JP2020027386A - Apparatus, system and method for determining abnormality - Google Patents

Abstract

To provide an abnormality determination apparatus adapted to perform preliminary processing capable of predicting a determination of normality and abnormality in the case of determining normality and abnormality using a learning model based on machine learning.SOLUTION: An abnormality determination apparatus 100 comprises: a first state data acquiring section 103 that acquires first state data detected by a sensor 11 attached on a processing apparatus 10; a quality data acquiring section 101 that acquires quality data related to normality and abnormality; a first learning model generating section 106 that generate a first learning model for generating second state data obtained by converting the first state data; a second learning model generating section 109 that generates a second learning model related to the second state data and a classification to which the quality data belongs; a first predicting section 111 that predicts the second state data corresponding to first input data using the first learning model; and a second predicting section 112 that predicts a classification to which the quality data corresponding to second input data belongs using the second learning model.SELECTED DRAWING: Figure 2

Description

  The present invention relates to an abnormality determination device, an abnormality determination system, and an abnormality determination method.

  Patent Literature 1 describes an apparatus that detects an abnormality of a tool of a machine tool. The device generates a normal model by machine learning using state data such as vibration information, cutting force information, and sound information, and detects the state based on the state data detected during machining after the normal model is created and the normal model. Diagnose whether the status data is normal or abnormal.

  Patent Literature 2 discloses an apparatus that monitors an abnormality during processing or a failure of a machine tool. The device detects the vibration of the machine tool, extracts the characteristic amount of the vibration before processing and the characteristic amount of the vibration during processing, obtains the classification result of the characteristic amount by applying a neural network, and performs the classification. The presence or absence of an abnormality is determined based on the result.

  Further, Patent Document 3 discloses a numerical control device having a processing condition adjusting function of suppressing occurrence of chatter vibration and tool wear / breakage. The apparatus determines the adjustment amount of the processing condition by machine learning so as to suppress occurrence of chatter vibration and the like.

JP, 2018-24055, A JP 2008-137102 A JP 2017-45300A

  As described in Patent Documents 1 and 2, normality / abnormality can be determined by machine learning using data detected by a sensor attached to a machine tool. However, in Patent Literature 1, machine learning is performed using the measured data itself, and the data has, for example, a very large number of variables. This may cause a decrease in prediction accuracy.

  On the other hand, in Patent Literature 2, as the input data of the neural network, not the vibration data itself but the feature amount of the vibration is used. As a result, as described above, the number of variables is reduced, so that the prediction accuracy of the learning model can be improved. Here, in Patent Document 2, the feature amount is a preset feature amount. However, how to set the feature amount is not easy. Depending on the method of extracting the feature, the prediction accuracy using the learning model may be reduced.

  Further, in the case where normality and abnormality are determined using a learning model by machine learning, various preliminary processes can be performed on the detected state data in addition to the extraction of the feature amount. Then, depending on the type of processing, the prediction accuracy using the learning model may be reduced depending on the processing method, as described above.

  The present invention relates to an abnormality determination that performs a preliminary process that can make a normal / abnormal determination prediction highly accurate when performing a preliminary process of the determination when performing a normal / abnormal determination using a learning model based on machine learning. It is an object to provide an apparatus, an abnormality determination system, and an abnormality determination method.

  An abnormality determination device according to the present invention includes a first state data acquisition unit that acquires first state data detected by a sensor attached to the processing device when the processing device performs a predetermined process; For the normal and abnormal quality data, a quality data acquisition unit that acquires the quality data associated with the first state data, by machine learning the first state data as the first learning data, A first learning model generation unit that generates a first learning model for generating second state data obtained by converting the first state data, and a machine that uses the second state data and the quality data as second learning data. A second learning model generation unit that generates a second learning model related to the classification to which the second state data and the quality data belong by learning; Using the first state data as the first input data, using the first input data and the first learning model, a first prediction unit that predicts the second state data corresponding to the first input data, A second prediction unit that predicts a class to which the quality data corresponding to the second input data belongs by using the second state data as the second input data and using the second input data and the second learning model. .

  According to the abnormality determination device, the second prediction unit determines, using the second state data and the second learning model, a classification to which the quality data belongs, that is, whether the predetermined process is normal or abnormal. . The second state data is not the first state data detected by the sensor attached to the processing device, but is data obtained by converting the first state data. The second state data is data obtained by the first prediction unit using the first state data and the first learning model.

  That is, the first stage processing of generating the second state data using the first state data and the first learning model is performed, and after the first stage processing, the second state data and the second learning model are used to perform normal processing. And a second-stage process for determining abnormality is performed. As described above, the abnormality determination device applies a learning model of machine learning in a plurality of stages. Therefore, it is possible to accurately and correctly determine and predict the normal and abnormal states.

  An abnormality determination system according to the present invention is provided with a plurality of processing devices that execute a predetermined process, and is provided so as to be able to communicate with the plurality of processing devices, and performs the process when each of the plurality of processing devices executes the predetermined process. A server that collects first state data detected by a sensor attached to the apparatus, and the abnormality determination apparatus described above, wherein the abnormality determination apparatus is processed based on the first state data collected by the server And Thus, since a large number of first state data can be collected, it is possible to determine the normal state and the abnormal state with higher accuracy.

  The abnormality determination method according to the present invention includes a first state data acquisition step of acquiring first state data detected by a sensor attached to the processing apparatus when the processing apparatus executes a predetermined processing; For the normal and abnormal quality data, the quality data acquisition step of acquiring the quality data associated with the first state data, by machine learning the first state data as the first learning data, A first learning model generating step of generating a first learning model for generating second state data obtained by converting the first state data, and a machine that uses the second state data and the quality data as second learning data. A second learning model generating step of generating a second learning model for the classification to which the second state data and the quality data belong by learning; The first state data in the first input data, using the first input data and the first learning model, a first prediction step of predicting the second state data corresponding to the first input data, The second state data as the second input data, using the second input data and the second learning model, a second prediction step of predicting a class to which the quality data corresponding to the second input data belongs. Prepare. According to the abnormality determination method, the same effect as that of the abnormality determination device can be obtained.

FIG. 1 is a diagram illustrating a configuration of an abnormality determination system 1. FIG. 2 is a block diagram illustrating an abnormality determination device 100 according to a first example. It is a figure showing the 1st example of the 1st state data. It is a figure showing the 2nd example of the 1st state data. 9 is a flowchart illustrating pre-processing performed by a pre-processing unit. It is a graph which shows first state data. 5 is a graph showing power spectrum information by frequency analysis. It is a graph which shows the power spectrum information after a normalization process. 5 is a graph showing power spectrum information that is output data (second state data) from a first learning model generation unit 106. FIG. 6 is a diagram illustrating first learning data and output data (second state data) in a first learning model generation unit 106. FIG. 8 is a diagram illustrating a first example of second learning data in a second learning model generation unit 109. FIG. 9 is a diagram illustrating a second example of second learning data in a second learning model generation unit 109. 9 is a flowchart illustrating a process performed by a second learning model generation unit 109. FIG. 7 is a block diagram illustrating an abnormality determination device 200 according to a second example. FIG. 7 is a diagram illustrating first learning data (first state data), first learning data (processing conditions), and output data (second state data) in a first learning model generation unit 206. FIG. 9 is a diagram illustrating a first example of second learning data in a second learning model generation unit 209. FIG. 13 is a diagram illustrating a second example of the second learning data in the second learning model generation unit 209.

(1. Application target of abnormality judgment device 100, 200)
The abnormality determination devices 100 and 200 determine whether the processing device 10 that performs the predetermined process is normal or abnormal with respect to the predetermined process performed by the processing device 10. The processing device 10 is, for example, a machine tool for processing a workpiece, a molding machine for molding a molded product, a transport device for transporting a transported product, and the like. The predetermined processing is cutting or grinding of a workpiece, molding of a molded product, transport of a transported object, and the like.

  In the case of cutting or grinding, the target of the abnormality is occurrence of chatter vibration, that is, occurrence of a chatter pattern on an object (workpiece) to be processed. In addition, it covers abnormalities caused by vibration. In the present embodiment, the processing device 10 to which the abnormality determination devices 100 and 200 are applied is, for example, a machine tool that performs cutting or grinding of an object (workpiece).

(2. Configuration of abnormality determination system 1)
The configuration of the abnormality determination system 1 will be described with reference to FIG. The abnormality determination system 1 includes a plurality of processing devices (machine tools) 10, 10, 10, 10 that execute predetermined processing, an inspection device 20, a server 30, and abnormality determination devices 100 and 200. The processing device 10 is, for example, a machining center or a lathe that performs a cutting process, or a grinding machine that performs a grinding process. In FIG. 1, the processing device 10 is a machining center as an example.

  The processing apparatus 10 performs a cutting process on the workpiece W using the processing tool T. The processing device 10 includes a sensor 11. The sensor 11 is, for example, a sound wave sensor that detects a sound wave generated by processing, a vibration sensor that detects vibration generated by processing, a sensor that detects a processing load, and the like. Examples of the sensor for detecting the processing load include a sensor for detecting a current of a motor for rotating and driving the processing tool T, a sensor for detecting a current of a motor for driving a moving body supporting the workpiece W, and a force sensor. And so on.

  The processing device 10 further includes a control device 12 that controls the position and the attitude of the processing tool T with respect to the workpiece W based on various information detected by the sensor 11. The control device 12 performs control based on preset processing conditions. The processing conditions include the type of the processing tool T, the material of the workpiece W, and the processing steps (roughing, finishing, etc.) from the start to the end of the processing.

  The inspection device 20 performs a quality inspection of the target (workpiece) W processed by the plurality of processing devices 10. The quality inspection includes shape inspection, surface roughness inspection, presence or absence of chatter pattern, and the like. The inspection device 20 can also acquire an image of the workpiece W in addition to the measurement.

  The server 30 is provided so as to be able to communicate with the plurality of processing devices 10 and the inspection devices 20. The server 30 collects processing conditions in each of the plurality of processing devices 10 and first state data detected by the sensor 11 attached to the processing device 10 when each of the plurality of processing devices 10 performs a predetermined process. I do. Further, the server 30 collects quality data that is an inspection result by the inspection device 20.

  The abnormality determination devices 100 and 200 are processed based on the processing conditions, the first state data, and the quality data collected by the server 30, and determine whether the predetermined processing by the plurality of processing devices 10 is normal or abnormal.

(3. Another example of abnormality determination system 1)
In FIG. 1, the server 30 collects various data. In addition, each of the plurality of processing devices 10 may be provided with a data collection device (not shown), and each data collection device may collect data on the corresponding processing device 10. In this case, each data collection device can also collect the quality data that is the inspection result by the inspection device 20. Then, the abnormality determination devices 100 and 200 can determine normality and abnormality using the data collected by each data collection device.

  Further, the abnormality determination devices 100 and 200 may be provided in each of the plurality of processing devices 10 similarly to the data collection device. In this case, each of the abnormality determination devices 100 and 200 can determine whether the processing performed by each processing device 10 is normal or abnormal.

(4. Configuration of Abnormality Determination Apparatus 100 of First Example)
The configuration of the abnormality determination device 100 of the first example will be described with reference to FIG. The method (the abnormality determination method) by the abnormality determination device 100 of the first example will also be described.

  The abnormality determination device 100 includes a part that functions in a learning phase of machine learning and a part that functions in an inference phase of machine learning. Even during the execution of the inference phase, the learning model is updated by executing the learning phase. Hereinafter, the learning phase and the inference phase will be described separately. In FIG. 2, solid arrows indicate the flow of processing that functions in the learning phase, and dashed arrows indicate the flow of processing that functions in the inference phase.

(4-1. Learning phase)
As illustrated in FIG. 2, the abnormality determination device 100 includes a quality data acquisition unit 101, a quality data storage unit 102, a first state data acquisition unit 103, a first state data storage unit 104, It includes a processing unit 105, a first learning model generation unit 106, a first learning model storage unit 107, a second state data storage unit 108, a second learning model generation unit 109, and a second learning model storage unit 110.

  When the processing device 10 executes a predetermined process, the quality data obtaining unit 101 obtains quality data related to normality and abnormality of the predetermined process (quality data obtaining step). Specifically, the quality data acquisition unit 101 acquires quality data as an inspection result by the inspection device 20. In the present embodiment, since the quality data is stored in the server 30 (shown in FIG. 1), the quality data acquisition unit 101 acquires the quality data from the server 30. However, the quality data obtaining unit 101 can also obtain the quality data directly from each processing device 10.

  The processing apparatus 10 is a machine tool that performs a cutting process or a grinding process on a workpiece (target object) W as a predetermined process. The quality data is data on the normality and abnormality of the workpiece W. In particular, the quality data is data relating to the degree of chatter pattern generated on the workpiece W.

  The quality data acquired by the quality data acquisition unit 101 is stored in the quality data storage unit 102 (quality data storage step). In the quality data storage unit 102, the quality data is stored in association with the workpiece W. That is, the quality data is stored in association with the first state data described later.

  The first state data acquisition unit 103 acquires first state data detected by the sensor 11 attached to the processing device 10 when the processing device 10 executes a predetermined process (first state data acquisition step). In the present embodiment, since the first state data is stored in the server 30 (shown in FIG. 1), the first state data acquisition unit 103 acquires the first state data from the server 30. However, the first state data acquisition unit 103 can also acquire directly from each processing device 10.

  Here, as described above, the processing device 10 is a machine tool that performs cutting or grinding as a predetermined process on the workpiece W. Then, the first state data is acquired for each workpiece W. The first state data acquired by the first state data acquisition unit 103 is stored in the first state data storage unit 104 in association with the workpiece W (first state data storage step).

  Although the quality data storage unit 102 and the first state data storage unit 104 are separate storage units (databases) as an example, they may be integrated storage units (databases). When stored in the integrated storage unit, the quality data and the first state data are stored in association with the respective workpieces W.

  Here, the first state data includes, as shown in FIG. 3, waveform data on any of a sound wave, a vibration, and a processing load generated due to the cutting processing or the grinding processing. That is, the first state data is data indicating a behavior with the passage of time from the start to the end of machining one workpiece W.

  Note that, as the first state data, waveform data relating to all of sound waves, vibrations, and processing loads can be used. Here, when the sensor 11 is a sound wave sensor, the first state data is waveform data on sound waves. When the sensor 11 is a vibration sensor, the first state data is waveform data on vibration. When the sensor 11 is a sensor that detects a processing load, the first state data is waveform data relating to the processing load.

  As shown in FIG. 3, the first state data may include waveform data on the surface roughness of the workpiece W in addition to the waveform data or in place of the waveform data. The waveform data related to the surface roughness of the workpiece W is usually information obtained by inspection by the inspection device 20. However, when a sensor capable of measuring the surface roughness is provided inside the processing apparatus 10, the sensor attached to the processing apparatus 10 acquires waveform data relating to the surface roughness.

  Further, the first state data may include non-waveform data relating to an image of the surface of the workpiece W in addition to the waveform data as shown in FIG. The non-waveform data relating to the image may be data capable of determining a chatter pattern based on luminance or data capable of determining a chatter pattern based on color. For example, when a chatter pattern exists on the surface of the workpiece W, the data relating to the image is luminance data or color data when scanned in the direction of the chatter pattern. As a result, the data regarding the image becomes data from which the chatter pattern can be determined.

  The preprocessing unit 105 performs a predetermined preprocessing on the first state data stored in the first state data storage unit 104 to convert the data format (preprocessing step). Then, the preprocessing unit 105 stores the converted first state data in the first state data storage unit 104.

  The pre-processing performed by the pre-processing unit 105 will be described with reference to FIG. The preprocessing unit 105 acquires first state data as waveform data stored in the first state data storage unit 104 (S11). The first state data as the waveform data is represented as a value with respect to time as shown in FIG.

  Subsequently, the preprocessing unit 105 performs frequency analysis on the first state data as the waveform data to generate power spectrum information for each frequency of the first state data (S12). In this way, first state data as power spectrum information is generated. The first state data as power spectrum information is represented as a value with respect to a frequency, as shown in FIG.

  Subsequently, the preprocessing unit 105 performs a normalization process to generate first state data as regularized power spectrum information (S13). Regularization is a process of setting the maximum value to 1. For example, when the maximum value of the first state data as the power spectrum information subjected to the frequency analysis is n (shown in FIG. 7), it is regularized by performing processing of 1 / n. The first state data as the regularized power spectrum information is expressed as a value with respect to the frequency as shown in FIG.

  Subsequently, the preprocessing unit 105 stores the first state data as regularized power spectrum information in the first state data storage unit 104 (S14). Here, first state data as power spectrum information before regularization may be stored in the first state data storage unit 104.

  The first learning model generation unit 106 generates a first learning model for generating second state data obtained by converting the first state data by machine learning using the first state data as first learning data (first learning model). Learning model generation step). As the first state data used by the first learning model generation unit 106, it is preferable to use the first state data generated by the preprocessing unit 105. That is, the first state data used in the first learning model generation unit 106 is preferably the first state data as regularized power spectrum information. By using the regularized data, it is possible to suppress the influence of the variation in data for each workpiece W. Further, the use of the power spectrum information facilitates the later-described machine learning processing. However, the first state data as power spectrum information before regularization can be used, and the first state data as waveform data can also be used.

  Then, the first learning model generation unit 106 reduces the number of variables included in the first state data by machine learning using the first state data as the first learning data, thereby reducing the dimension of the second state data. Generate a first learning model for generating. As shown in FIG. 9, the second state data generated by the first learning model is power spectrum information extracted as a part of the feature amount from the first state data as the power spectrum information shown in FIG. is there.

  That is, as shown in FIG. 10, the first learning data in the first learning model is P pieces of power spectrum information, whereas the output data (second state data) generated by the first learning model is , Q pieces of power spectrum information. Here, Q is a value much smaller than P. In other words, the first learning model is a model for generating the first state data having the dimension number P and the second state data having the reduced dimension number Q. Further, the number of dimensions here corresponds to the number of variables in a process using a second learning model described later.

  Here, in order to reduce the number of dimensions, when extracting some Q pieces from the P pieces of first state data, it is necessary to extract characteristic data from the first state data. However, it is not easy to determine characteristic data. Therefore, machine learning is used to determine characteristic data. That is, the first learning model extracts a feature variable (feature frequency) from a plurality of variables included in the first state data and generates a value corresponding to the feature variable (feature frequency) as the second state data. Model.

  The first learning model uses, for example, Lasso regression (also referred to as L1 regularization). In other words, the first learning model can perform feature selection by using Lasso regression as a means for reducing the dimension, thereby setting some parameters to zero. Then, by performing machine learning using a large number of first state data, it becomes possible to select a feature variable (feature frequency).

  Then, the first learning model generated by the first learning model generation unit 106 is stored in the first learning model storage unit 107 (first learning model storage step). Here, the first learning model generation unit 106 can continuously generate the first learning model even in the inference phase described later. In this case, the first learning model is updated successively. In machine learning, the more learning data, the higher the accuracy of the learning model. Therefore, by continuing to update the first learning model, it is possible to obtain more accurate results as the first learning model is used.

  Further, the second state data as output data obtained by the first learning model generation unit 106 learning in the learning phase is stored in the second state data storage unit 108. The second state data is stored in the second state data storage unit 108 in association with the workpiece W.

  The second learning model generation unit 109 performs machine learning using the second state data stored in the second state data storage unit 108 and the quality data stored in the quality data storage unit 102 as second learning data. Do. The second learning model generation unit 109 generates a second learning model related to the classification to which the second state data and the quality data belong by the machine learning (second learning model generation step). The generated second learning model is stored in the second learning model storage unit 110 (second learning model storage step). As the second learning model, a model for performing classification is applied. As the second learning model, for example, a support vector machine (SVM) or the like is applied.

  As shown in FIG. 11, the first example of the second learning data includes the second state data as Q power spectrum information obtained based on the waveform data and the quality as a value indicating the degree of the chatter pattern. And data. That is, the second learning model is a model representing the relationship between the power spectrum information and the value indicating the degree of the chatter pattern. Then, the quality data as a value indicating the degree of the chatter pattern is data for classification into two types, normal and abnormal. In other words, the second learning model is a model relating to two types of classification, normal and abnormal, as the classification to which the quality data belongs.

  In addition, the second learning model may be a model relating to a plurality of classifications in normal and a plurality of classifications in abnormal as classifications belonging to the quality data. In this case, even in the normal case, it is possible to distinguish between a case where the position is near the boundary of the normal range and a case where the position is near the center of the normal range. In addition, by classifying the abnormality into a plurality of levels, it is possible to distinguish between a minor abnormality and a severe abnormality. It is possible to execute a coping method according to the degree of the abnormality.

  In addition, as shown in FIG. 12, the second example of the second learning data includes the second state data as Q power spectrum information obtained based on the waveform data, and the second state data obtained based on the out-of-waveform data. Q1 pieces of second state data as power spectrum information and quality data as a value indicating the degree of chatter pattern are included. In this case, the second learning model is a model representing a relationship between power spectrum information based on waveform data, power spectrum information based on out-of-waveform data, and a value indicating the degree of chatter pattern.

  By using a plurality of types of power spectrum information as described above, the relationship with the quality data can be learned with higher accuracy. Here, the power spectrum information based on the out-of-waveform data is obtained in the same manner as the power spectrum information based on the above-described waveform data. That is, Q1 may have the same value as Q or a different value.

  Next, a detailed process of the second learning model generation unit 109 will be described with reference to FIG. The second learning model generation unit 109 switches the generated second learning model between a second learning model based on exclusive identification and a second learning model based on relative identification according to the cumulative number of abnormal data.

  As shown in FIG. 13, the second learning model generation unit 109 acquires the cumulative number of quality data (referred to as abnormal data) relating to the abnormality stored in the quality data storage unit 102 (S21). Subsequently, the second learning model generation unit 109 determines whether the cumulative number of abnormal data is larger than a reference value (S22). When the cumulative number of abnormal data is equal to or smaller than the reference value (S22: No), the second learning model generation unit 109 performs machine learning using quality data relating only to normal to thereby obtain a second learning model based on exclusive identification. Is generated (S23). On the other hand, when the cumulative number of abnormal data is larger than the reference value (S22: Yes), the second learning model generation unit 109 performs relative learning by performing machine learning using quality data on normality and quality data on abnormalities. Then, a second learning model is generated (S24).

  Here, the above switching process is based on the premise that the cumulative number of quality data relating to normality (referred to as normal data) is larger than the cumulative number of abnormal data. That is, in the early stage of data storage, even if a large number of normal data can be obtained, the abnormal data cannot be sufficiently obtained compared to the normal data. Therefore, in the initial stage of data storage, since normal data has been sufficiently stored, a second learning model for performing exclusive identification using only normal data is generated.

  On the other hand, by continuing to accumulate data, a state in which a sufficient cumulative number of abnormal data can be obtained is obtained. In that case, a second learning model for performing relative identification is generated using the normal data and the abnormal data. After the second learning model for performing the relative identification is generated, the second learning model is replaced with the second learning model for performing the exclusive identification and stored in the second learning model storage unit 110. Is done.

(4-2. Inference phase)
Next, a portion of the abnormality determination device 100 that functions in the inference phase of machine learning will be described with reference to FIG. As illustrated in FIG. 2, the abnormality determination device 100 includes a first state data acquisition unit 103, a first state data storage unit 104, a preprocessing unit 105, a first learning model storage unit 107, It includes a first prediction unit 111, a second state data storage unit 108, a second learning model storage unit 110, a second prediction unit 112, and an output unit 113.

  Here, the first learning model storage unit 107 and the second learning model storage unit 110 store the first learning model and the second learning model in the learning phase described above. Further, the learning phase is executed even while the inference phase is being executed. Therefore, the first learning model stored in the first learning model storage unit 107 and the second learning model stored in the second learning model storage unit 110 are updated.

  In the inference phase, the first state data at the determination timing is acquired by the first state data acquisition unit 103 and stored in the first state data storage unit 104. Further, the preprocessing unit 105 performs preprocessing on the first state data stored in the first state data storage unit 104, so that the first state data after the preprocessing is stored in the first state data storage unit 104. It is remembered. The quality data is acquired by the quality data acquisition unit 101 and stored in the quality data storage unit 102. This is the same as in the learning phase.

  The first prediction unit 111, the second prediction unit 112, and the output unit 113, which are configurations specific to the inference phase, will be described. The first prediction unit 111 acquires the first state data from the first state data storage unit 104. That is, the first prediction unit 111 acquires the first state data at the determination timing as the first input data. Here, the first input data that is the first state data uses the first state data as the P pieces of power spectrum information generated by the preprocessing unit 105.

  Then, the first prediction unit 111 predicts second state data corresponding to the first input data, using the obtained first input data and the first learning model stored in the first learning model storage unit 107. (First prediction step). As described above, the first learning model is a model that outputs the second state data when the first state data is input. Therefore, by inputting the first state data using the first learning model, the second state data is output.

  That is, the first prediction unit 111 obtains the second state data in which the dimension is reduced as shown in FIG. Note that the first prediction unit 111 uses the type of the first state data applied in the learning phase. That is, as shown in FIG. 3 or FIG. 4, the first prediction unit 111 can use power spectrum information based on one type of waveform data, or can use a plurality of pieces of information. The second state data output by the first prediction unit 111 is stored in the second state data storage unit 108.

  The second prediction unit 112 acquires the second state data predicted by the first prediction unit 111 from the second state data storage unit 108 as second input data. The second prediction unit 112 predicts the classification to which the quality data corresponding to the second input data belongs using the second input data and the second learning model stored in the second learning model storage unit 110 (the second prediction unit 112). Two prediction steps). As described above, the second learning model is a model relating to the second state data and the quality data. Therefore, by inputting the second state data using the second learning model, information relating to the quality data, that is, information that can be classified and belongs to the quality data is output.

  Therefore, the second predicting unit 112 can divide the first state data at the determination timing into two types of classification, that is, normal and abnormal, as the classification to which the quality data belongs. In other words, it is possible to classify whether the workpiece W associated with the first state data is in a normal state in which no chatter pattern exists or an abnormal state in which a chatter pattern exists.

  In addition, when the second learning model is a model relating to a multi-stage classification in normal and a multi-stage classification in abnormal as a classification belonging to the quality data, the workpiece W may be any one of the multi-stages. It can be predicted whether it belongs to a classification.

  The output unit 113 outputs the normal / abnormal state of the workpiece W determined by the second prediction unit 112. The output unit 113 performs output by display on a display device (not shown). The output unit 113 may output to the display device or the like of the abnormality determination device 100, may output to the display device or the like of the target processing device 10, or may output to the server 30 or other management device. You may make it output to a display device etc. The output unit 113 can also output to a mobile terminal owned by the worker or the administrator.

(5. Effect)
According to the abnormality determination device 100, the second prediction unit 112 determines, using the second state data and the second learning model, the classification to which the quality data belongs, that is, whether the predetermined process is normal or abnormal. I have. The second state data is not the first state data detected by the sensor 11 attached to the processing device 10, but is data obtained by converting the first state data. The second state data is data obtained by the first prediction unit 111 using the first state data and the first learning model.

  That is, the first stage processing of generating the second state data using the first state data and the first learning model is performed, and after the first stage processing, the second state data and the second learning model are used to perform normal processing. And a second-stage process for determining abnormality is performed. As described above, the abnormality determination device 100 applies a learning model of machine learning in a plurality of stages. Therefore, it is possible to accurately and correctly determine and predict the normal and abnormal states.

(6. Configuration of abnormality determination device 200 of second example)
The configuration of the abnormality determination device 200 of the second example will be described with reference to FIG. The method (the abnormality determination method) by the abnormality determination device 200 of the second example is also described.

  As in the first example, the abnormality determination device 200 includes a part that functions in a learning phase of machine learning and a part that functions in an inference phase of machine learning. Even during the execution of the inference phase, the learning model is updated by executing the learning phase. Hereinafter, the learning phase and the inference phase will be described separately. In FIG. 14, solid arrows indicate the flow of processing that functions in the learning phase, and dashed arrows indicate the flow of processing that functions in the inference phase. In addition, the same components as those in the first example are denoted by the same reference numerals, and detailed description is omitted.

(6-1. Learning phase)
As shown in FIG. 14, the abnormality determination device 200 includes a quality data acquisition unit 101, a quality data storage unit 102, a processing condition acquisition unit 221, a processing condition storage unit 222, a first state data acquisition Unit 103, first state data storage unit 104, preprocessing unit 105, first learning model generation unit 206, first learning model storage unit 107, second state data storage unit 108, second learning model generation unit 209, A second learning model storage unit 110 is provided.

  The processing condition acquisition unit 221 acquires the processing conditions stored in the control device 12 of the processing device 10 (processing condition acquisition step). The processing conditions are, for example, information on the type of the processing tool T, the material of the workpiece W, the timing from the start to the end of the processing, and the like. The processing conditions may be various operating conditions of the processing device 1 such as a set spindle speed. The processing conditions may include processing steps such as roughing and finishing. The acquired processing conditions are stored in the processing condition storage unit 222 (processing condition storage step).

  The first learning model generation unit 206 performs machine learning on the first learning model generation unit 106 of the first example according to the processing conditions, thereby generating a first learning model according to the processing conditions. Different. That is, the first learning model generation unit 206 generates the first learning model for generating the second state data obtained by converting the first state data by machine learning using the first state data and the processing condition as the first learning data. Generate (first learning model generation step).

  As shown in FIG. 15, the first learning data in the first learning model includes data relating to the first state data and data relating to the processing conditions. The first learning data as the first state data is P pieces of power spectrum information. The first learning data as the processing condition includes the type of the processing tool T, the material of the workpiece W, the processing timing, and the like. Output data (second state data) generated by the first learning model is Q pieces of power spectrum information. In addition, the first learning model extracts a feature variable (feature frequency) from a plurality of variables included in the first state data and generates a value corresponding to the feature variable (feature frequency) as second state data. Model. That is, the first learning model is a model for lowering the dimension as in the first example.

  The second learning model generation unit 209 performs machine learning on the second learning model generation unit 109 of the first example according to the processing conditions, thereby generating a second learning model according to the processing conditions. Different. That is, the second learning model generation unit 209 generates a second learning model for performing classification to which the quality data belongs by machine learning using the second state data and the processing condition as the second learning data.

  As shown in FIG. 16, the first example of the second learning data includes the second state data as Q power spectrum information obtained based on the waveform data and the quality as a value indicating the degree of the chatter pattern. Data and processing conditions are included. The processing conditions include the type of the processing tool T, the material of the workpiece W, the processing timing, and the like. That is, the second learning model is a model representing the relationship between the power spectrum information, the value indicating the degree of the chatter pattern, and the processing condition.

  As the classification to which the quality data belongs, the second learning model may be a model relating to two types of classification, normal and abnormal, or may be a model relating to multiple stages of classification in normal and multiple stages of classification in abnormal.

  As shown in FIG. 17, the second example of the second learning data includes the second state data as Q power spectrum information obtained based on the waveform data and the second state data obtained based on the out-of-waveform data. The second state data as Q1 pieces of power spectrum information, quality data as a value indicating the degree of chattering, and processing conditions are included. In this case, the second learning model is a model that represents a relationship between power spectrum information based on waveform data, power spectrum information based on out-of-waveform data, a value indicating a degree of chatter, and processing conditions.

  By using a plurality of types of power spectrum information as described above, the relationship with the quality data can be learned with higher accuracy. Here, the power spectrum information based on the out-of-waveform data is obtained in the same manner as the power spectrum information based on the above-described waveform data.

(6-2. Inference phase)
Next, a portion that functions in the inference phase of machine learning of the abnormality determination device 200 will be described with reference to FIG. As illustrated in FIG. 14, the abnormality determination device 200 includes a processing condition acquisition unit 221, a processing condition storage unit 222, a first state data acquisition unit 103, a first state data storage unit 104, The processing unit 105 includes a first learning model storage unit 107, a first prediction unit 211, a second state data storage unit 108, a second learning model storage unit 110, a second prediction unit 212, and an output unit 113.

  The first prediction unit 211 is different from the first prediction unit 111 of the first example in that the second prediction unit 211 predicts the second state data according to the processing conditions. The first predictor 211 acquires the first state data and the processing condition at the determination timing as the first input data, and uses the first input data and the first learning model to generate the second state data corresponding to the first input data. Is predicted (first prediction step).

  The second prediction unit 212 differs from the second prediction unit 112 of the first example in that the classification to which the quality data according to the processing condition belongs belongs. The second prediction unit 212 acquires the second state data and the processing conditions as the second input data, and predicts the classification to which the quality data corresponding to the second input data belongs using the second input data and the second learning model. (Second prediction step).

  Thus, the abnormality determination device 200 of the second example can determine the classification to which the quality data belongs, that is, whether the processing is normal or abnormal, in consideration of the processing conditions. That is, since the classification can be performed in more detail in consideration of the processing conditions, it is possible to determine whether the processing tool T is normal or abnormal according to the type.

1: Abnormality determination system, 10: Processing device, 11: Sensor, 12: Control device, 20: Inspection device, 30: Server, 100, 200: Abnormality determination device, 101: Quality data acquisition unit, 102: Quality data storage unit 103: first state data acquisition unit, 104: first state data storage unit, 105: preprocessing unit, 106, 206: first learning model generation unit, 107: first learning model storage unit, 108: second state Data storage unit, 109, 209: second learning model generation unit, 110: second learning model storage unit, 111, 211: first prediction unit, 112, 212: second prediction unit, 113: output unit, 221: processing Condition acquisition unit, 222: machining condition storage unit, T: machining tool, W: workpiece (object)

Claims (14)

At the time of execution of a predetermined process by the processing device, a first state data acquisition unit that acquires first state data detected by a sensor attached to the processing device,
Quality data related to normal and abnormal for the predetermined process, a quality data acquisition unit that acquires the quality data associated with the first state data,
A first learning model generating unit that generates a first learning model for generating second state data obtained by converting the first state data by machine learning using the first state data as first learning data,
A second learning model generation unit that generates a second learning model for the second state data and the classification to which the quality data belongs by machine learning using the second state data and the quality data as second learning data,
A first prediction unit that predicts the second state data corresponding to the first input data using the first state data at the determination timing as the first input data and using the first input data and the first learning model. When,
The second state data as the second input data, using the second input data and the second learning model, a second prediction unit that predicts the classification to which the quality data corresponding to the second input data belongs,
An abnormality determination device comprising:   The first learning model generation unit is configured to reduce the number of variables included in the first state data by machine learning using the first state data as first learning data. The abnormality determination device according to claim 1, wherein the first learning model for generating data is generated.   The first learning model is a model for extracting a feature variable from among a plurality of variables included in the first state data, and generating a value corresponding to the feature variable as the second state data. Item 3. The abnormality determination device according to item 2. The abnormality determination device further includes:
Preprocessing for generating power spectrum information for each frequency of the first state data by performing frequency analysis on the first state data as waveform data, and acquiring the first state data as the power spectrum information Part,
The first learning model generating unit generates the first learning model using the first state data as the power spectrum information generated by the preprocessing unit,
The said 1st prediction part predicts the said 2nd state data using the said 1st state data as the said power spectrum information produced | generated by the said pre-processing part, The Claims any one of Claims 1-3. Abnormality determination device.   The abnormality determination device according to claim 4, wherein the preprocessing unit generates the first state data as the regularized power spectrum information by performing a normalization process. The second learning model generation unit,
The generated second learning model, switching between the second learning model by exclusive identification and the second learning model by relative identification,
When the cumulative number of the quality data regarding the abnormality is equal to or less than a reference value, the second learning model is generated by exclusive identification by performing machine learning using the quality data regarding only the normal state,
If the cumulative number of the quality data relating to the abnormality is larger than the reference value, the second learning model based on relative identification is generated by performing machine learning using the quality data relating to the normal and abnormality. 6. The abnormality determination device according to claim 5. The second learning model generation unit, as the classification to which the quality data belongs, a plurality of normal classification, and the second learning model for a plurality of abnormal classification of abnormal, the second learning model,
7. The method according to claim 1, wherein the second predicting unit predicts, as a classification to which the quality data belongs, which classification among a plurality of normal classifications and a plurality of abnormal classifications. An abnormality determination device according to claim 1. The processing device is a machine tool that performs a cutting process or a grinding process on the object as the predetermined process,
The first state data includes a sound wave generated by cutting or grinding, vibration, waveform data related to any of the processing load,
The abnormality determination device according to any one of claims 1 to 7, wherein the quality data is data on normality and abnormality of the object.   The abnormality determination device according to claim 8, wherein the first state data includes out-of-waveform data on an image of the object in addition to the waveform data. The first learning model generation unit and the second learning model generation unit, by performing machine learning according to the processing conditions, to generate the first learning model and the second learning model according to the processing conditions,
The abnormality determination device according to claim 8, wherein the first prediction unit and the second prediction unit predict a classification to which the second state data and the quality data belong according to the processing condition.   The abnormality determination device according to claim 10, wherein the processing condition is any one of a type of a processing tool, a material of the object, and a timing from the start to the end of the processing.   The abnormality determination device according to claim 8, wherein the quality data is data relating to a degree of a chatter pattern generated in the target object. A plurality of processing devices that execute a predetermined process;
A server that is provided so as to be able to communicate with the plurality of processing devices, and collects first state data detected by a sensor attached to the processing device when the predetermined processing is performed by each of the plurality of processing devices,
The abnormality determination device according to any one of claims 1 to 12, wherein the abnormality determination device is processed based on the first state data collected by the server,
An abnormality determination system comprising: A first state data acquisition step of acquiring first state data detected by a sensor attached to the processing apparatus during execution of a predetermined process by the processing apparatus;
Quality data related to normal and abnormal for the predetermined processing, quality data acquisition step of acquiring the quality data associated with the first state data,
A first learning model generating step of generating a first learning model for generating second state data obtained by converting the first state data by machine learning using the first state data as first learning data,
A second learning model generating step of generating a second learning model for the second state data and the classification to which the quality data belongs by machine learning with the second state data and the quality data as second learning data,
A first prediction step of predicting the second state data corresponding to the first input data using the first state data at the determination timing as the first input data and using the first input data and the first learning model; When,
The second state data as the second input data, using the second input data and the second learning model, a second prediction step of predicting the classification to which the quality data corresponding to the second input data belongs,
An abnormality determination method comprising:

Images