evaluation unit 13, a parameter determination unit 14, a determination unit 15, a result output unit 16, and a control unit 17. .. The data acquisition unit 10, the feature amount extraction unit 11, the data reconstruction unit 12, the data evaluation unit 13, the parameter determination unit 14, the determination unit 15, the result output unit 16, and the control unit 17 are one or one stored in the arithmetic unit 1. It is a program or arithmetic circuit consisting of multiple elements.

<Functional configuration>
The data acquisition unit 10 has a function of acquiring identification target data via the data acquisition device 2 and temporarily recording the data in a memory or the like. Further, it is desirable that the data acquisition unit 10 has a function of storing the identification target data in a recording device inside or outside the arithmetic unit.

The feature amount extraction unit 11 extracts a low-dimensional feature amount indicating the identification target data by using the first neural network. The data acquired by the sensor as the identification target data can be treated as a multidimensional vector, and the value of the neuron in the first neural network is interpreted as one vector. Specifically, a neural network as shown in FIG. 3 can be considered. As a constraint on the first neural network used in the feature amount extraction unit 11, the number of neurons in the input layer needs to match the dimension n of the identification target data. Further, the number m of the feature amount to be extracted must be smaller than n. There are no restrictions on the number of layers of the first neural network or the number of neurons used in the intermediate layer, but it is desirable to increase the number according to the number of dimensions of the data to be identified. In addition, when the data to be identified is an image or a moving image, the interrelationship between adjacent neurons of the acquired data is important, so a convolution layer (Convolution Layer) with restrictions on the connection between neurons is used. This makes it possible to extract more important features.

The data reconstruction unit 12 reconstructs the same n-dimensional data as the identification target data from the feature amount extracted from the feature amount extraction unit 11 using the second neural network. Specifically, the neural network has the configuration shown in FIG. The constraint on the second neural network used in the data reconstructing unit 12 is that the number of neurons in the output layer is the same as the data to be identified. There are no restrictions on the number of layers in the second neural network or the number of neurons used in the intermediate layer, but when the data to be identified is an image or video, the interrelationship between adjacent neurons of the acquired data is important. By using a deconvolution layer with restrictions on the connections between neurons, it is possible to reconstruct data with higher accuracy. The feature amount extraction unit 11 and the data reconstruction unit 12 can also be used as one neural network as shown in FIG.

The data evaluation unit 13 evaluates the difference between the identification target data and the reconstruction data obtained by the data reconstruction unit 12. As a method for evaluating the difference between the two data, there are SSD (Sum of Squared Difference) or SAD (Sum of Absolute Difference). Each index is an index that becomes 0 when the identification target data and the reconstructed data match, and the larger the numerical value, the larger the difference between the data.

The parameter determination unit 14 determines the parameters of the neural network used for the feature amount extraction unit 11 and the data reconstruction unit 12 based on the results of the data evaluation unit 13. Here, the parameters of the neural network are weights and biases. The weight is a numerical value indicating the connection strength between each neuron of the neural network, and the bias is a constant independent of the value of the neuron added when information is transmitted between the neurons.

The parameters are corrected by the error back propagation method from the error between the identification target data and the reconstructed data calculated by the data evaluation unit 13. In this case, the identification target data is acquired, and the same process is repeated using the accumulated data. It is known that it is effective to use the mini-batch stochastic gradient descent method when learning with a neural network. By combining multiple pieces of data to be identified, a mini-batch is created and the error is calculated at the same time. It is effective to correct the parameters using the total value of them. When learning by the error back propagation method, it is preferable to determine the final parameter while gradually repeating the parameter correction.

The determination unit 15 compares the error value calculated by the error data evaluation unit 13 with the preset threshold value, and determines whether the identification target data is normal or abnormal. Since the data having a large error is the data in which the difference between the original identification target data and the reconstructed data is large, it is determined that the data has not been sufficiently learned in the past, that is, the abnormal data. Data with a small error is judged to be sufficiently learned data. The threshold value is the average value or median value of normal data plus a constant multiple of the standard deviation. If it is known that there is abnormal data in the identification target data or accumulated data, the minimum error calculated by the data evaluation unit 13 using the feature amount extraction unit 11 and the data reconstruction unit 12 from the abnormal data. You can set the value and so on.

The result output unit 16 notifies or displays the result of the determination unit 15. When the determination unit 15 determines that the abnormality is determined, the result output unit 16 notifies the outside by displaying the identification target data on the screen, outputting an abnormality signal from the arithmetic unit, or the like.

The control unit 17 controls each of the above-mentioned units according to the state to operate the data identification device of the present invention. Control is performed in two control modes, a learning mode (process 1) and an identification mode (process 2). The processing procedure or operation order of the data acquisition unit 10, the feature amount extraction unit 11, the data reconstruction unit 12, the data evaluation unit 13, the parameter determination unit 14, the determination unit 15, and the result output unit 16 is controlled according to each control mode. do.

<Operation of data identification device>
FIG. 6 shows an operation flow when the data identification device of the present invention operates in the learning mode (step 1), and FIG. 7 shows an operation flow when the data identification device operates in the identification mode (step 2).

In order to identify data using the data identification device of the present invention, first, the data is operated in the learning mode and then in the identification mode.

The operation flow in the case of operating in the learning mode (step 1) will be described with reference to FIG. Before entering the processing step in the learning mode, first, in step S10, the accumulated data is divided into units (2 to 64) of predetermined values. The divided unit is called a mini-batch. Let the number of mini-batch be N- _b and set the specified value of the number of mini-batch. Further, since it is necessary to repeatedly correct the parameters of the first neural network, a specified value _Ne for the number of repetitions is set in advance. First, the values of the mini-batch counter n _b indicating the number of mini-batch currently being processed and the repetition counter _ne indicating the current number of repetitions are initialized, and 1 is added to these values before starting the following step S11. do. Next, in step S11, the identification target data group is acquired from the data acquisition unit 10, and in step S12, the feature amount extraction unit 11 is used to extract the feature amount 1 group representing the identification target data group. In step S13, a reconstruction data group is generated from the obtained feature quantity 1 group by the second neural network of the data reconstruction unit 12. Next, in step S14, the error between the identification target data and the reconstructed data is calculated from the data evaluation unit 13. In step S15, when the error of all the identification target data included in the mini-batch has been calculated, the first neural network of the feature quantity extraction unit 11 and the data re-reproduction by the error back propagation method from the error obtained by calculating the average value of the error. The parameters of the second neural network of the component 12 are modified to determine the parameters.
These series of processes are repeated until the mini-batch number counter n _b being processed reaches the mini-batch number specified value N _b , and the processes of steps S11 to S15 are repeated for all mini-batch. When the number of mini-batch exceeds the specified value, it is determined whether the value of the repetition counter _ne is equal to or more than the specified number of repetitions. If _Ne is less than or equal to the specified number of repetitions, 1 is added to the value of _ne . After initializing the value of N _b , the processes of steps S11 to S15 are repeated again. This process is repeated and the process ends when the value of _{ne becomes N e or} more. Hereinafter, this series of flows is referred to as learning.

The control unit uses the input layer as the identification target data and the target output as the identification target data, and trains the neural network to extract important parameters expressing the identification target data from the identification target data and is close to the identification target data. Reconstruction data can be obtained. When the reconstructed data close to the identification target data can be obtained, the error value calculated for the learned data becomes smaller.

Next, data identification is performed by operating in the identification mode (step 2). The operation flow in the case of operating in the identification mode will be described with reference to FIG. 7. First, in step S20, the data acquisition unit 10 acquires data to be identified. Next, in step S21, the feature amount 2 is extracted from the identification target data acquired in step S20 by the first neural network of the feature amount extraction unit 11, and in step S22, the second neural network of the data reconstruction unit 12 is extracted. Create reconstruction data by. In step S23, the difference between the reconstructed data created by the data evaluation unit 13 and the identification target data is evaluated. In step S24, the determination unit 15 compares the error calculated in step S23 with the threshold value, and makes a determination. In step S25, the result output unit 16 outputs the determination result obtained in step S24. This series of flows is called identification. As the data acquired in step S20, real-time data may be acquired by a camera, a microphone, or the like, data acquired in advance may be read, and identification processing may be performed on them.

<Data identification method>
The data identification method performed by the data identification device of the present invention is performed by the following three methods.

-Data identification method A
The first data identification method operates in the learning mode using the learning data acquired in advance, and the parameter determination unit 14 determines the parameters of the feature amount extraction unit 11 and the data reconstruction unit 12. (Step 1). After that, it is operated in the identification mode using the determined parameters (step 2). At this time, the identification target data identified by the identification mode may be data-identified with respect to the data sequentially acquired by the data acquisition unit 10, or may be acquired in advance and stored inside or outside the arithmetic unit 1. Data identification may be performed by acquiring the identification target data group.

This data identification method is suitable for data in which the identification target data is always acquired under the same conditions or data acquired in the past.

-Data identification method B
The second data identification method operates the learning mode and the identification mode in parallel. At this time, the evaluation result calculated by the data evaluation unit 13 is transmitted to the parameter determination unit 14 and the determination unit 15, and the determination and the parameter correction can be performed at the same time. The modified parameters can be updated before executing a series of flows from the next data acquisition, but after learning a certain amount of identification target data, the feature quantity extraction unit 11 and the data reconstruction unit 12 Update the parameters used. This data identification method is effective for a system in which the production process is always in operation. In this data identification method, since the learning mode and the identification mode are operated in parallel, the state change can be more preferably dealt with than the data identification method A. When the learning mode and the identification mode are operated in parallel, the calculation load on the arithmetic unit becomes high, and if it affects the operation of the identification mode, two arithmetic units are prepared, and the learning mode and the identification mode are performed in parallel in each. It is also possible to operate it. In this case, since the operation of the learning mode does not affect the identification mode, the operation stability of the identification mode can be ensured.

・ Data identification method C
In the third data identification method, the control unit 17 operates the identification mode when the production process is in operation, especially when the identification target data is acquired from the production process and the production process is repeatedly operated and stopped. This is a method of recording the identification target data in a recording device inside or outside the arithmetic unit 1 and operating the learning mode when the production process is stopped. When the parameters obtained in the learning mode are used in the identification mode, it is desirable to prepare test data, confirm the obtained evaluation values in advance, make a judgment, and then update the parameters of the neural network used in the identification mode. This data identification method is suitable because the arithmetic unit 1 can be efficiently utilized when the production process is repeatedly stopped and operated. In this data identification method, the identification target data acquired when operating in the identification mode is saved, accumulated data is created, and this accumulated data is used when operating in the learning mode to perform learning. It is possible to reflect the latest production status at the next operation. As a result, even if a change occurs over time or suddenly in the production process, the change state can be reflected in the next and subsequent productions.

Hereinafter, the present invention will be specifically described with reference to Examples. However, the present invention is not limited to this.

<Example 1>
In this embodiment, the data identification method A is used. In this embodiment, according to the configuration shown in FIG. 8, the image of the product to be conveyed is acquired by the camera corresponding to the data acquisition device 2, and the appearance inspection of the product is performed using the data identification program stored in the PC as the arithmetic unit 1. It is carried out and the result is displayed on the display which is the display device 3. Here, the operation is performed in the identification mode for the identification target data acquired from the products that are continuously produced from the time T1 to the time T6 and transported by using the identification device learned by using the accumulated data up to the time T1. It was decided to identify whether the product to be transported was normal or abnormal.

At this time, a slight decrease in the amount of light occurred due to deterioration of the lighting. The captured image changed as shown in FIG. 9 at the time from T1 to T6 due to the change in the imaging condition generated by the decrease in the amount of light, and the time change of the evaluation value became as shown in FIG. Further, the deterioration of the lighting did not occur before the time T1. Since the identification device was trained to learn the image before the time T1, even if the product is normal, the value of the evaluation value calculated by the evaluation unit 13 from the change in the imaging condition occurring in the captured image is the time T1 in FIG. It was decided to gradually rise from time to time T3. As the conditions for acquiring the identification target data continued to change as shown in the figure, the difference between the accumulated data used for learning and the acquired identification target data became large, and the evaluation value continued to increase. For this reason, the evaluation value became equal to or higher than the threshold value after the time T5, but since the determination was normally made up to the aging T5, the change in the imaging conditions is gradual or the continuous production time is short. In some cases, the data identification method A can sufficiently identify whether the product is normal or abnormal.

<Example 2>
In this embodiment, the data identification method B is used. Here, using the configuration of FIG. 8, the products are continuously produced at times T1 to T6, images of the products to be transported are acquired, and the appearance inspection of the products is performed. In this embodiment, PC1 and PC2 are used as the arithmetic unit 1. Similar to the first embodiment, the PC1 is operated in the identification mode for the identification target data acquired from the camera, and the PC2 sequentially uses the identification target data group stored in the PC1 as the accumulated data at a predetermined time. It shall be operated in the learning mode in.

Here, since the data identification method B is used, it is assumed that the learning is performed at predetermined times T2 and T4. Further, as in Example 1, since the amount of light was slightly reduced due to the deterioration of the lighting after the time T1, the obtained identification target data was as shown in FIG. 9, and the time change of the evaluation value was as shown in FIG. ..

When the appearance inspection of the product was started from the time T1, the evaluation value continued to increase from the time T1 to the time T3 as in the first embodiment as shown in FIG. Here, at time T2, learning is performed using the accumulated data from time T1 to T2. By learning with the accumulated data of the latest time in this identification device, the difference in lighting conditions between the accumulated data used for learning and the acquired identification target data could be reduced. By updating the parameters of the feature amount extraction unit 11 and the data reconstruction unit 12 at the time T3 with the parameters obtained by the learning at this time, the value of the evaluation value is lowered after the time T3 than before the time T3. Was done. Therefore, it is possible to suppress the increase in the evaluation value after the time T3. Similarly, by learning again at time T4 and updating the parameters of the feature amount extraction unit 12 and the data reconstruction unit 13 at time T5, it became possible to suppress the increase in the evaluation value. As a result, at times T1 to T6, the evaluation value did not exceed the threshold value, and the determination could be performed accurately.

<Example 3>
In this embodiment, the data identification method C is used. Here, as in the first and second embodiments, the images of the products transported according to the configuration of FIG. 8 are acquired, and the appearance inspection of the products is performed. Especially in this example, the product was produced intermittently. Specifically, production is carried out from time T1 to time T2, time T3 to time T4, and time T5 to time T6, and the product is transported. Production is carried out from time T2 to time T3 and from time T4 to time T5. The delivery of the product was stopped. In this embodiment, the learning mode is operated during the product transport stop time.

Here, the identification target data obtained is as shown in FIG. 9 and the time change of the evaluation is as shown in FIG. 12 because a minute decrease in the amount of light occurs due to the deterioration of the lighting after the time T1. When the appearance inspection of the product was started from the time T1, the evaluation value continued to increase from the time T1 to the time T2. However, by using the learning mode of this data identification device for the accumulated data acquired between the time T1 and the time T2 between the time T2 and the time T3, the accumulated data used for the learning and the acquired identification target are used. Since the difference in lighting conditions in the data became smaller, the value of the evaluation value could be reduced. Therefore, it is possible to suppress the increase in the evaluation value after the time T3. Similarly, by performing learning from time T4 to time T5, it was possible to accurately perform the determination even after time T5.

From the above, the data identification method A is effective when it does not change due to accumulated data or the like, but the acquired data changes over time as in Example 1, and the time is such that continuous inspection is continued without learning. On the other hand, if the degree of change over time of the data is large, a problem may occur in the identification accuracy. Therefore, it is preferable to use the data identification method B and the data identification method C when the acquired data changes with time or when the acquired data is affected by the update of the production equipment or peripheral equipment. In particular, when the production equipment carries out continuous production, it is necessary to operate in the identification mode at all times, so it is necessary to operate in the learning mode in parallel with the identification mode. Therefore, it is preferable to use the data identification method B used in Example 2. When the production equipment is batch production, the arithmetic unit can be effectively utilized by performing learning when the equipment is stopped, so it can be said that it is desirable to use the data identification method C as in the third embodiment.

INDUSTRIAL APPLICABILITY The present invention is useful for detecting abnormalities in a transported product, resin or thread transported by a transport belt in a production process.

1 Arithmetic logic unit 2 Data acquisition device 3 Display device 10 Data acquisition unit 11 Feature quantity extraction unit 12 Data reconstruction unit 13 Data evaluation unit 14 Parameter determination unit 15 Judgment unit 16 Result output unit 17 Control unit

Claims (4)

A data identification device for a production process that identifies whether or not the data acquired and input in the production process is peculiar data in the data group.
A feature amount extraction means for extracting a feature amount from the data acquired and input in the production process using the first neural network, and
A data reconstruction means for generating reconstruction data having the same size as the input data using the second neural network from the feature amount,
A parameter determining means for determining the parameters of the first neural network and the parameters of the second neural network by the error back propagation method from the error between the reconstructed data and the input data.
A data evaluation means for evaluating the magnitude of the error between the reconstructed data and the input data, and
An identification means for discriminating whether or not the input data is singular data by comparing the evaluation value calculated by the evaluation means with a preset threshold value.
A learning mode in which the feature amount extraction means, the data reconstruction means, and the parameter determination means are operated, and a data identification mode in which the feature amount extraction means, the data reconstruction means, the data evaluation means, and the identification means are operated. It is equipped with a control means for operating the two modes of, independently or simultaneously .
The control means
While continuously operating the data identification mode, the learning mode is repeatedly operated at a predetermined cycle.
The data used in the learning mode is acquired and accumulated for a certain period of time obtained in the process of the data identification mode in which the learning mode is continuously operated between the learning mode and the learning mode operated immediately before the learning mode. Data
Control is performed so that the parameters set in the first neural network and the second neural network in the data identification mode are the parameters determined by the learning mode operated most recently in the data identification mode.
Data identification device for production process . It is a data identification device that identifies whether or not the input data is singular data in the data group.
A feature amount extraction means for extracting a feature amount from the input data using the first neural network,
A data reconstruction means for generating reconstruction data having the same size as the input data using the second neural network from the feature amount,
A parameter determining means for determining the parameters of the first neural network and the parameters of the second neural network by the error back propagation method from the error between the reconstructed data and the input data.
A data evaluation means for evaluating the magnitude of the error between the reconstructed data and the input data, and
An identification means for discriminating whether or not the input data is singular data by comparing the evaluation value calculated by the evaluation means with a preset threshold value.
A learning mode in which the feature amount extraction means, the data reconstruction means, and the parameter determination means are operated, and a data identification mode in which the feature amount extraction means, the data reconstruction means, the data evaluation means, and the identification means are operated. It is equipped with a control means for operating the two modes of, independently or simultaneously.
The control means
The learning mode is repeatedly operated at a predetermined cycle, and the data identification mode is continuously operated while the learning mode is not operated.
The data used in the learning mode is acquired and accumulated for a certain period of time obtained in the process of the data identification mode in which the learning mode is continuously operated between the learning mode and the learning mode operated immediately before the learning mode. Data
Control is performed so that the parameters set in the first neural network and the second neural network in the data identification mode are the parameters determined by the learning mode operated most recently in the data identification mode.
Data identification device. By having the first neural network and the second neural network learn the data acquired and accumulated for a certain period in the production process, it is possible to identify whether or not the data currently being acquired is singular data in the data group. It is a data identification method
Feature 1 is extracted from the data acquired and accumulated for a certain period in the production process using the first neural network, and the same size as the accumulated data is extracted from the feature 1 to the second neural network. Reconstruction data is generated, and the parameters of the first neural network and the parameters of the second neural network are determined by the error back propagation method from the error between the reconstruction data and the accumulated data. Step 1 and
Using the first neural network in which the parameters determined in the step 1 are set, the feature amount 2 is extracted from the data being acquired in the production process, and the parameters determined in the step 1 are set. Using the second neural network, reconstruction data having the same size as the data being acquired is generated from the feature amount 2, and the acquisition is based on the magnitude of the error between the reconstruction data and the data being acquired. It has a step 2 of identifying whether or not the data in the data is singular data.
While performing the step 2 continuously, the step 1 is repeated at a predetermined cycle.
The data used in the step 1 acquired and accumulated for a certain period of time is obtained in the process of the step 2 continuously performed between the step 1 and the step 1 performed immediately before the step 1. Data
A data identification method in which the parameters set in the first neural network and the second neural network in the step 2 are the parameters determined by the step 1 performed most recently in the step 2, respectively. A data identification method for discriminating whether or not the data currently being acquired is singular data in a data group by having the first neural network and the second neural network learn the data acquired and accumulated for a certain period of time. And
Feature 1 is extracted from the data acquired and accumulated for a certain period of time using the first neural network, and the reconstruction data of the same size as the accumulated data is extracted from the feature 1 to the second neural network. And the step 1 of determining the parameters of the first neural network and the parameters of the second neural network by the error back propagation method from the error between the reconstructed data and the accumulated data. ,
Using the first neural network in which the parameters determined in the step 1 are set, the feature amount 2 is extracted from the data being acquired, and the second one in which the parameters determined in the step 1 are set. Using a neural network, reconstruction data of the same size as the data being acquired is generated from the feature amount 2, and the data being acquired is based on the magnitude of the error between the reconstruction data and the data being acquired. Has a step 2 of identifying whether or not is singular data.
The step 1 is repeated at a predetermined cycle, and the step 2 is continuously performed while the step 1 is not performed.
The data used in the step 1 acquired and accumulated for a certain period of time is obtained in the process of the step 2 continuously performed between the step 1 and the step 1 performed immediately before the step 1. Data
A data identification method in which the parameters set in the first neural network and the second neural network in the step 2 are the parameters determined by the step 1 performed most recently in the step 2, respectively.

Images