JP2021022311A - Abnormality detecting device, abnormality detecting system, and program - Google Patents

Abstract

To provide an abnormality detecting device capable of streamlining labelling of sample data used for learning a learning model for detecting an object machine abnormality, keeping high abnormality detection precision, and achieving early detection of abnormalities and signs thereof, an abnormality detecting system, and a program.SOLUTION: An abnormality detecting device comprises: a frequency analysis processing unit which subjects a detection signal detected from an abnormality detection object to frequency analysis to obtain frequency characteristics; a clustering processing unit which classifies sample data of each frame included in the frequency characteristics into clusters; a multivariable analysis processing unit which extracts a feature quantity via multivariable analysis; a feature quantity display unit which visualizes classification of sample data into clusters; a presentation unit which presents information corresponding to the selected sample data among the displayed sample data; a labelling processing unit which sets a sample data label per cluster; and a model generation unit which generates a learning model using sample data in which a label is set.SELECTED DRAWING: Figure 4

Description

The present invention relates to anomaly detection devices, anomaly detection systems and programs.

A system that automatically detects abnormalities by monitoring the operating status of drive devices such as pumps or compressors with microphones or vibration sensors in machine tools or equipment such as processing machines (hereinafter sometimes referred to as machines). Is widely considered. In order to operate such machines and the like stably without trouble, it is important to detect the occurrence of failure and its sign as soon as possible by daily inspection and perform appropriate maintenance. Since malfunctions of machines and the like often appear as abnormalities in the sounds or vibrations generated by them, inspections are often performed by the ears of maintenance personnel. However, it is costly to constantly monitor machines, etc., and there is a possibility that detection may be delayed by regular inspections at regular intervals, and the judgment of abnormality detection also varies depending on the skill level of maintenance personnel. Therefore, there are problems such as the occurrence of an abnormality at an appropriate timing and the possibility of overlooking the sign thereof. In order to solve these problems, in recent years, an automatic detection system using a learning model that detects anomalies based on machine learning has been studied.

In general, a learning model in machine learning is constructed by a process called "learning" using a large amount of sample data (learning data) having the same properties as a judgment target collected in advance. There are several types of learning models for such machine learning, but when the purpose is to detect abnormalities in machines, etc., sample data at the time of abnormalities that occur infrequently (hereinafter, may be simply referred to as abnormal sample data). It is difficult to collect a large amount of). Therefore, outlier detectors that use only normal sample data or learning models called one-class classifiers are often used.

The outlier detector calculates the degree of deviation indicating the degree of deviation from the distribution of the normal sample data (hereinafter, may be simply referred to as normal sample data) used for learning in advance, and the deviation is obtained. If the degree is less than a certain threshold value, it is classified as normal, and if it is above the threshold value, it is classified as abnormal. Therefore, the quality of the normal sample data used when learning the outlier detector has a great influence on the performance of the outlier detector. When the number of sample data is insufficient, or when learning is performed with sample data that is significantly deviated from the population distribution of the original normal data, it is easy to erroneously determine that the input of normal sample data is abnormal. On the other hand, if sample data that should be considered abnormal is mixed in a part of the sample data, there is a high possibility that an input of similar sample data will be erroneously determined as normal. Therefore, in order to realize highly accurate judgment (abnormality detection), it is necessary for a person to select (label) the sample data to be given at the time of learning, and this labeling work often requires enormous labor, time, and cost. It becomes an issue.

As an anomaly detection system using such an outlier detector (1 class classifier), in the outlier detector for imaging data, in addition to the outlier detector learned only from normal sample data, an abnormality A system using an outlier detector learned only from sample data is disclosed (for example, Patent Document 1). It is described that the system can reduce the occurrence of an event in which abnormal sample data is erroneously determined to be normal, which cannot be reduced only by additional learning of normal sample data.

Further, as the above-mentioned abnormality detection system, the usefulness of additional learning as a means for dealing with aged deterioration of patterns in an outlier detector (1 class support vector machine) for abnormality detection is disclosed (). For example, Patent Document 2). The system uses an outlier detector learned from the initial sample data in combination with an outlier detector additionally learned from the latest sample data to determine whether the anomaly that has occurred is due to aging. It is disclosed that it has a mechanism to do it.

In general, it is known that the detection operation by the outlier detector (1 class classifier) is more difficult than the detection operation by the 2 class classifier, and it is difficult to make a highly accurate determination. The two-class classifier can make a relative judgment as to whether it is closer to A or B, whereas the outlier detector needs to make an absolute judgment as to whether it is A or not. Because. Nevertheless, in many anomaly detection systems targeting equipment, etc., it is generally necessary to take an outlier detector type approach because the probability of occurrence of anomalies is very low and a large number of anomalies are present in advance. This is because it is difficult to obtain sample data. Therefore, an approach that uses an outlier detector learned only from anomalous sample data, such as the system described in Patent Document 1, is impractical. Furthermore, even in a system environment where a large amount of abnormal sample data can be acquired, in that case, higher performance can be expected by performing abnormality detection using a normal and abnormal two-class classifier. Therefore, it is not necessary to use two types of outlier detectors. In the first place, the root cause of erroneously determining abnormal sample data as normal is the sample data that is actually abnormal in the normal sample data (sample data that is considered normal) used for learning the outlier detector. The point is that data close to is mixed. In order to solve this problem, it is judged whether the sample data input as the training data of the learning model such as the outlier detector is really normal sample data or actually abnormal sample data. However, a mechanism that enables appropriate labeling is required. Ultimately, humans have no choice but to make such judgments, but it is unrealistic for humans to make judgments one by one while visually or listening to a large amount of learning data. There should be room for technical intervention in order to reduce the load on the subject, but the technique described in Patent Document 1 cannot solve the problem.

Further, the system described in Patent Document 2 does not disclose a mechanism for guaranteeing the validity of sample data (learning data) used at the time of additional learning, and is valid for a large amount of learning data as described above. There is a problem that it is not possible to reduce the burden on the person that occurs when determining.

The present invention has been made in view of the above, and can streamline the labeling work of sample data used for learning a learning model for detecting anomalies of a target machine, maintain high anomaly detection accuracy, and have anomalies. It is an object of the present invention to provide an abnormality detection device, an abnormality detection system and a program capable of realizing early detection of a sign thereof.

In order to solve the above-mentioned problems and achieve the object, the present invention has a frequency analysis processing unit that performs frequency analysis on a detection signal detected from an abnormality detection target to obtain frequency characteristics, and the frequency analysis processing unit. Each frame included in the frequency characteristics obtained by the above is used as sample data, and the clustering processing unit that classifies the sample data into one or more clusters and a multivariate with respect to the frequency characteristics obtained by the frequency analysis processing unit. Based on the multivariate analysis processing unit that performs analysis and extracts the feature amount and the feature amount extracted by the multivariate analysis processing unit, the sample data is visualized so that it can be determined which cluster it is classified into. The feature amount display unit, the presentation unit that presents the information corresponding to the sample data selected by the input signal from the input unit among the sample data displayed by the feature amount display unit, and the presentation unit that presents the information. Based on the information provided, a labeling processing unit that sets a label for the sample data for each cluster according to an input signal from the input unit, and the sample data whose label is set by the labeling processing unit are used. It is characterized by including a model generation unit that generates a learning model by learning.

According to the present invention, it is possible to streamline the labeling work of sample data used for learning a learning model for detecting anomalies in a target machine, maintain high anomaly detection accuracy, and realize early detection of anomalies and their precursors. be able to.

FIG. 1 is a diagram showing an example of the overall configuration of the abnormality detection system according to the embodiment. FIG. 2 is a diagram showing an example of the hardware configuration of the abnormality detection device according to the embodiment. FIG. 3 is a diagram showing an example of the hardware configuration of the machine tool according to the embodiment. FIG. 4 is a diagram showing an example of a functional block configuration of the abnormality detection device according to the embodiment. FIG. 5 is a diagram illustrating the operation of the frequency analysis processing unit of the abnormality detection device according to the embodiment. FIG. 6 is a diagram illustrating the operation of the clustering processing unit of the abnormality detection device according to the embodiment. FIG. 7 is a diagram illustrating the operation of the multivariate analysis processing unit of the abnormality detection device according to the embodiment. FIG. 8 is a diagram illustrating the operation of the feature amount display unit of the abnormality detection device according to the embodiment. FIG. 9 is a diagram illustrating the operation of the basic information display unit of the abnormality detection device according to the embodiment. FIG. 10 is a diagram illustrating the operation of the reproduction unit of the abnormality detection device according to the embodiment. FIG. 11 is a diagram illustrating the operation of the labeling processing unit of the abnormality detection device according to the embodiment. FIG. 12 is a diagram illustrating the operation of the model generation unit of the abnormality detection device according to the embodiment. FIG. 13 is a diagram illustrating the operation of the abnormality detection unit of the abnormality detection device according to the embodiment. FIG. 14 is a diagram illustrating the operation of the goodness-of-fit calculation unit of the abnormality detection device according to the embodiment. FIG. 15 is a diagram illustrating the operation of the incidental information acquisition unit of the abnormality detection device according to the embodiment. FIG. 16 is a diagram illustrating the operation of the model management unit of the abnormality detection device according to the embodiment. FIG. 17 is a diagram illustrating the operation of the model selection unit of the abnormality detection device according to the embodiment. FIG. 18 is a diagram illustrating the operation of the data adjusting unit of the abnormality detection device according to the embodiment. FIG. 19 is a diagram showing an example of the functional block configuration of the machine tool according to the embodiment. FIG. 20 is a diagram showing an example of a flow of the overall operation of the abnormality detection device according to the embodiment. FIG. 21 is a diagram showing an example of a flow of the overall operation of the abnormality detection device according to the embodiment.

Hereinafter, embodiments of the abnormality detection device, the abnormality detection system, and the program according to the present invention will be described in detail with reference to the drawings. Further, the present invention is not limited by the following embodiments, and the components in the following embodiments include those easily conceived by those skilled in the art, substantially the same, and so-called equivalent ranges. Is included. Furthermore, various omissions, substitutions, changes and combinations of components can be made without departing from the gist of the following embodiments.

(Overall configuration of anomaly detection system)
FIG. 1 is a diagram showing an example of the overall configuration of the abnormality detection system according to the embodiment. The overall configuration of the abnormality detection system 1 according to the present embodiment will be described with reference to FIG.

As shown in FIG. 1, the abnormality detection system 1 includes an abnormality detection device 10, a machine tool 20 (an example of an abnormality detection target), and an A / D converter 30.

The abnormality detection device 10 is an information processing device that receives vibration data or acoustic data generated in a machining cycle by the machine tool 20 as a detection signal and performs analysis processing such as detection and determination of the presence or absence of an abnormality. The abnormality detection device 10 receives the detection signal as a digital signal converted by the A / D converter 30. The abnormality detection device 10 includes a display 58 (display device) for displaying a received detection signal or the like. The abnormality detection device 10 receives context information described later from the NC control device 25 of the machine tool 20, and receives an operation signal indicating that the machine tool 20 is in the process of machining via the A / D converter 30. Further, as shown in FIG. 1, the abnormality detection device 10 can be connected to the cloud 40 via a network, and the acquired detection signal or the like can be stored in the cloud 40. It is not essential that the abnormality detection device 10 is connected to the cloud 40.

The machine tool 20 is a machine that uses a tool 23 to perform machining such as cutting, grinding, or polishing on a machining target. The machine tool 20 is an example of an abnormality detection target to be detected by the abnormality detection device 10. The machine tool 20 is equipped with a sensor 24 that detects vibration or sound generated in a machining cycle, is held by a holder 22, and has a tool 23 that performs machining such as cutting, grinding, or polishing on a machining target. It includes an NC (Numerical Control) control device 25 that controls the operation of the machining cycle.

The target of abnormality detection by the abnormality detection device 10 is not limited to the machine tool 20, and any target that can detect an abnormality in operation by analyzing vibration or sound waveforms, for example, a printer. Alternatively, it can be targeted at automobiles (an example of anomaly detection target). Further, in the example shown in FIG. 1, the abnormality detection device 10 is a device separate from the machine tool 20, but the device is not limited to this, and may be mounted on the machine tool 20. ..

The sensor 24 is installed separately from the machine tool 20 and detects a physical quantity such as vibration or sound generated by a tool 23 such as a drill, an end mill, a face mill, a long drill, a tool tip or a grindstone, and information on the detected physical quantity. Is a sensor that outputs a detection signal (vibration data, acoustic data, etc.) to the A / D converter 30. The sensor 24 is composed of, for example, an acceleration sensor, an AE (Acoustic Emission) sensor, a microphone, or the like. The number of sensors 24 may be plural.

The sensor 24 that detects the physical quantity generated by the machine tool 20 is intended to detect vibration data, acoustic data, and the like, but is not limited to this, and is not limited to, for example, a torque sensor that detects the rotational torque of a tool. , An external sensor such as a load cell that detects a load related to a processing target or the like, a temperature sensor that detects temperature information, a camera that detects an image, or the like may be used.

The tool 23 is a machining tool such as a drill, an end mill, a face mill, a long drill, a tool tip, or a grindstone for performing machining such as cutting, grinding, or polishing on a machining target.

The NC control device 25 is a control device that controls the entire operation of the machining cycle in the machine tool 20 by executing the NC program. The NC control device 25 outputs context information described later and an operation signal indicating that processing is in progress.

The A / D converter 30 is a device that converts an analog detection signal (vibration data, acoustic data, etc.) output from the sensor 24 into a digital signal. Further, the A / D converter 30 converts an operation signal output from the NC control device 25 indicating that processing is in progress into a digital signal. That is, the analog detection signal output from the sensor 24 and the operation signal output from the NC control device 25 are input to different channels of the A / D converter 30. The A / D converter 30 outputs the converted digital signal to the abnormality detection device 10.

The A / D converter 30 is a device different from the abnormality detection device 10, but may be, for example, an A / D conversion board incorporated as an expansion board in the abnormality detection device 10. Further, the operation signal output from the NC control device 25 may not be transmitted to the A / D converter 30 but may be directly transmitted to the abnormality detection device 10 as an on / off signal.

(Hardware configuration of abnormality detection device)
FIG. 2 is a diagram showing an example of the hardware configuration of the abnormality detection device according to the embodiment. The hardware configuration of the abnormality detection device 10 according to the present embodiment will be described with reference to FIG.

As shown in FIG. 2, the abnormality detection device 10 includes a CPU (Central Processing Unit) 51, a ROM (Read Only Memory) 52, a RAM (Random Access Memory) 53, a communication I / F 54, and a sensor I / F 55. , The input / output I / F 56, the auxiliary storage device 59, and the voice I / F 61, and are connected by a bus 60 so that each unit can communicate with each other.

The CPU 51 is an arithmetic unit that controls the entire abnormality detection device 10. For example, the CPU 51 controls the operation of the entire abnormality detection device 10 and realizes a diagnostic function by executing a program stored in the ROM 52 or the like with the RAM 53 as a work area (work area).

The communication I / F 54 is an interface for communicating with an external device such as a machine tool 20. The communication I / F 54 is, for example, an interface compliant with Ethernet (registered trademark) and TCP (Transmission Protocol) / IP (Internet Protocol).

The sensor I / F 55 is an interface for receiving a detection signal (vibration data, acoustic data, etc.) and an operation signal from the sensor 24 installed in the machine tool 20. In reality, the sensor I / F 55 receives a digital signal whose detection signal and operation signal are A / D converted by the A / D converter 30. The interface for receiving the detection signal and the interface for receiving the operation signal may be separate.

The input / output I / F 56 is an interface for connecting various devices (for example, the input device 57 and the display 58) and the bus 60.

The input device 57 is a device such as a mouse or keyboard for inputting characters and numbers, selecting various instructions, and performing operations such as moving a cursor. The input device 57 may be realized by, for example, an input function (touch operation function) of the touch panel.

The display 58 is a display device such as an LCD (Liquid Crystal Display) or an organic EL (Electro-Luminence) display that displays various information such as cursors, menus, windows, characters, and images.

The auxiliary storage device 59 is an HDD (HDD) that stores various data such as setting information of the abnormality detection device 10, a detection signal (vibration data or acoustic data, etc.) received from the machine tool 20, an OS (Operating System), and an application program. It is a non-volatile storage device such as Hard Disk Drive), SSD (Solid State Drive), or EEPROM (Electrically Erasable Programmable Read-Only Memory). The auxiliary storage device 59 is provided in the abnormality detection device 10, but is not limited to this, and may be, for example, a storage device installed outside the abnormality detection device 10 or. It may be a server device capable of data communication with the abnormality detection device 10, or a storage device included in the cloud 40 or the like.

The voice I / F 61 is an interface for connecting a voice input / output device (for example, a speaker 62) and a bus 60. The speaker 62 (output device) is a device that outputs sound.

The hardware configuration of the abnormality detection device 10 shown in FIG. 2 is an example, and it is not necessary to include all the components shown in FIG. 2, or it may include other components.

(Hardware configuration of machine tools)
FIG. 3 is a diagram showing an example of the hardware configuration of the machine tool according to the embodiment. The hardware configuration of the machine tool 20 according to the present embodiment will be described with reference to FIG.

As shown in FIG. 3, the machine tool 20 has a CPU 71, a ROM 72, a RAM 73, a communication I / F 74, a drive control circuit 75, and a signal I / F 77, and each unit can communicate with each other. They are connected by bus 78. As shown in FIG. 1 above, the sensor 24 is installed at a predetermined position of the machine tool 20 capable of detecting vibration or sound generated in the machining cycle by the tool 23, but the sensor 24 is directly connected to the machine tool 20. It does not transmit or receive, and outputs a detection signal (vibration data, acoustic data, etc.) to the A / D converter 30 as described above.

The CPU 71 is an arithmetic unit that controls the entire machine tool 20. For example, the CPU 71 controls the operation of the entire machine tool 20 and realizes a machining function by executing a program (NC program) stored in the ROM 72 or the like with the RAM 73 as a work area (work area).

The communication I / F 74 is an interface for communicating with an external device such as the abnormality detection device 10. The drive control circuit 75 is a circuit that controls the drive of the motor 76. The motor 76 drives a tool 23 such as a drill, an end mill, a face mill, a long drill, a tool tip or a grindstone.

The signal I / F77 is an interface for transmitting an operation signal to the abnormality detection device 10 when the machine tool 20 is processing. A coaxial cable is connected to the signal I / F77 via, for example, a BNC (Bayonet Neil-Concelman connector) connector conforming to an Ethernet standard such as 10BASE-2.

The hardware configuration of the machine tool 20 shown in FIG. 3 is an example, and may include components other than the components shown in FIG. For example, the machine tool 20 may be provided with a display, and may be capable of displaying the same display contents as the display 58 included in the abnormality detecting device 10.

Further, the NC control device 25 shown in FIG. 1 described above is a device including, for example, a CPU 71, a ROM 72, a RAM 73, a communication I / F 74, and a drive control circuit 75. However, the present invention is not limited to this, and the machine tool 20 may be configured to include a CPU separately from the CPU 71 included in the NC control device 25. In this case, the CPU other than the CPU 71 has an operation other than the processing operation, for example, a lighting operation of a lamp or an LED (Light Emitting Diode) provided in the machine tool 20, and a rotation not directly related to the processing operation. The rotary motor that positions the unit 21 may be controlled.

(Configuration and operation of functional blocks of abnormality detection device)
FIG. 4 is a diagram showing an example of a functional block configuration of the abnormality detection device according to the embodiment. The configuration and operation of the functional block of the abnormality detection device 10 according to the present embodiment will be described with reference to FIG.

As shown in FIG. 4, the abnormality detection device 10 includes a frequency analysis processing unit 101, a clustering processing unit 102, a multivariate analysis processing unit 103, a feature quantity display unit 104, and a basic information display unit 105 (of the presentation unit). An example), a reproduction unit 106 (an example of a presentation unit), a labeling processing unit 107, a model generation unit 108, ancillary information acquisition unit 109, a model management unit 110, an abnormality detection unit 111, and a goodness-of-fit calculation unit. It has 112, a model selection unit 113, and a data adjustment unit 114. The abnormality detection device 10 further includes a communication unit 121, a signal reception unit 122, an input unit 123, a display unit 124 (display device), an output unit 125 (output device), and a storage unit 130.

The storage unit 130 is a functional unit that stores various types of information. The storage unit 130 is realized by at least one of the RAM 53 and the auxiliary storage device 59 shown in FIG. The storage unit 130 includes an acoustic data storage unit 131, a frequency characteristic storage unit 132, a feature quantity storage unit 133, a cluster storage unit 134, a model storage unit 135, and a detection result storage unit 136.

<Operation of frequency analysis processing unit>
FIG. 5 is a diagram illustrating the operation of the frequency analysis processing unit of the abnormality detection device according to the embodiment. The operation of the frequency analysis processing unit 101 of the abnormality detection device 10 will be described with reference to FIG.

The frequency analysis processing unit 101 of the abnormality detection device 10 stores acoustic data such as vibration data or acoustic data (hereinafter, referred to as acoustic data) which is a detection signal detected by the sensor 24 of the machine tool 20. This is a functional unit that reads acoustic data from the storage unit 131 and performs frequency analysis. Specifically, the frequency analysis processing unit 101 performs Fourier transform (FFT) or wavelet analysis as frequency analysis on the acoustic data, and obtains frequency characteristics such as a spectrogram as shown in FIG. obtain. For example, the frequency analysis processing unit 101 shifts a window length of 20 [ms] (the window function is, for example, a humming window) and 10 [ms] with respect to the one-minute acoustic data acquired at a sampling frequency of 48 [kHz]. When FFT was performed with a length and a frequency spectrum having 512-dimensional features per frame was calculated, a total of 5999 frames (the latter half 10 [ms] of the data of the final frame was processed by adding "0" data. In the case of 6000 frames) × 512-dimensional frequency characteristics (spectrogram) are obtained. The frequency analysis processing unit 101 stores the obtained frequency characteristic (spectrogram) in the frequency characteristic storage unit 132 of the storage unit 130, and also stores the 512-dimensional feature amount of each frame (frequency spectrum) included in the frequency characteristic. It is stored in the feature amount storage unit 133. The frequency analysis processing unit 101 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of clustering processing unit>
FIG. 6 is a diagram illustrating the operation of the clustering processing unit of the abnormality detection device according to the embodiment. The operation of the clustering processing unit 102 of the abnormality detection device 10 will be described with reference to FIG.

The clustering processing unit 102 of the abnormality detection device 10 reads out the frequency characteristics obtained by the frequency analysis processing unit 101 from the frequency characteristic storage unit 132, and reads the frequency spectrum of each frame included in the frequency characteristics into a plurality of groups (clusters). It is a functional part that classifies (clusters) into. This clustering process is realized by an algorithm such as, for example, the k-means method or GMM (Gaussian Mixture Model). Here, after the frequency spectrum of each frame is classified into a plurality of clusters by the clustering processing unit 102, the first group, the second group, and the third group (in FIG. 6) are arranged in descending order of the number of frames belonging to each cluster. It shall be referred to as cluster 1, cluster 2, cluster 3), ... The clustering processing unit 102 stores information about what kind of cluster each frame of the frequency characteristic is classified in the cluster storage unit 134 of the storage unit 130. The clustering processing unit 102 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

Data 1 to data n shown in FIG. 6 indicate spectrograms corresponding to each acoustic data (that is, n spectrograms), and each plot in the three-dimensional space shown in FIG. 6 is included in each spectrogram. The frame (frequency spectrum) is shown. Further, in FIG. 6, for convenience of display, it is expressed as a three-dimensional feature amount space, but in reality, the feature amount dimension of each frame (frequency spectrum) calculated by the frequency analysis processing unit 101 (for example, described above). It is a high-dimensional space of 512 dimensions).

<Operation of multivariate analysis processing unit>
FIG. 7 is a diagram illustrating the operation of the multivariate analysis processing unit of the abnormality detection device according to the embodiment. The operation of the multivariate analysis processing unit 103 of the abnormality detection device 10 will be described with reference to FIG. 7.

The multivariate analysis processing unit 103 of the abnormality detection device 10 reads out the frequency characteristics obtained by the frequency analysis processing unit 101 from the frequency characteristic storage unit 132, and each high-dimensional (for example, 512-dimensional) frequency spectrum included in the frequency characteristics. Is a functional unit that calculates a two-dimensional or three-dimensional feature amount that can be visualized by performing principal component analysis or multivariate analysis such as an autoencoder. For example, the multivariate analysis processing unit 103 performs principal component analysis on each frequency spectrum included in the frequency characteristics obtained by the frequency analysis processing unit 101, and extracts the first to third principal components as feature quantities. To do. This makes it possible to express each frame (frequency spectrum), which is a point on the high-dimensional feature space, as coordinates on the two-dimensional or three-dimensional feature space while retaining its main features. ..

The multivariate analysis processing unit 103 stores the two-dimensional or three-dimensional feature amount calculated by the multivariate analysis for each frequency spectrum in the feature amount storage unit 133 of the storage unit 130. For example, as shown in FIG. 7, the multivariate analysis processing unit 103 includes data (acoustic data, spectrogram), frames, clusters classified by the clustering processing unit 102, and a first principal component obtained by principal component analysis. The feature amount storage unit 133 stores a corresponding table in which the component to the third principal component are associated with each other. The multivariate analysis processing unit 103 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of feature amount display>
FIG. 8 is a diagram illustrating the operation of the feature amount display unit of the abnormality detection device according to the embodiment. The operation of the feature amount display unit 104 of the abnormality detection device 10 will be described with reference to FIG.

The feature amount display unit 104 of the abnormality detection device 10 refers to the correspondence table stored in the feature amount storage unit 133, and uses the feature amount calculated by the multivariate analysis processing unit 103 for the frequency spectrum of each frame as the source. This is a functional unit to be displayed on the display unit 124 in such a manner that the cluster to which the frequency spectrum of the above belongs can be discriminated. For example, as shown in FIG. 8, the feature amount display unit 104 is on a three-dimensional feature amount space centered on the feature amount (first principal component to third principal component) calculated by the multivariate analysis processing unit 103. A plot showing each frame (frequency spectrum) is displayed (presented) in different colors or shapes according to the cluster to which the frame belongs. As a result, the user can grasp the distribution state of each frame (frequency spectrum) in the three-dimensional feature space of the first principal component to the third principal component, and the plot is out of the upper large group. The state of the above can be visually recognized, and the frame to be noted becomes clear. When each frame is displayed as a plot on the three-dimensional feature space, FIG. 8 shows the state of the plot on the feature space from a fixed viewpoint, but it depends on the operation input to the input unit 123. Therefore, the direction of the viewpoint may be freely changed.

The feature amount display unit 104 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of basic information display unit>
FIG. 9 is a diagram illustrating the operation of the basic information display unit of the abnormality detection device according to the embodiment. The operation of the basic information display unit 105 of the abnormality detection device 10 will be described with reference to FIG.

The basic information display unit 105 of the anomaly detection device 10 selects from plots showing each frame (frequency spectrum) displayed in the feature amount space by the feature amount display unit 104 by operating the input unit 123 from the user. This is a functional unit that displays and presents basic information about the frame corresponding to the plotted plot on the display unit 124. Each plot on the feature space is mapped on a low-dimensional (two-dimensional or three-dimensional) feature space that can be visualized by multivariate analysis, and the frame corresponding to each plot originally has. Not all information can be expressed, and the meaning of each dimension is not always clear. Therefore, as shown in FIG. 9, the basic information display unit 105 uses basic information such as the original waveform of the acoustic data corresponding to the plot (frame) selected by the user and the time series (spectrogram) of the frequency spectrum. It is intended to be presented to and to support intuitive understanding. By visually displaying (presenting) the basic information of the plot (frame) selected by the user in this way, it is possible to assist the labeling work described later in which the knowledge of the user is reflected.

The basic information display unit 105 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of playback unit>
FIG. 10 is a diagram illustrating the operation of the reproduction unit of the abnormality detection device according to the embodiment. The operation of the reproduction unit 106 of the abnormality detection device 10 will be described with reference to FIG.

The reproduction unit 106 of the abnormality detection device 10 is selected by an operation from the user to the input unit 123 from the plots showing each frame (frequency spectrum) displayed in the feature amount space by the feature amount display unit 104. As shown in FIG. 10, this is a functional unit that reproduces acoustic data including a frame corresponding to a plot by sound from the output unit 125 for a predetermined time before and after (for example, 1 second before and after) including the time corresponding to the plot. The display operation by the basic information display unit 105 described above visually confirms the basic information of the selected frame, but at the site where the machine tool 20 is installed, the maintenance person listens to the state of the machine tool 20. It is captured, and by presenting the acoustic data as sound by the reproduction unit 106 and listening to it by ear, it can be a great clue to judge whether it is normal or abnormal. In this way, by reproducing (presenting) the acoustic data corresponding to the plot (frame) selected by the user as sound, it is possible to assist the labeling work described later in which the knowledge of the user is reflected.

It should be noted that the present invention is not limited to reproducing only acoustic data as sound, and when the detection signal detected by the signal receiving unit 122 is vibration data, the waveform of the vibration data is reproduced as sound data. May be.

The reproduction unit 106 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of labeling processing unit>
FIG. 11 is a diagram illustrating the operation of the labeling processing unit of the abnormality detection device according to the embodiment. The operation of the labeling processing unit 107 of the abnormality detection device 10 will be described with reference to FIG.

The labeling processing unit 107 of the abnormality detection device 10 is based on the presentation of information (information corresponding to sample data) corresponding to the plot selected by the user by at least one of the basic information display unit 105 and the reproduction unit 106. This is a functional unit that reflects the label change corresponding to the cluster to which the selected plot belongs by operating the input unit 123 from. For example, when the frequency spectrum of each frame included in the frequency characteristics is classified into clusters, the clustering processing unit 102 described above stores a labeling table for associating the classified clusters with labels as shown in FIG. It shall be stored in part 134. At this time, for example, it is assumed that the label corresponding to each cluster in the labeling table saved by the clustering processing unit 102 is written as "normal" as an initial value. Then, when the user determines that the data of the plot selected by the operation to the input unit 123 (and the plot of the same cluster around it) is "abnormal" data, the clustering processing unit 102 sets the labeling table. , The label corresponding to the cluster of the plot is changed to "abnormal" and updated by the operation from the user to the input unit 123. In the labeling table, the label corresponding to each cluster is set to "normal" as the initial value, but it is not limited to this, and the label is set to "abnormal" as the initial value. As a matter of fact, the label of the cluster judged to be normal may be changed to "normal" by the user as appropriate. Alternatively, the label may be left blank by default, and the user may set the label for each cluster as appropriate.

In the above description, the labeling processing unit 107 sets the label for each cluster in the labeling table, but the label is not limited to this, and the label is set for each plot (for each frame) selected by the user. Although it may be possible, it is more realistic to set a label for each cluster because the number of frames obtained from the acoustic data is enormous.

The labeling processing unit 107 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of model generator>
FIG. 12 is a diagram illustrating the operation of the model generation unit of the abnormality detection device according to the embodiment. The operation of the model generation unit 108 of the abnormality detection device 10 will be described with reference to FIG.

As shown in FIG. 12, the model generation unit 108 of the abnormality detection device 10 includes a correspondence table for associating frames and clusters stored in the feature amount storage unit 133, and clusters and labels stored in the cluster storage unit 134. It is a functional unit that generates a learning model by learning the acoustic data stored in the acoustic data storage unit 131 as labeled learning data by referring to the labeling table associated with the above. Here, the training data corresponds to each frame (frequency spectrum) included in the spectrogram, which is a frequency characteristic. Here, it is assumed that the model generation unit 108 generates an outlier detector as a learning model. In this case, the model generation unit 108 uses only the training data (normal sample data) to which the “normal” label is attached among the labeled training data, and trains the one-class support vector machine (SVM), the autoencoder, or the like. Learn the model (outlier detector). The learning model (outlier detector) can calculate how much the input training data deviates from the normal sample data.

The model generation unit 108 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of abnormality detection unit>
FIG. 13 is a diagram illustrating the operation of the abnormality detection unit of the abnormality detection device according to the embodiment. The operation of the abnormality detection unit 111 of the abnormality detection device 10 will be described with reference to FIG.

The abnormality detection unit 111 of the abnormality detection device 10 is input using a learning model (outlier detector) generated by the model generation unit 108 and stored in the model storage unit 135 by the model management unit 110 as described later. This is a functional unit that detects an abnormality in the acoustic data (acoustic data stored in the acoustic data storage unit 131) (data 1'to n'shown in FIG. 13), that is, determines whether the data is normal or abnormal. The abnormality detection unit 111 stores the result of abnormality detection performed on the acoustic data using the learning model in the detection result storage unit 136 of the storage unit 130. By storing the abnormality detection result in the detection result storage unit 136 in this way, the person in charge of storage can confirm it later. When the abnormality detection unit 111 is determined to be abnormal using the learning model, the abnormality detection unit 111 notifies the maintenance person (for example, the display unit 124) according to the severity of the abnormality (for example, the degree of deviation from normal). It may be displayed or notified by e-mail.

The abnormality detection unit 111 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of the goodness-of-fit calculation unit>
FIG. 14 is a diagram illustrating the operation of the goodness-of-fit calculation unit of the abnormality detection device according to the embodiment. The operation of the goodness-of-fit calculation unit 112 of the abnormality detection device 10 will be described with reference to FIG.

The goodness-of-fit calculation unit 112 of the abnormality detection device 10 is how suitable the learning model (outlier detector) currently used for abnormality detection by the abnormality detection unit 111 is for the acoustic data currently input. It is a functional part that calculates the goodness of fit, which is an index indicating whether the model is in use. For example, the goodness-of-fit calculation unit 112, as shown in FIG. 14, has centroids in each cluster for plots showing the training data used to train the training model currently in use, and the acoustic data currently input. The goodness of fit is calculated based on the average value of the distances (for example, Euclidean distance or Mahalanobis distance, etc.) of each frame to the plot based on, or the average likelihood calculated by GMM.

The goodness-of-fit calculation unit 112 determines a threshold value for the calculated goodness of fit, so that the learning model currently in use is based on an appropriate learning model for detecting an abnormality in the input acoustic data. The fact that there is a divergence (for example, a numerical value indicating the degree of divergence, or the goodness of fit itself) (an example of information on the goodness of fit) is displayed on the display unit 124, and the model selection unit 113 described later is relearned. As a learning model to be used for, or as a learning model for performing appropriate abnormality detection on the acoustic data, it may be prompted to select from the learning models generated in the past. Further, the goodness-of-fit calculation unit 112 was used for learning the learning model currently used, for example, in a low-dimensional feature space as an indication that the learning model deviates from the appropriate learning model for detecting anomalies. In addition to displaying the sample data, the sample data based on the acoustic data input after that may be displayed in another color.

The goodness-of-fit calculation unit 112 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of incidental information acquisition unit>
FIG. 15 is a diagram illustrating the operation of the incidental information acquisition unit of the abnormality detection device according to the embodiment. The operation of the incidental information acquisition unit 109 of the abnormality detection device 10 will be described with reference to FIG.

The incidental information acquisition unit 109 of the abnormality detection device 10 is a functional unit that acquires incidental information for management in association with the learning model (outlier detector) generated by the model generation unit 108. Here, the incidental information is information that can be used as a clue for grasping the applicable range of the learning model, and is information that is expected to have some correlation with the input acoustic data. For example, the incidental information is information that can be acquired via a network such as "temperature", "weather", "time", etc. when the learning data used for learning the learning model is acquired, or the machine tool 20. Various parameters such as operating status and set values that can be obtained from, and arbitrary labels that have been input by the user in advance. Further, in the example of the incidental information shown in FIG. 15, the learning model ID is "Classifier A", the generated date and time is "YYYY / MM / DD HH: mm: ss", the learning model type is "AutoEncoder", and the epoch. As "1000", as a feature amount, "Log Spectram", as a weather, "Rainy", and as a model of a machine tool 20, "Middle Rotation" is included.

The incidental information acquisition unit 109 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of model management unit>
FIG. 16 is a diagram illustrating the operation of the model management unit of the abnormality detection device according to the embodiment. The operation of the model management unit 110 of the abnormality detection device 10 will be described with reference to FIG.

The model management unit 110 of the anomaly detection device 10 acquired the learning model (outlier value detector) generated by the model generation unit 108 by the learning data used for learning the learning model and the incidental information acquisition unit 109. It is a functional unit that stores and manages incidental information in the model storage unit 135 of the storage unit 130 in association with the incidental information. Further, as the learning data associated with the learning model, for example, as shown in FIG. 16, even if it is a correspondence table generated by the multivariate analysis processing unit 103 and a labeling table with a label written by the labeling processing unit 107. Good.

The model management unit 110 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of model selection section>
FIG. 17 is a diagram illustrating the operation of the model selection unit of the abnormality detection device according to the embodiment. The operation of the model selection unit 113 of the abnormality detection device 10 will be described with reference to FIG.

The model selection unit 113 of the abnormality detection device 10 is currently using a learning model as a learning model for detecting an abnormality in acoustic data for which the goodness of fit calculated by the goodness of fit calculation unit 112 is currently input. Is not suitable, it is a functional unit that selects an initial model for re-learning based on the input acoustic data from the learning models managed by the model storage unit 135. As a specific scale for selection, for example, the incidental information associated with the learning model managed in the model storage unit 135 is the same as or similar to the incidental information newly acquired by the incidental information acquisition unit 109. The distance (eg, Euclidean distance or Maharanobis) between the distribution of training data used to train a learning model or a managed learning model and sample data (frames) based on newly input acoustic data. The learning model with the smallest distance, etc.), etc. In this way, the learning time is re-learned by the model selection unit 113 as an initial model of a learning model having the same or similar incidental information or a learning model whose distribution of learning data is close to the new input data. Can be shortened, and the risk of falling into a locally optimal solution can also be reduced.

Note that the re-learning of the learning model is not limited to re-learning only with the newly input acoustic data. For example, the learning used for learning the learning model used for abnormality detection until now. The data may be mixed and used for re-learning. Further, the operation of selecting the learning model stored in the model storage unit 135 by the model selection unit 113 as the initial model for re-learning has been described, but the operation is not limited to this. For example, the incidental information associated with the model storage unit 135 was used for learning a learning model in which the incidental information newly acquired by the incidental information acquisition unit 109 is substantially the same as the learning model or a learning model managed. If there is a learning model in which the distance (for example, Euclidean distance or Maharanobis distance) can be regarded as almost 0 between the distribution of the training data and the sample data (frame) based on the newly input acoustic data, retrain. Instead, the learning model may be used as it is.

The model selection unit 113 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of data adjustment unit>
FIG. 18 is a diagram illustrating the operation of the data adjusting unit of the abnormality detection device according to the embodiment. The operation of the data adjusting unit 114 of the abnormality detection device 10 will be described with reference to FIG.

The data adjustment unit 114 of the abnormality detection device 10 is a functional unit that adjusts the number of learning data used for the re-learning when re-learning using the learning model selected by the model selection unit 113 as the initial model. As a method of adjusting the training data, the data adjustment unit 114 adjusts the data by, for example, a method of removing the oldest time stamps attached to the training data in order, or a method of removing the learning data randomly selected using a random number. I do. Alternatively, when the data adjustment unit 114 clusters the entire training data based on the newly input acoustic data by the clustering processing unit 102, the number of training data belonging to each cluster (count number shown in FIG. 18) is large. By preferentially removing the training data from the data, it is possible to effectively remove similar training data while ensuring variations in the training data.

As described above, when the learning model is retrained, the training data increases monotonically as new training data is added in sequence, and eventually the capacity of available resources (mainly memory capacity) is exceeded. It is possible to prevent the calculation from becoming impossible.

The data adjustment unit 114 is realized, for example, by the CPU 51 shown in FIG. 2 executing a program.

<Operation of other functional parts>
Returning to FIG. 4, the operation of other functional units of the abnormality detection device 10 will be described.

The communication unit 121 is a functional unit that performs data communication with the machine tool 20 and receives the above-mentioned incidental information. For example, the communication unit 121 receives context information from the machine tool 20. Here, the context information is information including identification information, operation information, and the like about the machine tool 20. The context information includes, for example, the identification information (model, etc.) of the machine tool 20, the identification information of the drive unit 204, the operating state at that time, the rotation speed and the machining speed, the diameter and material of the tool 23 driven by the drive unit 204, and the like. Processing conditions, sequence number, cycle number, tool number, etc. are included. The communication unit 121 is realized by the communication I / F 54 shown in FIG.

The signal receiving unit 122 detects an operation signal indicating that the machine tool 20 is in the process of machining, and a detection signal (vibration data) that detects vibration or sound generated during the machining process from the sensor 24 installed in the machine tool 20. Alternatively, it is a functional unit that receives (or acoustic data, etc.) via the A / D converter 30. The signal receiving unit 122 is realized by the sensor I / F55 shown in FIG.

The input unit 123 is a functional unit that receives an operation input from the user. The input unit 123 is realized by the input device 57 shown in FIG.

The display unit 124 has a plot showing each frame (frequency spectrum) on the feature space by the feature display 104, basic information corresponding to the plot selected on the feature space by the basic information display 105, and a goodness of fit. It is a functional unit that displays the goodness of fit and the like calculated by the calculation unit 112. The display unit 124 is realized by the display 58 shown in FIG.

The output unit 125 is a functional unit that reproduces acoustic data including a frame corresponding to a selected plot in the feature space by sound according to an instruction from the reproduction unit 106. The output unit 125 is realized by the speaker 62 shown in FIG.

The acoustic data storage unit 131 is a functional unit that stores acoustic data received by the signal reception unit 122. The frequency characteristic storage unit 132 is a functional unit that stores frequency characteristic (spectrogram) data obtained by the frequency analysis processing unit 101. The feature amount storage unit 133 includes the feature amount of each frame included in the frequency characteristics obtained by the frequency analysis processing unit 101, and the two-dimensional or three-dimensional feature calculated by the multivariate analysis processing unit 103 by the multivariate analysis processing unit 103. It is a functional unit that stores the amount and so on. The cluster storage unit 134 is a functional unit that stores cluster information classified by the clustering processing unit 102. The model storage unit 135 associates the learning model (outlier value detector) generated by the model generation unit 108 with the learning data used for learning the learning model and the incidental information acquired by the incidental information acquisition unit 109. It is a functional part that remembers. The detection result storage unit 136 is a functional unit that stores the result of abnormality detection performed by the abnormality detection unit 111 on the acoustic data using the learning model.

The frequency analysis processing unit 101, clustering processing unit 102, multivariate analysis processing unit 103, feature amount display unit 104, basic information display unit 105, reproduction unit 106, and labeling processing unit 107 of the abnormality detection device 10 shown in FIG. 4 , Model generation unit 108, incidental information acquisition unit 109, model management unit 110, abnormality detection unit 111, conformity calculation unit 112, model selection unit 113, and data adjustment unit 114, the CPU 51 shown in FIG. 2 executes a program. It is not limited to being realized by, but may be realized by hardware such as IC (Integrated Cluster).

Further, each functional unit of the abnormality detection device 10 shown in FIG. 4 conceptually shows a function, and is not limited to such a configuration. For example, a plurality of functional units illustrated as independent functional units in the abnormality detection device 10 shown in FIG. 4 may be configured as one functional unit. On the other hand, in the abnormality detection device 10 shown in FIG. 4, the function of one functional unit may be divided into a plurality of functions and configured as a plurality of functional units.

Further, each functional unit included in the abnormality detection device 10 does not necessarily have to be realized on the abnormality detection device 10, and for example, a part of the functional unit may be realized by the cloud 40. For example, a signal receiving unit 122 that acquires a detection signal from the machine tool 20, a communication unit 121 and an incidental information acquisition unit 109 that receive incidental information, and an abnormality detecting unit 111 that detects an abnormality in acoustic data using a learning model. The functional units other than the above may be realized on the cloud 40.

(Structure and operation of functional blocks of machine tools)
FIG. 19 is a diagram showing an example of the functional block configuration of the machine tool according to the embodiment. The configuration and operation of the functional block of the machine tool 20 according to the present embodiment will be described with reference to FIG.

As shown in FIG. 19, the machine tool 20 has a numerical control unit 201, a communication unit 202, a drive control unit 203, a drive unit 204, a signal transmission unit 205, and a detection unit 211.

The numerical control unit 201 is a functional unit that executes machining by the drive unit 204 by numerical control (NC). For example, the numerical control unit 201 generates and outputs numerical control data for controlling the operation of the drive unit 204. Further, the numerical control unit 201 transmits the context information to the abnormality detection device 10 via the communication unit 202, and the operation signal is transmitted to the signal transmission unit 205 during the processing of the processing cycle defined by the NC program. It is transmitted to the abnormality detection device 10 via. When machining an object to be machined, the numerical control unit 201 changes the type of the drive unit 204 to be driven or the drive state (rotation speed, rotation speed, etc.) of the drive unit 204 according to the machining process. Each time the numerical control unit 201 changes the operation type, the numerical control unit 201 sequentially transmits the context information corresponding to the changed operation type to the abnormality detection device 10 via the communication unit 202. The numerical control unit 201 is realized, for example, by the CPU 71 shown in FIG. 3 executing a program (NC program).

The communication unit 202 is a functional unit that performs data communication with the abnormality detection device 10. For example, the communication unit 202 transmits the context information corresponding to the operation at that time to the abnormality detection device 100 under the control of the numerical control unit 201. The communication unit 202 is realized by, for example, a communication I / F 74 shown in FIG. 3 and a program running on the CPU 71.

The drive control unit 203 is a functional unit that drives and controls the drive unit 204 based on the numerical control data obtained by the numerical control unit 201. The drive control unit 203 is realized by, for example, the drive control circuit 75 shown in FIG.

The drive unit 204 is a functional unit that is the target of drive control by the drive control unit 203. The drive unit 204 drives the tool 23 under the control of the drive control unit 203. The drive unit 204 may be, for example, any motor or the like, which is used for processing and is subject to numerical control. Two or more drive units 204 may be provided. The drive unit 204 is an actuator that is drive-controlled by the drive control unit 203, and is realized by, for example, the motor 76 shown in FIG.

The signal transmission unit 205 is a functional unit that transmits an operation signal to the abnormality detection device 10 when the machining process of the machining cycle defined by the NC program is being executed by the numerical control unit 201. The signal transmission unit 205 is realized by a program that operates on the signal I / F77 and the CPU 71 shown in FIG.

The detection unit 211 is a functional unit that detects a physical quantity such as vibration or sound generated by the tool 23 held by the machine tool 20 and outputs the information of the detected physical quantity as a detection signal to the A / D converter 30. The detection unit 211 is realized by the sensor 24 shown in FIG. The number of detection units 211 is arbitrary. For example, a plurality of detection units 211 that detect the same physical quantity may be provided, or a plurality of detection units 211 that detect different physical quantities may be provided.

It should be noted that each functional unit of the machine tool 20 shown in FIG. 19 conceptually shows a function, and is not limited to such a configuration. For example, a plurality of functional units illustrated as independent functional units in the machine tool 20 shown in FIG. 19 may be configured as one functional unit. On the other hand, in the machine tool 20 shown in FIG. 19, the functions possessed by one functional unit may be divided into a plurality of functions and configured as a plurality of functional units.

(Overall operation of abnormality detection device)
20 and 21 are diagrams showing an example of the flow of the overall operation of the abnormality detection device according to the embodiment. The overall operation flow of the abnormality detection device 10 according to the present embodiment will be described with reference to FIGS. 20 and 21.

<Step S11>
The frequency analysis processing unit 101 of the abnormality detection device 10 reads acoustic data from the acoustic data storage unit 131 in which the detection signal (here, acoustic data) detected by the sensor 24 of the machine tool 20 is stored, performs frequency analysis, and performs frequency analysis. Obtain frequency characteristics such as spectral grams. The frequency analysis processing unit 101 stores the obtained frequency characteristic (spectrogram) in the frequency characteristic storage unit 132 of the storage unit 130, and stores the feature amount of each frame (frequency spectrum) included in the frequency characteristic in the feature amount storage. Store in unit 133. Then, the process proceeds to step S12.

<Step S12>
The clustering processing unit 102 of the abnormality detection device 10 reads out the frequency characteristics obtained by the frequency analysis processing unit 101 from the frequency characteristic storage unit 132, and reads the frequency spectrum of each frame included in the frequency characteristics into a plurality of groups (clusters). Classify (clustering) into. The clustering processing unit 102 stores information about what kind of cluster each frame of the frequency characteristic is classified in the cluster storage unit 134 of the storage unit 130. Then, the process proceeds to step S13.

<Step S13>
The multivariate analysis processing unit 103 of the abnormality detection device 10 reads out the frequency characteristics obtained by the frequency analysis processing unit 101 from the frequency characteristic storage unit 132, and each high-dimensional (for example, 512-dimensional) frequency spectrum included in the frequency characteristics. Is subjected to principal component analysis or multivariate analysis such as an autoencoder to calculate a two-dimensional or three-dimensional feature amount that can be visualized. The multivariate analysis processing unit 103 stores the two-dimensional or three-dimensional feature amount calculated by the multivariate analysis for each frequency spectrum in the feature amount storage unit 133 of the storage unit 130. For example, as shown in FIG. 7, the multivariate analysis processing unit 103 includes data (acoustic data, spectrogram), frames, clusters classified by the clustering processing unit 102, and a first principal component obtained by principal component analysis. The feature amount storage unit 133 stores a corresponding table in which the component to the third principal component are associated with each other. Either of the clustering by the clustering processing unit 102 and the multivariate analysis by the multivariate analysis processing unit 103 may be executed first. Then, the process proceeds to step S21.

<Step S21>
The feature amount display unit 104 of the abnormality detection device 10 refers to the correspondence table stored in the feature amount storage unit 133, and uses the feature amount calculated by the multivariate analysis processing unit 103 for the frequency spectrum of each frame as the source. This is a functional unit to be displayed on the display unit 124 in such a manner that the cluster to which the frequency spectrum of the above belongs can be discriminated. For example, as shown in FIG. 8, the feature amount display unit 104 is on a three-dimensional feature amount space centered on the feature amount (first principal component to third principal component) calculated by the multivariate analysis processing unit 103. A plot showing each frame (frequency spectrum) is displayed (presented) in different colors or shapes according to the cluster to which the frame belongs. Then, the process proceeds to step S22.

<Step S22>
The basic information display unit 105 of the anomaly detection device 10 is selected by an operation from the user to the input unit 123 from the plots showing each frame (frequency spectrum) displayed in the feature amount space by the feature amount display unit 104. Basic information about the frame corresponding to the plot is displayed and presented on the display unit 124. Then, the process proceeds to step S23.

<Step S23>
The reproduction unit 106 of the abnormality detection device 10 is selected by an operation from the user to the input unit 123 from the plots showing each frame (frequency spectrum) displayed in the feature amount space by the feature amount display unit 104. Acoustic data including a frame corresponding to the plot is reproduced and presented by sound from the output unit 125 for a predetermined time before and after (for example, 1 second before and after) including the time corresponding to the plot. The display of basic information by the basic information display unit 105 and the reproduction of acoustic data by the reproduction unit 106 are not limited to the above-mentioned order, and only one of them may be presented. Then, the process proceeds to step S24.

<Step S24>
The labeling processing unit 107 of the abnormality detection device 10 is a labeling table stored in the cluster storage unit 134 based on the presentation of information corresponding to the plot selected by the user by at least one of the basic information display unit 105 and the reproduction unit 106. In, the label corresponding to the cluster to which the selected plot belongs is set by the operation from the user to the input unit 123. Further, when a different plot is selected by the operation input to the input unit 123 by the user, the process returns to step S22, and when the work of setting the label corresponding to the cluster is completed, the process proceeds to step S31.

<Step S31>
As shown in FIG. 12, the model generation unit 108 of the abnormality detection device 10 includes a correspondence table for associating frames and clusters stored in the feature amount storage unit 133, and clusters and labels stored in the cluster storage unit 134. Refers to the labeling table associated with the above, learns the acoustic data stored in the acoustic data storage unit 131 as labeled learning data, and generates a learning model. Then, the process proceeds to step S32.

<Step S32>
The incidental information acquisition unit 109 of the abnormality detection device 10 acquires incidental information for management in association with the learning model (outlier detector) generated by the model generation unit 108. Then, the process proceeds to step S33.

<Step S33>
The model management unit 110 of the anomaly detection device 10 acquired the learning model (outlier value detector) generated by the model generation unit 108 by the learning data used for learning the learning model and the incidental information acquisition unit 109. It is stored and managed in the model storage unit 135 of the storage unit 130 in association with the incidental information. Further, as the learning data associated with the learning model, for example, as shown in FIG. 16, even if it is a correspondence table generated by the multivariate analysis processing unit 103 and a labeling table with a label written by the labeling processing unit 107. Good. Then, the learning model generated by the model generation unit 108 and stored in the model storage unit 135 by the model management unit 110 is used for the subsequent abnormality detection for the acoustic data. Then, the process proceeds to step S34.

<Step S34>
The abnormality detection unit 111 of the abnormality detection device 10 is the acoustic data (outlier detector) generated by the model generation unit 108 and stored in the model storage unit 135 by the model management unit 110. Anomaly detection, that is, determination of normality or abnormality is performed on the acoustic data stored in the acoustic data storage unit 131). After that, the abnormality is detected by using the acoustic data input (received) from the signal receiving unit 122 and stored in the acoustic data storage unit 131. Then, the process proceeds to step S41.

<Step S41>
The goodness-of-fit calculation unit 112 of the abnormality detection device 10 indicates how suitable the learning model (outlier detector) currently used for abnormality detection by the abnormality detection unit 111 is for the currently input acoustic data. It is a functional part that calculates the goodness of fit, which is an index indicating whether the model is a model. For example, the goodness-of-fit calculation unit 112, as shown in FIG. 14, has centroids in each cluster for plots showing the training data used to train the training model currently in use, and the acoustic data currently input. The goodness of fit is calculated based on the average value of the distances (for example, Euclidean distance or Mahalanobis distance, etc.) of each frame to the plot based on, or the average likelihood calculated by GMM. Then, the process proceeds to step S42.

<Step S42>
The model selection unit 113 of the abnormality detection device 10 is currently using a learning model as a learning model for detecting an abnormality in acoustic data for which the goodness of fit calculated by the goodness of fit calculation unit 112 is currently input. Is not suitable, it is a functional unit that selects an initial model for re-learning based on the input acoustic data from the learning models managed by the model storage unit 135. As a specific scale for selection, for example, the incidental information associated with the learning model managed in the model storage unit 135 is the same as or similar to the incidental information newly acquired by the incidental information acquisition unit 109. The distance (eg, Euclidean distance or Maharanobis) between the distribution of training data used to train a learning model or a managed learning model and sample data (frames) based on newly input acoustic data. The learning model with the smallest distance, etc.), etc. Then, the process proceeds to step S43.

It should be noted that the incidental information associated with the model storage unit 135 was used for learning a learning model in which the incidental information newly acquired by the incidental information acquisition unit 109 is substantially the same as the learning model or a learning model managed. If there is a learning model in which the distance (for example, Euclidean distance or Maharanobis distance) can be regarded as almost 0 between the distribution of the training data and the sample data (frame) based on the newly input acoustic data, retrain. Instead, the learning model may be used as it is. In this case, the process proceeds to step S34.

<Step S43>
The data adjustment unit 114 of the abnormality detection device 10 adjusts the number of learning data used for the re-learning when the learning model selected by the model selection unit 113 is used as the initial model for re-learning. Then, the process returns to step S31, and the learning model is relearned.

In the processing flow shown in FIGS. 20 and 21, the entire operation of the abnormality detection device 10 is executed.

As described above, in the abnormality detection device 10 according to the present embodiment, a multivariate analysis such as principal component analysis is performed on a high-dimensional feature amount calculated by frequency analysis of a detection signal (acoustic data, vibration data, etc.). To visualize and display their distribution on a 2D or 3D low-dimensional feature space, and select any plot on the low-dimensional feature space to plot it. The information corresponding to is to be presented. As a result, the degree of variation of individual training data and the degree of deviation from the typical normal sample data forming the largest cluster can be visually confirmed, and the data is selected on the low-dimensional feature amount space. Based on the presentation of information corresponding to the plot, the labeling work of the sample data used for learning the learning model for abnormality detection of the target machine (machine machine 20) can be streamlined, and high abnormality detection accuracy can be maintained. Early detection of anomalies and their precursors can be achieved. The presentation of the information corresponding to the selected plot includes the display operation of the basic information regarding the frame corresponding to the selected plot by the basic information display unit 105, and the sound including the frame corresponding to the selected plot by the reproduction unit 106. The data is realized by the operation of reproducing the data by sound for a predetermined time before and after (for example, 1 second before and after) including the time corresponding to the plot.

Further, when the training data of each cluster is plotted on the low-dimensional feature space, the color or shape is changed and displayed according to the cluster to which the training data belongs. As a result, it is possible to recognize at a glance which cluster each plot belongs to, and the labeling work can be further made more efficient.

In addition, there is no guarantee that the learning model once generated can be used permanently thereafter. This is because the operating noise of machines, etc. gradually changes according to their aging deterioration even during normal operation, and it depends on various fluctuation factors such as temperature, weather, time, and operating conditions of machines, etc. It can also fluctuate periodically and irregularly. In order to deal with this, in the abnormality detection device 10 according to the present embodiment, the learning model (outlier detector) currently used for abnormality detection by the abnormality detection unit 111 is currently input by the goodness-of-fit calculation unit 112. The goodness of fit, which is an index showing how suitable the model is for the existing acoustic data, is calculated, and the learning model currently in use is appropriate for detecting anomalies in the input acoustic data. The fact that there is a deviation from the learning model (for example, a numerical value indicating the degree of deviation, the goodness of fit itself, etc.) is displayed on the display unit 124. As a result, it is possible to quantitatively grasp how well the already generated learning model fits the current situation, and it is possible to know the appropriate timing of re-learning. Further, by performing the threshold value determination on the goodness of fit calculated by the goodness-of-fit calculation unit 112, it is possible to automatically relearn the learning model at an appropriate timing without intervening human judgment. ..

Further, the abnormality detection device 10 according to the present embodiment is currently used as a learning model for detecting an abnormality in the acoustic data for which the degree of conformity calculated by the degree of conformity calculation unit 112 is currently input. When it is shown that the learning model is not suitable, the initial model for re-learning based on the input acoustic data is selected from the learning models managed by the model storage unit 135. As a specific scale for selection, for example, the incidental information associated with the learning model managed in the model storage unit 135 is the same as or similar to the incidental information newly acquired by the incidental information acquisition unit 109. The distance (eg, Euclidean distance or Maharanobis) between the distribution of training data used to train a learning model or a managed learning model and sample data (frames) based on newly input acoustic data. The learning model with the smallest distance, etc.), etc. In this way, the learning time is re-learned by the model selection unit 113 as an initial model of a learning model having the same or similar incidental information or a learning model whose distribution of learning data is close to the new input data. Can be shortened, and the risk of falling into a locally optimal solution can also be reduced. Further, the incidental information associated with the model storage unit 135 was used for learning a learning model in which the incidental information newly acquired by the incidental information acquisition unit 109 is substantially the same as the learning model or a learning model managed. If there is a learning model in which the distance (for example, Euclidean distance or Maharanobis distance) can be regarded as almost 0 between the distribution of the training data and the sample data (frame) based on the newly input acoustic data, retrain. Instead, the learning model may be used as it is. This makes it possible to deal with periodic fluctuation factors and irregular fluctuation factors having some causal relationship with various incidental information without re-learning.

The outlier detector has been described as a learning model generated by the abnormality detection device 10 according to the present embodiment, but if the labeling processing unit 107 can add a plurality of types of labels to the sample data, It may generate a different learning model (for example, a two-class classifier) or the like.

Further, each function of the above-described embodiment can be realized by one or a plurality of processing circuits. Here, the "processing circuit" is a processor programmed to execute each function by software such as a processor implemented by an electronic circuit, or an ASIC (Application Specific Integrated) designed to execute each function described above. Devices including devices such as Circuits, DSPs (Digital Signal Processors), FPGAs (Field-Programmable Gate Arrays), SoCs (System on a chip), GPUs (Graphics Processing Units), and conventional circuit modules.

Further, the programs executed by the abnormality detection device 10 and the machine tool 20 of the above-described embodiment are files in an installable format or an executable format, such as a CD-ROM (Computer Disc Read Only Memory) and a flexible disk (FD). , CD-R (Compact Disk-Recordable), DVD (Digital Versaille Disk), or the like, which may be recorded on a computer-readable recording medium and provided as a computer program product.

Further, the program executed by the abnormality detection device 10 and the machine tool 20 of the above-described embodiment is stored on a computer connected to a network such as the Internet, and is configured to be provided by downloading via the network. May be good. Further, the program executed by the abnormality detection device 10 and the machine tool 20 of the above-described embodiment may be configured to be provided or distributed via a network such as the Internet.

Further, the program executed by the abnormality detection device 10 and the machine tool 20 of the above-described embodiment has a module configuration including each of the above-mentioned functional units, and the CPU (processor) is the above-mentioned ROM as the actual hardware. By reading and executing the program from, each of the above parts is loaded on the main storage device, and each part is generated on the main storage device.

1 Anomaly detection system 10 Anomaly detection device 20 Machine tool 22 Holder 23 Tool 24 Sensor 25 NC control device 30 A / D converter 40 Cloud 51 CPU
52 ROM
53 RAM
54 Communication I / F
55 Sensor I / F
56 I / O I / F
57 Input device 58 Display 59 Auxiliary storage device 60 Bus 61 Voice I / F
62 Speaker 71 CPU
72 ROM
73 RAM
74 Communication I / F
75 Drive control circuit 76 Motor 77 Signal I / F
78 Bus 101 Frequency analysis processing unit 102 Clustering processing unit 103 Multivariate analysis processing unit 104 Feature quantity display unit 105 Basic information display unit 106 Playback unit 107 Labeling processing unit 108 Model generation unit 109 Ancillary information acquisition unit 110 Model management unit 111 Abnormality detection Unit 112 Conformity calculation unit 113 Model selection unit 114 Data adjustment unit 121 Communication unit 122 Signal reception unit 123 Input unit 124 Display unit 125 Output unit 130 Storage unit 131 Acoustic data storage unit 132 Frequency characteristic storage unit 133 Feature quantity storage unit 134 Cluster Storage unit 135 Model storage unit 136 Detection result storage unit 201 Numerical control unit 202 Communication unit 203 Drive control unit 204 Drive unit 205 Signal transmission unit 211 Detection unit

Japanese Patent No. 4369661 Japanese Patent No. 6097517

Claims (16)

A frequency analysis processing unit that performs frequency analysis on the detection signal detected from the abnormality detection target and obtains frequency characteristics,
A clustering processing unit that classifies the sample data into one or more clusters by using each frame included in the frequency characteristics obtained by the frequency analysis processing unit as sample data.
A multivariate analysis processing unit that performs multivariate analysis on the frequency characteristics obtained by the frequency analysis processing unit and extracts features, and a multivariate analysis processing unit.
Based on the feature amount extracted by the multivariate analysis processing unit, a feature amount display unit that visualizes the sample data so that it can be determined which cluster it is classified into.
Among the sample data displayed by the feature amount display unit, a presentation unit that presents information corresponding to the sample data selected by an input signal from the input unit, and a presentation unit.
Based on the information presented by the presenting unit, a labeling processing unit that sets a label for the sample data for each cluster according to an input signal from the input unit, and a labeling processing unit.
A model generation unit that generates a learning model by learning using the sample data labeled by the labeling processing unit, and a model generation unit.
Anomaly detection device equipped with. The presenting unit receives at least one of the waveform of the detection signal corresponding to the sample data and the time series data of the frequency spectrum of the sample data as the information corresponding to the sample data selected by the input unit. The abnormality detection device according to claim 1, which is a basic information display unit that displays included basic information on a display device. The abnormality according to claim 1, wherein the presenting unit is a reproduction unit that reproduces the detection signal corresponding to the sample data as sound from an output device as information corresponding to the sample data selected by the input unit. Detection device. Any one of claims 1 to 3, further comprising an abnormality detection unit that detects an abnormality in a detection signal input from the abnormality detection target using the learning model generated by the model generation unit. Anomaly detection device described in. The goodness of fit of the learning model used in the abnormality detection by the abnormality detection unit is calculated for the detection signal currently input from the abnormality detection target, and the information on the goodness of fit is displayed on the display device. The abnormality detection device according to claim 4, further comprising a unit. A model management unit that stores the learning model generated by the model generation unit in the storage unit in association with incidental information that is predicted to have a correlation with the detection signal input for learning the learning model. The abnormality detection device according to claim 5, further provided. The abnormality detection device according to claim 6, wherein the model management unit further stores the learning model generated by the model generation unit in the storage unit in association with sample data used for learning the learning model. The learning model currently used is not suitable as the learning model for detecting the abnormality in the detection signal currently input from the abnormality detection target by the degree of conformity calculated by the conformity calculation unit. In the case of showing that, a model selection unit for selecting a learning model suitable for the detection signal as an initial model from the learning models stored in the storage unit is further provided.
The abnormality detection device according to claim 6 or 7, wherein the model generation unit relearns a new learning model using the initial model and at least the detection signal currently input from the abnormality detection target. The model generation unit relearns a new learning model using the initial model, sample data used for learning the initial model, and sample data based on the detection signal currently input from the abnormality detection target. The abnormality detection device according to claim 8. The learning model currently used is not suitable as a learning model for detecting an abnormality in a detection signal currently input from the abnormality detection target with the degree of conformity calculated by the conformity calculation unit. In the case of showing that, a model selection unit for selecting a learning model suitable for the detection signal from the learning models stored in the storage unit as a learning model to be used next for abnormality detection is further provided.
The abnormality detection device according to claim 6 or 7, wherein the abnormality detection unit uses a learning model selected by the model selection unit to detect an abnormality in a detection signal input from the abnormality detection target. The method 8 or 9 further includes a data adjustment unit that adjusts the number of sample data used for the re-learning when the learning model selected by the model selection unit is used as the initial model for re-learning. Anomaly detection device. The abnormality detection device according to claim 11, wherein the data adjustment unit removes the oldest time stamps attached to the sample data in order. When the data adjustment unit classifies the sample data input from the abnormality detection target into one or more clusters by the clustering processing unit, the sample data is based on the number of sample data belonging to each cluster. The abnormality detection device according to claim 11. The abnormality detection device according to any one of claims 1 to 13, wherein the model generation unit generates an outlier detector as the learning model by learning. A frequency analysis processing unit that performs frequency analysis on the detection signal detected from the abnormality detection target and obtains frequency characteristics,
A clustering processing unit that classifies the sample data into one or more clusters by using each frame included in the frequency characteristics obtained by the frequency analysis processing unit as sample data.
A multivariate analysis processing unit that performs multivariate analysis on the frequency characteristics obtained by the frequency analysis processing unit and extracts features, and a multivariate analysis processing unit.
Based on the feature amount extracted by the multivariate analysis processing unit, a feature amount display unit that visualizes the sample data so that it can be determined which cluster it is classified into.
Among the sample data displayed by the feature amount display unit, a presentation unit that presents information corresponding to the sample data selected by an input signal from the input unit, and a presentation unit.
Based on the information presented by the presenting unit, a labeling processing unit that sets a label for the sample data for each cluster according to an input signal from the input unit, and a labeling processing unit.
A model generation unit that generates a learning model by learning using the sample data labeled by the labeling processing unit, and a model generation unit.
Anomaly detection system with. On the computer
A frequency analysis processing step that performs frequency analysis on the detection signal detected from the abnormality detection target to obtain frequency characteristics, and
A clustering processing step of classifying the sample data into one or more clusters by using each frame included in the frequency characteristics as sample data.
A multivariate analysis processing step that performs multivariate analysis on the frequency characteristics and extracts features,
Based on the extracted features, a feature display step that visualizes the sample data so that it can be determined which cluster it is classified into, and
Among the displayed sample data, a presentation step for presenting information corresponding to the sample data selected by an input signal from the input unit, and a presentation step.
Based on the presented information, a labeling process step of setting a label of the sample data for each cluster according to an input signal from the input unit, and
A model generation step of generating a learning model by learning using the sample data with a label set, and
A program to execute.