JP2020086786A - Detection device and machine learning method - Google Patents

Abstract

To provide a detection device supporting selection of an algorithm and adjustment of parameters of feature extraction used for preprocessing of data utilized for detection processing.SOLUTION: A detection device 1 of the present invention comprises: a plurality of feature extraction units 36 extracting data according to feature of data from the data according to an industrial machinery 2; a preprocessing unit 34 generating a pair of learning data and evaluation data from the data according to the industrial machinery 2 using each of the feature extraction units 36, respectively; a learning unit 110 performing machine learning based on each of the learning data to generate a plurality of learning models, respectively; an estimation unit 120 estimating with the learning models generated by using the learning data corresponding to the evaluation data, based on the evaluation data; and an evaluation unit 38 evaluating the accuracy of learning/estimation of the learning models based on the estimation result by the estimation unit 120, and outputting the evaluation result of the corresponding feature extraction unit 36 based on the evaluation result.SELECTED DRAWING: Figure 2

Description

The present invention relates to a detection device and a machine learning method.

At manufacturing sites such as factories, non-defective products/defective products of manufactured products are detected, objects of products flowing in a line are detected, and failures of industrial machines are detected. Such various detection operations have conventionally been performed visually by an experienced operator or with reference to the values detected by the sensor. However, in the manual detection work, there arises a problem in that the detection accuracy of the detection work varies due to a difference in judgment criteria based on a difference in experience of each worker, lack of concentration due to a change in physical condition, and the like. For this reason, many manufacturing sites have introduced a detection device for various detection operations, which makes a mechanical automatic determination based on a value detected by a sensor or the like.

In many cases, not only numerical data but also non-numeric data such as images, signals, and characters are handled as data used by the detection device installed in the manufacturing site for automatic determination. When such non-numeric data is used for automatic judgment, it cannot be input as it is into the judgment formula, etc., so it is subjected to the feature extraction processing to convert it to numeric data indicating the characteristic part of the data. Used for automatic judgment.

As a conventional technique related to a detection device used at a manufacturing site or the like, for example, Patent Documents 1 to 3 disclose a technique of performing a detection process using a technique such as machine learning.

JP, 2017-097718, A JP, 2017-076289, A JP, 2005-185149, A

A deep learning method has been used in the detection processing by machine learning. If deep learning is used in the detection process, learning can be performed by using the structure of a multi-layer neural network for large-scale labeled data and automatically extracting and extracting feature parts from the data for learning and estimation. Will be able to. However, in order to perform such a process, it is necessary to perform a large-scale process using a GPU (Graphical Processing Unit) or the like, which causes a problem that the device performance and the calculation time are heavy.

On the other hand, if feature detection processing specialized for data is performed in the pre-processing stage and the detection device is configured to use a small-scale machine learning machine for learning and estimation, deep learning technology automatically The calculation load is often smaller than that when extraction is performed. When performing such feature extraction processing, in order to extract feature data suitable for learning from each data, an appropriate feature extraction algorithm is selected, and the parameters of the selected feature extraction algorithm are appropriately set. By adjusting, the accuracy can be improved while suppressing the calculation load of learning and estimation in the machine learning device. However, when manually selecting an appropriate feature extraction algorithm and appropriately adjusting parameters, optimization takes time, and there is a problem that the basis for selecting algorithms and adjusting parameters becomes ambiguous. Further, even when a new feature extraction method is examined, it is necessary to verify learning and inference accuracy improvement with respect to the existing method, but this work also requires man-hours.

Therefore, an object of the present invention is to provide a detection device that supports selection of a feature extraction algorithm used for preprocessing of data used for detection processing and adjustment of parameters.

According to the present invention, a non-numerical data (image, signal, character, etc.) is detected by a plurality of methods in a detection device that performs a feature extraction (preprocessing) step from data and a learning or inference (learning core) step. Preprocessing (edge detection, spectrum analysis, etc.), and learning or inference (unsupervised classification, regression with teacher, etc.) with the combined result (N-dimensional vector) automatically selects a preprocessing method that improves inference accuracy. The above problem is solved by providing a function (providing accuracy equivalent to that when selecting the optimum extraction method).

An aspect of the present invention is a detection device that performs a detection process based on data related to an industrial machine, and includes a plurality of feature extraction units that extract data related to the features of the data from the data related to the industrial machine. , A pre-processing unit that respectively creates a set of learning data and evaluation data from the data related to the industrial machine using each feature extraction unit, and performs machine learning based on each of the learning data to generate a plurality of learning models. Based on the estimation part by the estimation part by which the learning part each produced|generated and the said evaluation data, the estimation using the learning model produced|generated using the learning data corresponding to the said evaluation data, and the said estimation part, An evaluation unit that evaluates the accuracy of learning or estimation of the learning model and outputs the evaluation result of the corresponding feature extraction unit based on the evaluation result.

Another aspect of the present invention is a machine learning method related to a detection process based on data related to an industrial machine, and a plurality of feature extraction methods for extracting data related to a feature of the data from the data related to the industrial machine, respectively. A pre-processing step of creating a set of learning data and evaluation data from the data related to the industrial machine, and a learning step of performing machine learning based on each of the learning data and generating a plurality of learning models, respectively. Based on the evaluation data, an estimation step of performing estimation using a learning model generated using learning data corresponding to the evaluation data, and learning or estimation of the learning model based on the estimation result of the estimation step. Is performed, and an evaluation step of outputting the evaluation result of the corresponding feature extraction unit based on the evaluation result is performed.

According to the present invention, it becomes possible to perform learning or inference in a relatively short time even in a device having low device performance. Further, since it becomes possible to select the optimum parameter in a short time in a fixed procedure regardless of the ability and experience of the worker, it is expected that the inference accuracy will be improved. Furthermore, when introducing the new pretreatment method, it is sufficient to simply add the new pretreatment method to the existing processing group, and therefore the number of verification steps can be reduced.

It is a schematic hardware block diagram of the detection apparatus by one Embodiment. FIG. 3 is a schematic functional block diagram of a detection device according to an embodiment. It is a figure which shows the example of the feature extraction part with which a preprocessing part is provided. It is a figure which shows the other example of the feature extraction part with which a preprocessing part is equipped. It is a figure which shows the example of preparation of the learning data and evaluation data by a pre-processing part. It is a figure which shows the other example of preparation of the learning data and evaluation data by a preprocessing part. It is a figure explaining the flow of evaluation of the feature extraction part by an evaluation part.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a schematic hardware configuration diagram showing a main part of a detection device including a machine learning device according to an embodiment. The detection device 1 of the present embodiment can be mounted on a control device that controls an industrial machine 2 such as a transport machine or a robot. Further, the detection device 1 of the present embodiment includes a personal computer installed together with a control device that controls the industrial machine 2 such as a transport machine or a robot, an edge computer connected to the control device via a wired/wireless network, It can be implemented as a computer such as a cell computer, a host computer, or a cloud server. In the present embodiment, an example in which the detection device 1 is implemented as a computer connected to a control device that controls an industrial machine 2 such as a transport machine or a robot via a wired/wireless network is shown.

The CPU 11 included in the detection device 1 according to the present embodiment is a processor that controls the detection device 1 as a whole. The CPU 11 reads out the system program stored in the ROM 12 via the bus 20 and controls the entire detection device 1 according to the system program. The RAM 13 temporarily stores temporary calculation data, various data input by the operator via the input device 71, and the like.

The non-volatile memory 14 is composed of, for example, a memory backed up by a battery (not shown), an SSD (Solid State Drive), or the like, and retains the storage state even when the power of the detection device 1 is turned off. In the non-volatile memory 14, a setting area in which setting information related to the operation of the detection device 1 is stored, data input from the input device 71, data acquired from the industrial machine 2 (image data, audio data, time series). Data, character data, etc.), data acquired from the machine learning device 100, data read via an external storage device (not shown) or a network, and the like are stored. The programs and various data stored in the non-volatile memory 14 may be expanded in the RAM 13 at the time of execution/use. A system program including a known analysis program for analyzing various data is written in the ROM 12 in advance.

The industrial machine 2 can be, for example, a machine tool, a carrier machine, a robot, a mining machine, a woodworking machine, an agricultural machine, a construction machine, or the like. The industrial machine 2 is configured such that the industrial machine 2 can acquire information related to the operation of each unit such as a prime mover. Further, a sensor such as an image sensor or a voice sensor is separately attached to the industrial machine 2. It is configured so that information required for the operation of the machine 2 can be acquired. The information acquired by the industrial machine 2 is acquired by the detection device 1 via the wired/wireless network 5 and the interface 16 and stored in the RAM 13, the nonvolatile memory 14, or the like. Further, the detection device 1 outputs a predetermined control signal to the industrial machine 2 via the interface 16 and the network 5 as necessary.

On the display device 70, each data read in the memory, data obtained as a result of execution of a program, etc., data output from the machine learning device 100 described later, etc. are output via the interface 17 and displayed. To be done. Further, the input device 71 including a keyboard, a pointing device, and the like passes commands, data, and the like based on the operation of the operator to the CPU 11 via the interface 18.

The interface 21 is an interface for connecting the detection device 1 and the machine learning device 100. The machine learning device 100 includes a processor 101 that controls the entire machine learning device 100, a ROM 102 that stores a system program and the like, a RAM 103 that temporarily stores each process related to machine learning, and a storage of a learning model and the like. A non-volatile memory 104 used for The machine learning device 100 can observe each information that can be acquired by the detection device 1 via the interface 21 (for example, image data, voice data, time series data, character data, etc. acquired from the industrial machine 2). .. Further, the detection device 1 acquires the processing result output from the machine learning device 100 via the interface 21, stores and displays the acquired result, and displays the acquired result to another device via the network 5 or the like. Send.

FIG. 2 is a schematic functional block diagram of the detection device 1 and the machine learning device 100 according to the embodiment. The detection device 1 of this embodiment has a configuration required when the machine learning device 100 performs learning (learning mode). In each functional block shown in FIG. 2, the CPU 11 included in the detection device 1 shown in FIG. 1 and the processor 101 of the machine learning device 100 execute respective system programs, and the detection device 1 and the machine learning device 100 have the same function. It is realized by controlling the operation of each unit.

The detection device 1 of the present embodiment includes a data acquisition unit 30, a data set generation unit 32, a preprocessing unit 34, and an evaluation unit 38, and the preprocessing unit 34 includes a plurality of feature extraction units 36. Further, the acquired data storage unit 50 acquired by the data acquisition unit 30 is provided on the nonvolatile memory 14, and the acquired data storage unit 50 acquired by the data acquisition unit 30 is constructed on the nonvolatile memory 104 of the machine learning device 100 by machine learning by the learning unit 110. A learning model storage unit 130 for storing the learning model is provided.

The data acquisition unit 30 is a functional unit that acquires various data from the industrial machine 2, the input device 71, and the like. The data acquisition unit 30 includes, for example, image data of a detection target acquired from the industrial machine 2, operation noise of the industrial machine 2, voltage value/current value of the electric motor, temperature distribution data of each part of the industrial machine 2, and the like. A set of label values given by the operator to the data is acquired and stored in the acquired data storage unit 50. The data acquisition unit 30 may acquire data from another device via an external storage device (not shown) or a wired/wireless network.

The data set generation unit 32 sets a predetermined number or a predetermined ratio of a plurality of data stored in the acquired data storage unit 50 as a learning data set, and other data as an evaluation data set. Is a functional means of dividing as. It is desirable that each of the sub-data sets divided by the data set generation unit 32 be divided so that bias is unlikely to occur, and for example, data to be used as a learning data set from the data stored in the acquired data storage unit 50. May be randomly selected or may be divided into each data set by using a statistical sampling method such as a systematic sampling method.

The preprocessing unit 34 is a functional unit that performs data conversion by the feature extraction unit 36 on the learning data set and the evaluation data set generated by the data set generation unit 32. The pre-processing unit 34 includes a plurality of feature extraction units 36, and sequentially inputs each data included in the data set to the feature extraction unit 36 according to a preset pre-processing requirement to perform data conversion.

The preprocessing requirement set in the preprocessing unit 34 defines the conversion order of the data included in the data set by the one or more feature extraction units 36. An example of the preprocessing requirement is a procedure of applying the feature extraction unit 36 that performs so-called corner detection processing on image data included in a data set. Further, as another example of the preprocessing requirement, after applying the feature extraction unit 36 that performs a predetermined level correction process on the image data included in the data set, the feature extraction unit 36 that performs the edge detection process is applied. There are procedures. Furthermore, as another example of the preprocessing requirement, a procedure of applying the feature extraction unit 36 that performs the noise removal process to the voice data included in the data set and then the feature extraction unit 36 that performs the sound feature extraction process. Is mentioned. The preprocessing requirements are set in advance in the preprocessing unit 34 according to the type of data detected by the detection device and the purpose of detection.

The feature extraction unit 36 included in the preprocessing unit 34 is a functional unit that extracts a feature portion of each data included in the data set for learning or estimation. The plurality of feature extraction units 36 are each for performing a predetermined feature extraction process, and those performing the same type of feature extraction process have different parameters or algorithms used in the process.

FIG. 3 is a diagram illustrating an example of the feature extraction unit 36 that performs the feature extraction processing of image data. In the example of FIG. 3, the preprocessing unit 34 includes the feature extraction units A1 to A3 as the feature extraction unit 36 that performs the feature detection process for corner detection, and all of them are corner detection algorithms AKAZE (Accelerated KAZE algorithm). The sensitivity parameter is set to 1.0 times the standard sensitivity, the characteristic extraction unit A2 is set to 1.5 times, and the characteristic extraction unit A3 is set to 2.0 times the standard sensitivity.

In addition, FIG. 4 is a diagram illustrating an example of the feature extraction unit 36 that performs feature extraction processing of audio data. In the example of FIG. 4, the preprocessing unit 34 includes the feature extraction units B1 to B3 as the feature extraction unit 36 that performs the feature extraction process of noise removal, and all of them use the spectral subtraction (SS) method as the noise removal algorithm. The window width parameter is set to 1.0 times the standard window width for the feature extraction unit B1, 1.5 times for the feature extraction unit B2, and 2.0 times for the feature extraction unit B3. Further, the preprocessing unit 34 includes feature extraction units C1 and C2 as the feature extraction unit 36 that performs the feature extraction process of the sound feature extraction, and the feature extraction unit C1 uses the mel frequency cepstrum coefficient (MFCCC) as the sound feature extraction algorithm. ), and the feature extraction unit C2 uses an FFT (Fast Fourier Transform) and a cepstrum as a sound feature extraction algorithm.

The preprocessing unit 34 uses the plurality of feature extraction units 36 prepared according to the type of data to be preprocessed as described above, and according to the preset preprocessing requirements, the respective feature extraction units 36. Generate learning data by applying. FIG. 5 is a diagram illustrating a process performed by the preprocessing unit 34 including the feature extraction units A1 to A3 illustrated in FIG. In the example of FIG. 5, the preprocessing unit 34 creates learning data a1 to a3 by converting each data included in the learning data set generated by the data set generation unit 32 using the feature extraction units A1 to A3. Also, each data included in the evaluation data set is converted by using the feature extraction units A1 to A3 to create evaluation data a1 to a3.

FIG. 6 is a diagram illustrating a process performed by the preprocessing unit 34 including the feature extraction units B1 to B3 and C1 to C2 illustrated in FIG. The pre-processing unit 34 is set as a pre-processing requirement to perform noise removal processing on voice data in advance and then execute sound feature extraction processing. In the example of FIG. 6, the preprocessing unit 34 creates a set of the feature extraction unit selected from the feature extraction units B1 to B3 and the feature extraction unit selected from the feature extraction units C1 to C2 according to the preprocessing requirements. , The learning data sets b1 to a6 are created from the learning data set and the evaluation data sets b1 to a6 are created from the evaluation data set.

The learning unit 110 performs machine learning using each set of learning data created by the preprocessing unit 34 and generates a plurality of learning models. The learning unit 110 uses each of a plurality of sets of learning data acquired from the industrial machine 2 and preprocessed by the preprocessing unit 34 by a known machine learning method such as unsupervised learning and supervised learning. To generate a plurality of learning models and store the generated plurality of learning models in the learning model storage unit 130. The unsupervised learning method performed by the learning unit 110 is, for example, the k-means method, and the supervised learning method is, for example, a general neural network or SVM.

On the other hand, the estimation unit 120 estimates the state using the learning model stored in the learning model storage unit 130 based on each set of evaluation data created by the preprocessing unit 34. In the estimation unit 120 of the present embodiment, for the set of evaluation data created by the pre-processing unit 34 using the predetermined feature extraction unit 36, the set of learning data created by the same feature extraction unit 36 is used. The estimation using the learning model generated by the learning unit 110 is performed.

The evaluation unit 38 evaluates the learning or inference accuracy of each learning model based on the result of the inference with respect to the evaluation data estimated by the estimation unit 120, and generates the learning model based on the evaluation result. It is a functional unit that evaluates the feature extraction unit 36 used at that time. The evaluation unit 38 compares the inference result by the estimation unit 120 with respect to each evaluation data included in the evaluation data set with the label value assigned to each evaluation data, and thereby the inference result of the estimation unit 120 is obtained. Determine correctness. Next, the evaluation unit 38 uses a known statistical method (for example, the minimum circumstance error amount for the label) for the inference result regarding each evaluation data, and a known machine learning evaluation index such as an ROC curve and AUC. Then, the accuracy of learning or inference of the learning model used for the inference is evaluated. Then, the evaluation unit 38 evaluates the feature extraction unit 36 used when generating the learning model based on the accuracy of learning or inference of the learning model. The evaluation unit 38 performs such processing on each learning model stored in the learning model storage unit 130. The evaluation result by the evaluation unit 38 may be displayed and output on the display device 70, or may be transmitted and output to a host computer, a cloud computer or the like via the network 5 for use. The evaluation unit 38 may perform display output or the like only for the feature extraction unit 36 used for generating the learning model with the highest learning or inference accuracy.

FIG. 7 is a diagram showing an example of evaluation by the evaluation unit 38 based on the learning data and the evaluation data created based on the feature extraction unit shown in the examples of FIGS. As illustrated in FIG. 7, it is assumed that the learning model a1 is generated by the learning unit 110 based on the learning data a1 created using the feature extraction unit A1. At this time, the estimation unit 120 performs the estimation process using the learning model a1 based on the evaluation data a1 created using the feature extraction unit A1 and outputs the estimation result a1. Then, the evaluation unit 38 applies a known statistical process or a known machine learning evaluation index to the estimation result a1 to evaluate the accuracy of the learning model a1, and outputs the result as the evaluation of the feature extraction unit A1. To do. Similarly, the evaluation unit 38 evaluates the learning models a2 and a3 from the estimation results a2 and a3, and outputs the results as the evaluations of the feature extraction units A2 and A3, respectively.

The detection device 1 according to the above-described embodiment creates learning data and evaluation data in which the features of the data are extracted by the plurality of feature extraction units 36 with respect to the data acquired for performing the detection, and the created learning data is obtained. Each feature extraction unit 36 is evaluated from the estimation result obtained by performing learning and estimation based on the data and the evaluation data. The evaluation result output by the evaluation unit 38 indicates the suitability of each feature extraction unit 36 with respect to the data acquired to perform the detection, and the operator who has grasped this shows the feature extraction unit 36 to be used for the detection process. It is possible to easily determine the set of the feature extraction unit 36 (that is, the parameter or algorithm used for the preprocessing, the set thereof). Further, even when the feature extraction unit 36 is newly added, it is possible to easily verify the improvement in learning/estimation accuracy with respect to the existing feature extraction unit 36.

As a modification of the above-described embodiment, the evaluation unit 38 newly adds the feature extraction unit 36 whose parameters are adjusted to the preprocessing unit 34 based on the evaluation result of each feature extraction unit 36, and again, Even if the data set generation unit 32, the preprocessing unit 34, the learning unit 110, and the estimation unit 120 are caused to execute the processing using the added feature extraction unit 36, the feature extraction unit 36 may be evaluated. good. When the evaluation unit 38 evaluates each of the feature extraction units 36, no matter which of the feature extraction units 36 is used, the learning that exceeds the predetermined threshold value set in advance or the learning that indicates the accuracy of inference is performed. When a model is not obtained, the change tendency of the parameter value set in each feature extraction unit 36 and the change tendency of the learning or inference accuracy of the learning model generated using the feature extraction unit 36 are described. By comparison, a feature extraction unit 36 in which a parameter value that is expected to increase the learning accuracy of the learning model is set is newly added, and each functional unit is instructed to perform processing using the added feature extraction unit 36. To do.

By causing the evaluation unit 38 to perform the above operation, when the feature extraction unit 36 prepared in advance does not have the feature extraction unit 36 suitable for the data acquired for the detection. Even if there is, the feature extraction unit 36 whose parameter value is adjusted by the evaluation unit 38 is automatically added and re-evaluated, so that a suitable feature extraction unit 36 is searched semi-automatically. ..

Although the embodiments of the present invention have been described above, the present invention is not limited to the examples of the above-described embodiments and can be implemented in various modes by making appropriate changes.
For example, although the detection device 1 and the machine learning device 100 are described as devices having different CPUs (processors) in the above-described second and third embodiments, the machine learning device 100 includes the CPU 11 and the ROM 12 included in the detection device 1. It may be realized by a stored system program.

Further, although the configuration for learning and the configuration for estimation are described as different embodiments in the second and third embodiments described above, the detection device 1 may be configured to include them at the same time. In this case, the detection device 1 operates so as to update (learn) the learning model as necessary while estimating the attachment state of the tool.

1 Detection Device 2 Industrial Machine 5 Network 11 CPU
12 ROM
13 RAM
14 non-volatile memory 16, 17, 18 interface 20 bus 21 interface 30 data acquisition unit 32 data set generation unit 34 preprocessing unit 36 feature extraction unit 38 evaluation unit 50 acquired data storage unit 70 display device 71 input device 100 machine learning device 101 Processor 102 ROM
103 RAM
104 non-volatile memory 110 learning unit 120 estimation unit 130 learning model storage unit

Claims (5)

A detection device that performs a detection process based on data related to an industrial machine,
A plurality of feature extraction units that extract data relating to the features of the data from the data relating to the industrial machine are provided, and respective sets of learning data and evaluation data are created from the data relating to the industrial machine using the respective feature extraction units. A pre-processing unit,
A learning unit that performs machine learning based on each of the learning data to generate a plurality of learning models,
An estimation unit that performs estimation based on the evaluation data using a learning model generated using learning data corresponding to the evaluation data;
Based on the estimation result by the estimation unit, to evaluate the accuracy of learning and estimation of the learning model, based on the evaluation result, an evaluation unit that outputs the evaluation result of the corresponding feature extraction unit,
A detection device equipped with. The data related to the industrial machine is at least one of image data, time series data, and character data,
The detection device according to claim 1. The machine learning is either unsupervised learning or supervised learning,
The detection device according to claim 1. The evaluation unit outputs an evaluation result of the feature extraction unit corresponding to a learning model with the highest learning/estimation accuracy,
The detection device according to claim 1. A machine learning method for detection processing based on data of an industrial machine,
A preprocessing step of respectively creating a set of learning data and evaluation data from the data relating to the industrial machine by respectively using a plurality of feature extracting means for extracting data relating to the feature of the data from the data relating to the industrial machine,
A learning step of performing machine learning based on each of the learning data and generating a plurality of learning models,
An estimation step of performing estimation using a learning model generated using learning data corresponding to the evaluation data based on the evaluation data;
Based on the estimation result by the estimation step, evaluate the accuracy of learning and estimation of the learning model, based on the evaluation result, an evaluation step of outputting the evaluation result of the corresponding feature extraction means,
A machine learning method with.

Images