Classification device, learning model generation device, classification method, and learning model generation method

Abstract

A classification device according to the present invention is provided with: an acquisition unit that acquires time-axis waveform data of an object of interest; a conversion unit that converts the time-axis waveform data into first frequency characteristics data; a spectrum calculation unit that divides the first frequency characteristics data into a plurality of divided bins by a prescribed bandwidth and that calculates the maximum values of the spectrum in the individual divided bins; an approximation processing unit that approximates the first frequency characteristics data on the basis of the maximum values of the spectrum in the individual divided bins and that outputs second frequency characteristics data; a generation unit that generates third frequency characteristics data from the second frequency characteristics data by using a learning model; and a classification unit that classifies the object of interest by using the second frequency characteristics data and the third frequency characteristics data. The learning model is prepared through training with approximated training data.

Description

[Technical Field]

[0001] This disclosure relates to a classification device, a learning model generation device, a classification method, and a learning model generation method. [Background technology]

[0002] Patent Document 1 discloses a method for inspecting abnormal noise in rotating machinery, which determines whether the machine is good or bad based on the difference between the spectral distance of the machine under inspection and the average value of the spectral distance of good machines. The method for inspecting abnormal noise in rotating machinery involves detecting the operating sounds of a predetermined number of good rotating machines as good sample data, performing frequency analysis on each good sample data to express it as a power spectrum, performing octave analysis on the power spectrum to obtain predetermined octave band spectra, and then calculating the average value of the good product spectral distance, which is obtained by averaging these average spectra and the sum of the squares of the errors between the average spectrum and the octave band spectra based on each good sample data. The method for inspecting abnormal noise in rotating machinery involves calculating the spectral distance of the machine under inspection obtained by the same process using the operating sound of the machine under inspection as the sample, and then determining whether the machine is good or bad based on the difference between the spectral distance of the machine under inspection and the average value of the spectral distance of good machines.

[0003] Furthermore, Patent Document 2 discloses a sound inspection method for electric razors, which involves dividing the sound components generated from an electric razor during operation into a frequency band occupied by the sound generated from the drive source and a frequency band occupied by the sound generated from the shaving unit, and determining the quality of the electric razor by comparing the volume values ​​in each frequency band with preset volume reference values ​​for the drive source and the shaving unit for quality determination. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 10-002789 [Patent Document 2] Japanese Patent Application Publication No. 61-073031 [Overview of the Initiative]

[0005] This disclosure was devised in view of the conventional circumstances described above, and aims to provide a classification device, a learning model generation device, a classification method, and a learning model generation method that can perform good quality determination of inspected objects regardless of individual differences in the inspected objects.

[0006] This disclosure provides a classification device for classifying objects having a periodically driven drive source, comprising: an acquisition unit for acquiring time-axis waveform data of the object; a conversion unit for converting the time-axis waveform data into first frequency characteristic data; a spectrum calculation unit for dividing the first frequency characteristic data into a plurality of division sections with a predetermined bandwidth and calculating the maximum value of the spectrum for each division section; an approximation processing unit for approximating the first frequency characteristic data based on the maximum value of the spectrum for each division section and outputting second frequency characteristic data; a generation unit for generating third frequency characteristic data from the second frequency characteristic data using a learning model; and a classification unit for classifying the object based on the second frequency characteristic data and the third frequency characteristic data, wherein the learning model is a model that has learned learning data approximated by at least the approximation processing unit.

[0007] Furthermore, this disclosure provides a learning model generation device that performs learning with respect to an object having a periodically driven drive source, comprising: an acquisition unit that acquires time-axis waveform data of the object; a conversion unit that converts the time-axis waveform data into first frequency characteristic data; a spectrum calculation unit that divides the first frequency characteristic data into a plurality of division intervals with a predetermined bandwidth and calculates the maximum value of the spectrum for each division interval; an approximation processing unit that approximates the first frequency characteristic data based on the maximum value of the spectrum for each division interval and outputs second frequency characteristic data; and a learning model generation unit that trains a learning model with a plurality of the second frequency characteristic data.

[0008] Also, the present disclosure provides a classification method performed by a classification device that classifies an object having a periodically driven drive source, the method including: acquiring time-axis waveform data of the object; converting the time-axis waveform data into first frequency characteristic data; dividing the first frequency characteristic data into a plurality of divided sections with a predetermined bandwidth, calculating a maximum value of a spectrum for each of the divided sections; outputting second frequency characteristic data approximating the first frequency characteristic data based on the maximum value of the spectrum for each of the divided sections; generating third frequency characteristic data from the second frequency characteristic data using a learning model; classifying the object based on the second frequency characteristic data and the third frequency characteristic data; and the learning model being a model that has learned at least learning data approximated by the classification device.

[0009] Also, the present disclosure provides a learning model generation method performed by a learning model generation device that performs learning on an object having a periodically driven drive source, the method including: acquiring time-axis waveform data of the object; converting the time-axis waveform data into first frequency characteristic data; dividing the first frequency characteristic data into a plurality of divided sections with a predetermined bandwidth, calculating a maximum value of a spectrum for each of the divided sections; outputting second frequency characteristic data approximating the first frequency characteristic data based on the maximum value of the spectrum for each of the divided sections; and causing a plurality of the second frequency characteristic data to be learned by a learning model.

[0010] According to the present disclosure, regardless of individual differences in the inspection object, it is possible to execute a determination of whether the inspection object is a non-defective product.

Brief Description of the Drawings

[0011] [Figure 1] A diagram for explaining an example of an inspection object and an inspection jig in Embodiment 1 [Figure 2] A diagram for explaining an example of an inspection object and an inspection jig in Embodiment 1 [Figure 3] X-Z cross-sectional view of the inspection jig [Figure 4]Block diagram showing an example of the internal configuration of a non-defective product determination device according to Embodiment 1 [Figure 5] Flowchart showing an example of the overall operation procedure of the terminal device in Embodiment 1 [Figure 6] Flowchart showing an example of the learning model generation processing procedure of the terminal device in Embodiment 1 [Figure 7] Flowchart showing an example of the determination error distribution generation processing procedure of the terminal device in Embodiment 1 [Figure 8] Flowchart showing an example of the non-defective product determination processing procedure of the terminal device in Embodiment 1 [Figure 9] Diagram showing an example of frequency characteristic data [Figure 10] Partial enlarged view of the frequency characteristic data [Figure 11] Diagram showing a comparison example of frequency characteristic data between non-defective products [Figure 12] Diagram showing a comparison example of the low-frequency region of frequency characteristic data between non-defective products [Figure 13] Diagram showing a comparison example of the mid-frequency region of frequency characteristic data between non-defective products [Figure 14] Diagram showing comparison examples of frequency characteristic data of non-defective products and defective products respectively [Figure 15] Diagram for explaining an example of calculating the maximum value for each interval in the frequency characteristic data [Figure 16] Diagram showing a comparison example of frequency characteristic data before and after approximation processing [Figure 17] Diagram showing a comparison example of the frequency characteristic data (input data) of non-defective products and the frequency characteristic data (output data) of the first learning model [Figure 18] Diagram showing a comparison example of the frequency characteristic data (input data) of defective products and the frequency characteristic data (output data) of the first learning model [Figure 19] Diagram showing an example of the first determination error distribution and the second determination error distribution [Figure 20] Diagram showing an example of the correlation between the determination performance α and the bandwidth [Figure 21] Correlation graph showing an example of the correlation between the determination performance α and the bandwidth [Modes for carrying out the invention]

[0012] (Background leading to this disclosure) Traditionally, the analysis of sounds collected using measuring instruments has been used for quality control (judgment) of products equipped with motors and actuators produced in factories, as well as equipment inspections. However, because the collected sounds include individual differences in products and equipment, as well as ambient noise, it has been difficult to quantitatively evaluate abnormal sounds. For this reason, abnormal sound evaluation has been carried out through sensory evaluation using the hearing of inspectors who perform quality control of products and equipment inspections.

[0013] However, sensory evaluation had the problem of inconsistent results depending on the inspector. Furthermore, even with evaluations by a single inspector, there was the problem of inconsistency in the results due to the inspector's physical condition, fatigue, etc.

[0014] To address these challenges, conventional literature 1 performs octave analysis on good sample data of the operating sound of the rotating machinery under inspection. Based on the average value of the good spectral distance using the octave band spectrum and the spectral distance of the machine under inspection based on the operating sound of the rotating machinery, the quality of the machine under inspection is determined.

[0015] However, in conventional literature 1, the frequency characteristics of operating noise are calculated using a bandwidth filter with a constant ratio width that is close to the characteristics of human hearing, so the bandwidth widens in the high-frequency range. Therefore, conventional literature 1 has the problem that the evaluation (inspection) accuracy decreases when evaluating (inspecting) rotating machinery based on differences in noise levels caused by harmonic components of the drive frequencies of motors, actuators, etc.

[0016] On the other hand, a common frequency analysis method using FFT (Fast Fourier Transform) has a narrow bandwidth of frequency resolution, which is uniquely determined by the sampling frequency and the number of sampling points. Therefore, if the harmonic components of the drive frequency of motors, actuators, etc., fluctuate due to individual differences in products and equipment, measurement errors, etc., it may not be possible to evaluate the high-frequency components after the fluctuation in each bandwidth. Consequently, a problem with frequency analysis methods using FFT is that the difference in frequency characteristics between products and equipment evaluated as good becomes large, and when a machine learning model is used that uses the frequency characteristics of products evaluated as good as training data, the accuracy of determining good and defective products in good product inspection decreases.

[0017] Hereinafter, embodiments specifically disclosing the configuration and operation of the classification device, learning model generation device, classification method, and learning model generation method relating to this disclosure will be described in detail, with reference to the drawings as appropriate. However, unnecessarily detailed explanations may be omitted. For example, detailed explanations of already well-known matters and redundant explanations of substantially identical configurations may be omitted. This is to avoid the following explanation becoming unnecessarily verbose and to facilitate understanding by those skilled in the art. The accompanying drawings and the following explanation are provided to enable those skilled in the art to fully understand this disclosure and are not intended to limit the subject matter described in the claims.

[0018] First, an example of the object to be inspected 1 and the inspection jig will be described with reference to Figure 1. Figure 1 is a diagram illustrating an example of the object to be inspected 1 and the inspection jig in Embodiment 1. Figure 2 is a diagram illustrating an example of the object to be inspected 1 and the inspection jig in Embodiment 1.

[0019] In this embodiment 1, an example is given where the object to be inspected 1 is an electric razor, but it goes without saying that the object to be inspected 1 is not limited to this. The object to be inspected 1 can be any device equipped with a drive source such as a motor or actuator that is driven periodically.

[0020] The inspection jig comprises at least a vibration sensor 2, a cradle CDL, and a foamed rubber RB. The inspection jig is fixed to a frame STG made of a material such as aluminum by screwing the cradle CDL, which supports the object to be inspected 1 with the foamed rubber RB in between, to the frame STG.

[0021] The cradle CDL has a mounting surface CDL1 formed to support the object to be inspected 1, conforming to its shape. The object to be inspected 1 is placed on the mounting surface CDL1 of the cradle CDL.

[0022] The vibration sensor 2 is housed in the cradle CDL, partially exposed from the mounting surface CDL1. The portion of the vibration sensor 2 exposed from the mounting surface CDL1 comes into contact with the object to be inspected 1. The vibration sensor 2 is equipped with a piezoelectric element 2A (see Figure 3). The vibration sensor 2 converts the drive vibration of the drive source of the object to be inspected 1, such as a motor or actuator, into time-domain waveform vibration data (an example of time-domain waveform data) using the piezoelectric element 2A, and transmits the converted vibration data to the analog filter 3 (see Figure 4).

[0023] The foamed rubber RB is a so-called vibration-damping rubber that prevents vibrations other than those of the object being inspected 1 from being transmitted to the cradle CDL via the stand STG. As a result, the vibration sensor 2 can more efficiently acquire only the drive vibrations of the object being inspected 1.

[0024] Next, with reference to Figure 3, the mounting surface CDL1 of the vibration sensor 2 and cradle CDL will be described. Figure 3 is an XZ cross-sectional view of the inspection fixture.

[0025] Figure 3 shows a partial cross-section of the cradle CDL with the object to be inspected 1 placed on it, cut at approximately halfway along the Y direction (width direction of the object to be inspected 1), and a partial cross-sectional view (XZ cross-section) of the object to be inspected 1 and the inspection jig, cut at approximately halfway along the Y direction (width direction of the object to be inspected 1).

[0026] The object to be inspected 1 has text, diagrams, etc. printed on its printed surface 1A. The mounting surface CDL1 of the cradle CDL is formed in a shape that does not conform to the shape of the object to be inspected 1 in part, based on the shape of the object to be inspected 1 and the position of the printed surface 1A. When the object to be inspected 1 is placed on the mounting surface CDL1, a gap SP is formed between the mounting surface CDL1 and the surface of the object to be inspected 1 corresponding to the printed surface 1A, resulting in a non-contact state between the mounting surface CDL1 and the printed surface 1A of the object to be inspected 1.

[0027] This prevents the inspection jig from causing the printed surface 1A of the object to be inspected to rub against the mounting surface CDL1 due to the driving vibration of the object to be inspected 1, thereby preventing deterioration of print quality.

[0028] The vibration sensor 2 is equipped with a piezoelectric element 2A in the portion exposed from the mounting surface CDL1. The piezoelectric element 2A converts the vibration of the object to be inspected 1 into an electrical signal.

[0029] Next, with reference to Figure 4, the good product determination system 100 according to Embodiment 1 will be described. Figure 4 is a block diagram showing an example of the internal configuration of the good product determination system 100 according to Embodiment 1. Note that the good product determination system 100 shown in Figure 4 is just one example and is not limited thereto.

[0030] The good product determination system 100 comprises an object to be inspected 1, a vibration sensor 2, an analog filter 3, an A / D (Analog / Digital) converter 4, and a terminal device P1 (an example of a classification device or learning model generation device). It goes without saying that the good product determination system 100 shown in Figure 4 is just one example and is not limited thereto. For example, there may be multiple objects to be inspected 1, vibration sensors 2, analog filters 3, and A / D converters 4.

[0031] The analog filter 3 acquires vibration data (analog signal) of the object under inspection 1 transmitted from the vibration sensor 2. The analog filter 3 performs a process on the acquired vibration data to remove frequency components that are unnecessary for determining whether the object under inspection 1 is good, or to extract only the frequency components that are necessary for determining whether the object under inspection 1 is good. The analog filter 3 transmits the vibration data after performing the above processing to the A / D converter 4.

[0032] The A / D converter 4 converts the vibration data, which is an analog signal transmitted from the analog filter 3, into a digital signal. The A / D converter 4 then transmits the converted vibration data to the communication unit 10 of the terminal device P1.

[0033] The terminal device P1 can be implemented as, for example, a PC (Personal Computer), a notebook PC, or a tablet device. The terminal device P1 comprises a communication unit 10, a processor 11, memory 12, a display unit 13, and an input unit 14. The input unit 14 is not a mandatory component and may be omitted.

[0034] The communication unit 10 (an example of an acquisition unit) is connected to the A / D converter 4 via wired or wireless communication. That is, one example of the communication unit 10 is a wired communication-enabled connector, and another example of the communication unit 10 is a wireless communication-enabled transmitter and receiver. The wireless communication referred to here is not particularly limited to any specific type, such as Wi-Fi® or other wireless LAN technologies.

[0035] Furthermore, the communication unit 10 may be connected to an external storage medium (not shown), such as a CD-ROM (Compact Disc Read Only Memory), USB memory, or SD® card, and may be able to read learning models, vibration data of good or defective products, frequency characteristic data, learning data, etc., stored on the external storage medium.

[0036] The processor 11 is configured using, for example, a CPU (Central Processing Unit) or an FPGA (Field Programmable Gate Array), and works in cooperation with the memory 12 to perform various processing and control. Specifically, the processor 11 refers to the programs and data held in the memory 12 and executes those programs to realize the functions of each part. The parts referred to here include the vibration waveform measurement unit 111, the frequency characteristic conversion unit 112, the drive frequency calculation unit 113, the interval maximum value calculation unit 114, the frequency characteristic bandwidth approximation unit 115, the waveform prediction unit 116, the machine learning model generation unit 117, and the data analysis unit 118.

[0037] The vibration waveform measurement unit 111 acquires vibration data (digital signal) of the time-domain waveform output from the communication unit 10 and outputs it to the frequency characteristic conversion unit 112.

[0038] The frequency response conversion unit 112 (an example of a conversion unit) performs FFT processing on the vibration data output from the vibration waveform measurement unit 111, converting the vibration data into frequency response data FCG1 (see Figure 9). The frequency response conversion unit 112 outputs the converted frequency response data FCG1 to the drive frequency calculation unit 113.

[0039] The drive frequency calculation unit 113 performs a peak search process on the frequency characteristic data FCG1 output from the frequency characteristic conversion unit 112. Based on the peak search processing results, the drive frequency calculation unit 113 calculates the drive frequency DF1 (see Figure 10) of the motor, actuator, etc. of the object under inspection 1. The drive frequency calculation unit 113 outputs the calculated drive frequency information and the frequency characteristic data FCG1 to the section maximum value calculation unit 114.

[0040] The section maximum value calculation unit 114 (an example of a spectrum calculation unit) divides the frequency band of the frequency characteristic data FCG1 by the bandwidth based on the drive frequency output from the drive frequency calculation unit 113, and calculates the maximum value of the spectrum (dB) in each divided section (i.e., each frequency band). The section maximum value calculation unit 114 outputs the information of the maximum value of the spectrum in each divided section to the frequency characteristic bandwidth approximation unit 115.

[0041] Furthermore, if there is a frequency band in the frequency characteristic data where the difference between the frequency characteristics of good products and defective products is sufficiently small, the interval maximum value calculation unit 114 may extract frequency regions other than this frequency band and divide the frequency band of the frequency characteristic data FCG1 by bandwidth. This reduces the processing load required for machine learning on the terminal device P1, allowing it to perform good product determination more efficiently.

[0042] The frequency response bandwidth approximation unit 115 (an example of an approximation processing unit) generates frequency response data AG1 (see Figure 16) by approximating each divided section of the frequency response data FCG1 with the maximum value of the spectrum, based on the information of the maximum value of the spectrum of each divided section output from the section maximum value calculation unit 114. The frequency response bandwidth approximation unit 115 may output the generated frequency response data AG1 to the waveform prediction unit 116, the machine learning model generation unit 117, and the data analysis unit 118, respectively, and may also record it in the memory 12.

[0043] In the following explanation, the frequency response data AG1, which is approximated and generated by the frequency response bandwidth approximation unit 115, may be referred to as "input data" or "approximated data." Furthermore, approximated frequency response data for good products will be described as AG11, approximated frequency response data for defective products as AG12, and approximated frequency response data where good or defective products are not specified, or where the distinction between good and defective products is unknown, will be described as AG1. Similarly, in the frequency response data before approximation, the frequency response data for good products will be described as FCG11, the frequency response data for defective products as FCG12, and the frequency response data where good or defective products are not specified will be described as FCG1.

[0044] The machine learning model generation unit 117 (an example of a learning model generation unit) performs machine learning using one or more recorded and collected frequency characteristic data AG11 (approximate data) of good products as training data. Specifically, the machine learning model generation unit 117 performs machine learning to determine (classify) whether an inspected object 1 is a good product based on the frequency characteristic data AG11 (approximate data) of the inspected object 1. As a result of the learning, the machine learning model generation unit 117 generates a first learning model for determining whether an inspected object 1 is a good product and records it in memory 12.

[0045] Examples of statistical classification techniques include linear classifiers, support vector machines, quadratic classifiers, kernel density estimation, decision trees, artificial neural networks, Bayesian techniques and / or networks, hidden Markov models, binary classifiers, multi-class classifiers, clustering techniques, random forest techniques, logistic regression techniques, linear regression techniques, and gradient boosting techniques. However, the statistical classification techniques used are not limited to these.

[0046] The waveform prediction unit 116 (an example of a prediction unit) inputs the frequency characteristic data AG1 (input data) output from the frequency characteristic bandwidth approximation unit 115 to the first learning model recorded in the memory 12, and acquires the frequency characteristic data FG1 (output data, predicted data) output from the first learning model. The waveform prediction unit 116 outputs the acquired frequency characteristic data FG1 (output data, predicted data) to the data analysis unit 118.

[0047] Furthermore, when the waveform prediction unit 116 receives frequency characteristic data AG11, which represents a good product, it acquires the frequency characteristic data FG11 output from the first learning model. When it receives frequency characteristic data AG12, which represents a defective product, it acquires the frequency characteristic data FG12 output from the first learning model.

[0048] In the following explanation, the frequency response data FG1, FG11, and FG12, which are input to and output by each learning model, may be referred to as "output data" or "prediction data."

[0049] The data analysis unit 118 (an example of a classification unit) calculates the mean squared error between the frequency characteristic data AG1 (input data) output from the frequency characteristic bandwidth approximation unit 115 and the frequency characteristic data FG1 (output data) output from the waveform prediction unit 116 and output from the first learning model. Based on the calculated mean squared error, the first judgment error distribution GDD, and the second judgment error distribution BDD, the data analysis unit 118 determines whether the object to be inspected 1 is a good product or not (i.e., whether it is classified as a good product or not).

[0050] Here, the first judgment error distribution GDD is generated by the data analysis unit 118 based on multiple frequency characteristic data AG11 (approximate data) that are known in advance to be good products and multiple frequency characteristic data FG11 (output data) output from the first learning model, and is, for example, a histogram of the first judgment error distribution GDD shown in Figure 19.

[0051] Furthermore, the second judgment error distribution BDD is generated by the data analysis unit 118 based on multiple frequency characteristic data AG12 (approximate data) that are known in advance to be defective products and multiple frequency characteristic data FG12 (output data) output from the first learning model, and is, for example, a histogram of the second judgment error distribution BDD shown in Figure 19.

[0052] The data analysis unit 118 uses the first judgment error distribution GDD and the second judgment error distribution BDD to determine whether the object to be inspected 1 is a good product or not. However, the determination may also be made using only the first judgment error distribution GDD, or by checking whether the calculated mean squared error is within a predetermined threshold.

[0053] The data analysis unit 118 generates inspection (classification) results indicating whether the object 1 under inspection is a good product or a defective product, and outputs them to the display unit 13 for display. The data analysis unit 118 also outputs the generated histogram and the calculated mean squared error to the display unit 13 for display, allowing the user to determine whether the object is a good product or not.

[0054] Memory 12 (an example of a storage unit) has a storage device that includes semiconductor memory such as RAM (Random Access Memory) and ROM (Read Only Memory), and a storage device such as an SSD (Solid State Drive) or HDD. Memory 12 stores vibration data of good or defective products, training data, a first training model generated by the machine learning model generation unit 117, a first judgment error distribution GDD, a second judgment error distribution BDD, etc.

[0055] The display unit 13 is configured using a display such as an LCD (Liquid Crystal Display) or an organic EL (Electroluminescence). The display unit 13 displays the judgment (classification) results, histograms, and other notifications and data output from the processor 11.

[0056] The input unit 14 is a user interface that can accept user input and is configured using, for example, a mouse, keyboard, or touch panel. The input unit 14 converts the accepted user input into electrical signals (control commands) and outputs them to the processor 11.

[0057] Furthermore, the good product judgment system 100 according to Embodiment 1 may be configured to analyze vibration data acquired from each of multiple inspection targets using a single terminal device P1, and to execute machine learning processing and good product judgment processing, respectively. This allows the good product judgment system 100 to collect and learn training data used for machine learning more efficiently.

[0058] Furthermore, the terminal device P1 of the good product judgment system 100 may be implemented by, for example, a cloud server. In such a case, the cloud server acquires vibration data of at least one object to be inspected 1 via wireless communication or a network (not shown) connected via wired communication. The cloud server analyzes the acquired vibration data and performs machine learning processing and good product judgment processing, respectively. The cloud server may also transmit the good product judgment result to another terminal device managed by the worker performing the inspection of the object to be inspected 1 for display.

[0059] Next, the overall operation procedure of terminal device P1 will be described with reference to Figure 5. Figure 5 is a flowchart showing an example of the overall operation procedure of terminal device P1 in Embodiment 1.

[0060] In Figure 5, the learning model generation process (step St100), the judgment error distribution generation process (step St200), and the good product judgment process (step St300) are shown as a series of processes (flows), but each of these processes can be completed as an independent flow.

[0061] For example, the learning model generation process (step St100), the judgment error distribution generation process (step St200), and the good product judgment process (step St300) may each be executed by one or more different terminal devices P1. Furthermore, the data generated in each process (learning model, judgment error distribution) may be arbitrarily shared among multiple different terminal devices P1. Specifically, terminal device P1 may acquire data of the first learning model and the first judgment error distribution GDD that have been generated in advance from other terminal devices, external storage media, etc., and use the acquired first learning model and first judgment error distribution GDD to perform good product inspection (judgment) of the object to be inspected 1.

[0062] Terminal device P1 performs machine learning using frequency characteristic data AG11 (approximate data) based on each of the multiple vibration data collected from the inspected object 1 which is a good product, and generates a first learning model for determining whether the inspected object 1 is a good product (St100).

[0063] Terminal device P1 calculates the mean squared error between the frequency characteristic data AG11 (approximate data) and the frequency characteristic data FG11 (output data) of the good product inspection target 1, using frequency characteristic data AG11 (approximate data) based on each of the multiple vibration data collected from multiple good product inspection target 1, and the frequency characteristic data FG11 (output data) of the good product obtained (output) by inputting the frequency characteristic data AG11 (approximate data) into the first learning model. Based on the calculated mean squared error, terminal device P1 generates a first judgment error distribution GDD (see Figure 19) for the good product inspection target 1 (St200). Note that the first learning model used here may be obtained from other terminal devices, external storage media, etc.

[0064] Terminal device P1 calculates the mean squared error between the frequency characteristic data AG11 (approximate data) of the object under inspection 1 and the frequency characteristic data FG1 (output data) of the object under inspection 1 output from the first learning model. Based on the calculated mean squared error and the first judgment error distribution GDD, it determines whether the object under inspection 1 is a good product or not (i.e., whether it is classified as a good product or not) (St300). Note that the first learning model used here may be obtained from another terminal device, external storage medium, etc.

[0065] As described above, the terminal device P1 according to Embodiment 1 can perform good product determination of the object to be inspected 1. Note that the process in step St100 may be omitted if there is a first learning model that has been trained in advance, or if the first learning model that has been generated in advance by another terminal device is acquired and used. Similarly, the process in step St200 may be omitted if there is a first judgment error distribution GDD that has been generated in advance, or if the first judgment error distribution GDD that has been generated in advance by another terminal device is acquired and used.

[0066] Furthermore, if the terminal device P1 determines that the object to be inspected 1 is a good product as a result of the processing in step St300, it may further perform machine learning using the frequency characteristic data (approximate data) of the object to be inspected 1 (good product) and update the first learning model.

[0067] This allows terminal device P1 to more efficiently perform the process of determining whether object 1 is good (inspection) and updating the first learning model used to perform the good product determination of object 1.

[0068] Next, with reference to Figure 6, the learning model generation process (step St100) shown in Figure 5 will be explained. Figure 6 is a flowchart showing an example of the learning model generation process procedure for terminal device P1 in Embodiment 1. Here, as described above, learning is performed using vibration data of inspection target object 1, which has been determined to be a good product.

[0069] Terminal device P1 acquires vibration data (digital signal) of the time-domain waveform of object 1 under inspection (St101).

[0070] Terminal device P1 performs FFT processing on the acquired vibration data and converts the vibration data into frequency characteristic data FCG1 (see Figure 9) (St11).

[0071] Terminal device P1 calculates the drive frequency DF1 of the object under inspection 1 (see Figure 10) from the frequency characteristic data FCG1 by peak search processing (St12A). Terminal device P1 associates the calculated drive frequency DF1 with information about the object under inspection 1 (for example, type of object under inspection 1, product name, model number, etc.) and records it in memory 12.

[0072] Terminal device P1 divides the frequency characteristic data FCG1 into bandwidth BW1 (see Figure 15) segments and calculates the maximum value of the spectrum in each of the divided segments (St13).

[0073] Terminal device P1 determines the frequency domain to be approximated by the maximum value of each divided section (St14A).

[0074] Terminal device P1 uses the maximum value of the vibration data for each divided section to perform approximation processing on the frequency domain, which is the target of approximation processing, and converts the frequency characteristic data FCG1 into frequency characteristic data AG11 (approximation data) (St15).

[0075] The terminal device P1 repeatedly performs the processes described in steps St11 to St15 for each vibration waveform that has been identified as a good product, and records and collects the converted frequency characteristic data AG11 (approximate data) in memory 12 as training data used for machine learning to generate a learning model (St102).

[0076] Terminal device P1 uses multiple collected training data (frequency characteristic data AG11 (approximate data)) to perform machine learning (DNN) processing and generate a first training model (St103).

[0077] Based on the above, the terminal device P1 in Embodiment 1 generates a first learning model. The terminal device P1 can also generate a second learning model from vibration waveforms known to be defective, and this can be done using a similar procedure.

[0078] Next, with reference to Figure 7, the judgment error distribution generation process (step St200) shown in Figure 5 will be described. Figure 7 is a flowchart showing an example of the judgment error distribution generation process procedure for terminal device P1 in Embodiment 1. Note that the data conversion process (step St10B) shown in Figure 7 is almost the same as the data conversion process (step St10A) described in Figure 6, so only the differences will be explained.

[0079] Terminal device P1 acquires vibration data (digital signal) of the time-domain waveform of object 1 under inspection (St201).

[0080] Terminal device P1 performs FFT processing on the acquired vibration data to convert the vibration data into frequency characteristic data FCG1 (see Figure 9) (St11), and reads the drive frequency DF1 (see Figure 10) of the object to be inspected 1 recorded in memory 12 based on the information of the object to be inspected 1 (e.g., type of object to be inspected 1, product name, model number, etc.) (St12B). Also, terminal device P1 reads the frequency range that is the target of the approximation processing that approximates by the maximum value of each division interval based on the information of the object to be inspected 1 (e.g., type of object to be inspected 1, product name, model number, etc.) (St14B).

[0081] Terminal device P1 inputs the approximated frequency response data AG1 (input data) into the first learning model. Terminal device P1 acquires the frequency response data FG1 (output data, predicted data) output from the first learning model (St202).

[0082] Terminal device P1 calculates the mean squared error between the frequency response data AG1 (input data) and the frequency response data FG1 (output data) output from the first learning model (St203). Terminal device P1 records and collects the calculated mean squared error data in memory 12 (St204).

[0083] The terminal device P1 generates a first judgment error distribution GDD and a second judgment error distribution BDD (see Figure 19) for the good product inspection object 1, based on multiple mean square error data collected from the good product inspection object 1, which are recorded in the memory 12 (St205).

[0084] As described above, the terminal device P1 in Embodiment 1 can generate a first judgment error distribution GDD and a second judgment error distribution BDD that can determine whether the object to be inspected 1 is a good product or not (i.e., whether it is classified as a good product or not) based on the mean squared error.

[0085] Next, with reference to Figure 8, the good product determination process (step St300) shown in Figure 5 will be described. Figure 8 is a flowchart showing an example of the good product determination process procedure for terminal device P1 in Embodiment 1.

[0086] Note that the processes in steps St201 to St203 shown in Figure 8 are the same as the processes in steps St201 to St203 explained in Figure 7, so their explanation is omitted. Also, the data conversion process (step St10B) shown in Figure 8 is the same as the data conversion process (step St10B) explained in Figure 7, so its explanation is omitted.

[0087] Terminal device P1 determines that the mean squared error between the frequency characteristic data AG1 (input data) and frequency characteristic data FG1 (output data) calculated by the processing in step St203 is the judgment error (μA-3σ) of the first judgment error distribution GDD for good products. A ) ~ Judgment error (μA+3σ A Based on whether or not it falls within the specified range, a determination is made as to whether or not the object to be inspected 1 is a good product (i.e., whether or not it is classified as a good product) (St301), and the determination result is output to the display unit 13 (St302).

[0088] In this case, terminal device P1 determines that the mean squared error calculated by the processing in step St203 is the judgment error (μA-3σ). A ) ~ Judgment error (μA+3σA If it is determined that the value is within the specified range, then the inspected object 1 is determined to be a good product (i.e., classified as a good product). On the other hand, terminal device P1 determines that the mean squared error calculated by the processing in step St203 is the determination error (μA-3σ). A ) ~ Judgment error (μA+3σ A If it is determined that the object is outside the normal range, then the object being inspected (1) is determined not to be a good product (i.e., it is not classified as a good product).

[0089] As described above, the terminal device P1 in Embodiment 1 divides the frequency characteristic data FCG1 based on the bandwidth of the drive frequency of the object to be inspected 1, and approximates the frequency characteristic data FCG1 based on the maximum value of the spectrum of the divided section, thereby enabling the inspection object 1 to be judged as good (classified) regardless of individual differences in the drive source such as the motor or actuator equipped in the object to be inspected 1.

[0090] Furthermore, this allows terminal device P1 to suppress misjudgments in the individual differences between inspected objects 1 (classification judgment) and improve the accuracy of good product judgment.

[0091] Next, the frequency response data FCG1 of the object under inspection 1 will be described with reference to Figures 9 and 10, respectively. Figure 9 is a diagram showing an example of the frequency response data FCG1. Figure 10 is a magnified view of a portion of the frequency response data FCG1.

[0092] The frequency characteristic data FCG1 is generated by the frequency characteristic conversion unit 112 and is a graph showing the frequency characteristics of the drive source of the object under inspection 1, such as a motor or actuator, obtained by performing FFT processing on the vibration data (digital signal) of the time-domain waveform of the object under inspection 1.

[0093] The drive source of the object under inspection 1 is driven at a predetermined period. As a result, the frequency characteristic data FCG1 shows a complex frequency characteristic in which harmonic components of the drive frequency are superimposed at approximately equal intervals on the frequency characteristics inherent to the drive source, and each harmonic component forms a peak in the spectrum. Furthermore, the superposition of harmonic components is formed over a high-frequency band.

[0094] The frequency response data FCG1 shown in Figure 10 is a partially enlarged view of the frequency response data FCG1 shown in Figure 9. The drive frequency calculation unit 113 detects the drive frequency DF1 = 180 Hz of the object under inspection 1 from the frequency response data FCG1 by peak search processing.

[0095] Next, we will explain the individual differences in the frequency characteristic data of good quality inspected object 1 by referring to Figures 11 to 13. Figure 11 shows a comparison of the frequency characteristic data FCG11 and FCG12 of two good quality objects. Figure 12 shows a comparison of the low-frequency region EX1 of the frequency characteristic data FCG11 and FCG12 of two good quality objects. Figure 13 shows a comparison of the mid-frequency region EX2 of the frequency characteristic data FCG11 and FCG12 of two good quality objects.

[0096] The frequency response data FCG11 and FCG12 are graphs of frequency response data obtained by processing vibration data acquired from two inspected objects 1 that were judged to be good products using FFT. In the low-frequency region EX1, the peaks of harmonic components that occur at approximately equal multiples of the drive frequency DF1 are almost identical in the frequency response data FCG11 and FCG12, but there is a difference in the frequencies at which the peaks of harmonic components are formed in the mid-frequency region EX2.

[0097] Furthermore, with reference to Figure 14, the frequency characteristic data of a good product and a defective product will be explained. Figure 14 shows comparative examples of the frequency characteristic data of good and defective products.

[0098] Frequency response data FCG21 shows the frequency response data for inspection object 1, which is a good product. Frequency response data FCG22 shows the frequency response data for inspection object 1, which is a defective product.

[0099] The frequency response data FCG21 and FCG22 show that the peaks of harmonic components occurring at approximately equal multiples of the drive frequency DF1 in the low-frequency region are nearly identical. However, in the mid-frequency and high-frequency regions (region AR1 shown in Figure 14), there are significant differences in the frequencies at which the peaks of harmonic components are formed.

[0100] Therefore, in the machine learning process that generates a learning model (first learning model or second learning model) for determining whether a product is good or bad, when comparing the magnitude of the spectrum for each frequency, the difference in the frequencies at which the peaks of harmonic components are formed between good product inspection targets 1 cancels out the differences in the spectrum in the mid-frequency and high-frequency regions between good product inspection target 1 and defective product inspection target 1, potentially reducing the accuracy of the good product determination.

[0101] Therefore, as described above, the terminal device P1 in this embodiment 1 divides the frequency characteristic data by bandwidth and approximates the frequency characteristic data with the maximum value of the spectrum in each divided section, thereby generating training data that ignores the shift in the peak position of harmonic components. Hereinafter, specific examples of approximation of frequency characteristic data will be described.

[0102] Next, the method for calculating the maximum value of the spectrum in each division interval will be explained with reference to Figures 15 and 16, respectively. Figure 15 is a diagram illustrating an example of calculating the maximum value for each division interval in the frequency characteristic data FCG1. Figure 16 is a diagram showing a comparative example of the frequency characteristic data FCG1 before and after approximation processing.

[0103] The frequency response data FCG1 shown in Figure 15 illustrates an example where the drive frequency DF1 = 180 Hz, but it goes without saying that the drive frequency is just one example and is not limited to this.

[0104] The bandwidth BW1 is determined based on the drive frequency DF1. The bandwidth BW1 is set to a value greater than the drive frequency DF1 = 180 Hz and less than the drive frequency DF1 × 2 = 360 Hz. In this embodiment 1, an example is shown where the bandwidth BW1 is set to a sampling frequency of 44100 Hz ÷ sampling points of 4096 × number of samples of 32 = 345 Hz.

[0105] As a result, terminal device P1 ensures that each of the divided sections SC1 to SC6, separated by the bandwidth BW1, contains a maximum of two peaks of harmonic components of the drive frequency. For example, in the example shown in Figure 15, divided section SC1 contains one peak. Divided sections SC2 to SC6 each contain two peaks. Note that each divided section only needs to contain at least one peak.

[0106] Terminal device P1 calculates the peaks Pk1, Pk2, Pk3, Pk4, Pk5, and Pk6, which are the maximum values ​​of the spectrum in each divided section, and then approximates the corresponding divided section based on the calculated peaks Pk1 to Pk6, thereby converting it into frequency response data AG1. As a result, frequency response data FCG1 is converted into frequency response data AG1, in which the spectrum is shown in a stepped manner.

[0107] Next, the frequency response data (output data) produced by the first learning model will be explained with reference to Figures 17 and 18, respectively. Figure 17 shows a comparative example of input data for a good product (frequency response data AG11) and predicted data by the first learning model (frequency response data FG11). Figure 18 shows a comparative example of input data for a defective product (frequency response data AG12) and predicted data by the first learning model (frequency response data FG12). Here, the inspected object 1 is described as a good product.

[0108] In FIG. 17, the frequency characteristic data AG11 is approximate data of the inspection object 1 that is a non-defective product. The frequency characteristic data FG11 is output data (predicted data) that is obtained by inputting the frequency characteristic data AG11 as input data into the first learning model and outputting from the first learning model. In FIG. 18, the frequency characteristic data AG12 is approximate data of the inspection object 1 that is a defective product. The frequency characteristic data FG12 is output data (predicted data) that is obtained by inputting the frequency characteristic data AG12 as input data into the first learning model and outputting from the first learning model.

[0109] The spectrum of the frequency characteristic data AG11 of the non-defective product and the frequency characteristic data FG11 output by the first learning model generated using the frequency characteristic data (approximate data) of the non-defective product matches more with the spectrum of the frequency characteristic data AG12 of the defective product and the frequency characteristic data FG12 output by the first learning model generated using the frequency characteristic data (approximate data) of the non-defective product.

[0110] Thereby, the terminal device P1 can determine whether the inspection object 1 is a non-defective product (that is, whether it is classified as a non-defective product) by evaluating the mean square error in the frequency band between the frequency characteristic data AG1 as input data and the predicted data (output data) as the determination error.

[0111] Next, referring to FIG. 19, the mean square error will be described. FIG. 1 is a diagram showing an example of the first determination error distribution GDD and the second determination error distribution BDD.

[0112] The first determination error distribution GDD shows the normal distribution of the mean square error between the frequency characteristic data AG11 (input data) of the non-defective product and the frequency characteristic data FG11 (output data) output from the first learning model. The determination error range MSEA is centered on the average value μA of the determination error, from the determination error (μA - 3σ A ) to the determination error (μA + 3σ A ) range. Here, σ A is the standard deviation σ A of the first determination error distribution GDD.

[0113] The second judgment error distribution BDD shows the normal distribution of the mean squared error between the frequency characteristic data AG12 (input data) of defective products and the frequency characteristic data FG12 (output data) output from the first learning model. The judgment error range MSEB is centered on the mean value μB of the judgment error, and represents the judgment error (μB-3σ). B ) from the judgment error (μB + 3σ B This indicates the range up to ). Note that σ here refers to σ B This is the standard deviation σ of the second decision error distribution BDD. B That is the case.

[0114] The judgment performance α is the judgment error (μA + 3σ) of the first judgment error distribution GDD. A ) and the judgment error of the second judgment error distribution BDD (μB-3σ B This shows the difference from ).

[0115] Next, the judgment performance α will be explained with reference to Figures 20 and 21, respectively. Figure 20 is a diagram showing an example of the correlation between judgment performance and bandwidth. Figure 21 is a correlation graph DPG showing an example of the correlation between judgment performance and bandwidth. In Figures 20 and 21, the correlation between judgment performance and bandwidth is explained as an example when the drive frequency DF1 = 180 Hz.

[0116] First, we will explain the judgment performance α1 when the bandwidth BW1 is not set in case (i).

[0117] The first judgment error distribution (not shown) is centered around the mean value of the judgment error μA1, and the judgment error (μA-3σ) A ) from the judgment error (μA + 3σ A The range up to ) is the judgment error range MSEA1. Similarly, the second judgment error distribution (not shown) is centered on the mean value of the judgment error μB1, and the judgment error (μB-3σ) B ) from the judgment error (μB + 3σ B The range up to ) is the judgment error range MSEB1. In such cases, the standard deviation of the first judgment error distribution that determines a product to be good is σ AThe standard deviation σ of the second judgment error distribution that determines whether a product is defective. B As these two values ​​increase, the judgment error range MSEA1 of the first judgment error distribution and the judgment error range MSEB1 of the second judgment error distribution overlap.

[0118] Therefore, if the bandwidth BW1 is not set, the judgment performance α1 < 0, and a misjudgment occurs in this overlapping region. As a result, the terminal device P1 cannot determine whether the object to be inspected 1 is a good product or a defective product based on the mean squared error.

[0119] Next, we will explain the determination performance α2 when the bandwidth is set such that the drive frequency > bandwidth in case (ii).

[0120] The first judgment error distribution (not shown) is centered around the mean value of the judgment error μA², with the judgment error (μA-3σ) as the center. A ) from the judgment error (μA + 3σ A The range up to ) is the judgment error range MSEA2. Similarly, the second judgment error distribution (not shown) is centered on the mean value of the judgment error μB2, and the judgment error (μB-3σ) B ) from the judgment error (μB + 3σ B The range up to ) becomes the judgment error range MSEB2. In such cases, the standard deviation of the first judgment error distribution that determines a product to be good is σ A The standard deviation σ of the second judgment error distribution that determines whether a product is defective. B As these values ​​decrease, the judgment error range MSEA2 of the first judgment error distribution and the judgment error range MSEB2 of the second judgment error distribution no longer overlap.

[0121] Therefore, when a bandwidth is set such that the drive frequency DF1 (=180Hz) > bandwidth, the judgment performance α2 > 0, and a difference is created between the judgment error range MSEA2 and the judgment error range MSEB2. As a result, the terminal device P1 can determine whether the object to be inspected 1 is a good product or a defective product based on the mean squared error.

[0122] Next, we will explain the judgment performance α3 when the bandwidth is set such that the drive frequency DF1 (=180Hz) < bandwidth < drive frequency DF1 × 2 (=360Hz) in case (iii).

[0123] The first judgment error distribution (not shown) is centered around the mean value of the judgment error μA3, with the judgment error (μA-3σ) as the center. A ) from the judgment error (μA + 3σ A The range up to ) is the judgment error range MSEA3. Similarly, the second judgment error distribution (not shown) is centered around the mean judgment error μB3, and the judgment error (μB-3σ) B ) from the judgment error (μB + 3σ B The range up to ) becomes the judgment error range MSEB3. In such cases, the standard deviation of the first judgment error distribution that determines a product to be good is σ A The standard deviation σ of the second judgment error distribution that determines whether a product is defective. B As these values ​​become even smaller, the gap between the judgment error range MSEA3 of the first judgment error distribution and the judgment error range MSEB3 of the second judgment error distribution widens further.

[0124] Therefore, when a bandwidth is set such that drive frequency DF1 (=180Hz) < bandwidth < drive frequency DF1 × 2 (=360Hz), judgment performance α3 > judgment performance α2, and a difference is created between the judgment error range MSEA3 and the judgment error range MSEB3. As a result, terminal device P1 can determine with higher accuracy whether the object to be inspected 1 is a good product or a defective product based on the mean squared error.

[0125] Next, we will explain the judgment performance α4 when the bandwidth is set such that the drive frequency DF1 × 2 (= 360 Hz) < bandwidth in case (iv).

[0126] The first judgment error distribution (not shown) is centered around the mean value of the judgment error μA4, with the judgment error (μA-3σ) as the center. A ) from the judgment error (μA + 3σ AThe range up to ) is the judgment error range MSEA4. Similarly, the second judgment error distribution (not shown) is centered on the mean value of the judgment error μB4, and the judgment error (μB-3σ) B ) from the judgment error (μB + 3σ B The range up to ) is the judgment error range MSEB4. In such cases, the standard deviation of the first judgment error distribution that determines a product to be good is σ A The standard deviation σ of the second judgment error distribution that determines whether a product is defective. B As both become even smaller, the gap between the mean value μA4 and the mean value μB4 becomes smaller, thus reducing the difference between the judgment error range MSEA4 of the first judgment error distribution and the judgment error range MSEB4 of the second judgment error distribution.

[0127] Therefore, when a bandwidth is set such that the drive frequency DF1 × 2 (= 360 Hz) < bandwidth, the judgment performance α3 > judgment performance α4. However, a difference arises between the judgment error range MSEA4 and the judgment error range MSEB4, so the terminal device P1 can determine whether the object to be inspected 1 is a good product or a defective product based on the mean squared error.

[0128] The correlation graph DPG shows the changes in judgment performance α1 to α4 for each case (i) to case (iv) based on bandwidth. As described above, when judgment performance α < 0, terminal device P1 is unable to determine good products based on the mean squared error between the input data (approximate data), which is the frequency characteristic data, and the output data (predicted data), which is the frequency characteristic data. When judgment performance α ≥ 0, good product determination based on the mean squared error becomes possible. Judgment performance α is maximized when the bandwidth is set such that BW1 (= 180 Hz) < bandwidth < drive frequency DF1 × 2 (= 360 Hz), and it can be seen that the judgment accuracy is best improved in good product determination (classification determination) based on the mean squared error.

[0129] As described above, the terminal device P1 (an example of a classification device) in Embodiment 1 is a device for determining whether an object to be inspected 1 (an example of an object) having a periodically driven drive source is a good product, and comprises a communication unit 10 (an example of an acquisition unit) that acquires vibration data (an example of time-axis waveform data) of the object to be inspected 1 (motor, actuator, etc.), a frequency characteristic conversion unit 112 (an example of a conversion unit) that converts the vibration data into frequency characteristic data FCG1 (an example of first frequency characteristic data), and a section maximum value calculation unit 114 (an example of a spectrum calculation unit) that divides the frequency characteristic data FCG1 into predetermined bandwidths (for example, 180 Hz) and calculates the maximum value of the spectrum for each divided section, The system includes a frequency response bandwidth approximation unit 115 (an example of an approximation processing unit) that approximates frequency response data FCG1 based on the maximum value of the spectrum for each divided interval and outputs frequency response data AG1 (input data, approximate data) (an example of a second frequency response data); a machine learning model generation unit 117 (an example of a learning model generation unit) that generates frequency response data FG1 (output data, predicted data) (an example of a third frequency response data) from frequency response data AG1 (input data, approximate data) using a learning model; and a data analysis unit 118 (an example of a classification unit) that classifies the object to be inspected 1 based on frequency response data AG1 and frequency response data FG1. The learning model is trained on learning data approximated by at least the frequency response bandwidth approximation unit 115.

[0130] As a result, the terminal device P1 in Embodiment 1 can further improve the accuracy of determining whether an item is good, regardless of individual differences in the frequency characteristic data acquired from the object to be inspected 1.

[0131] Furthermore, the terminal device P1 in Embodiment 1 further includes a drive frequency calculation unit 113 that calculates the drive frequency DF1 of the drive source based on the frequency characteristic data FCG1. As a result, the terminal device P1 in Embodiment 1 can calculate the drive frequency DF1 used for setting the bandwidth BW1.

[0132] Furthermore, the terminal device P1 in Embodiment 1 further includes a memory 12 (an example of a storage unit) that stores the drive frequency DF1 of the drive source. As a result, the terminal device P1 in Embodiment 1 can automatically set the bandwidth BW1 by using the drive frequency of the drive source of the object to be inspected 1 recorded in the memory 12.

[0133] Furthermore, the predetermined bandwidth BW1 in the terminal device P1 in Embodiment 1 is greater than the driving frequency DF1. As a result, the terminal device P1 in Embodiment 1 includes at least one peak Pk1 to Pk6 due to high-frequency components in each divided section, and the frequency characteristic bandwidth approximation unit 115 can approximate each divided section of the frequency characteristic data FCG1 with the maximum value of the spectrum of each divided section.

[0134] Furthermore, the predetermined bandwidth BW1 in the terminal device P1 in Embodiment 1 is less than twice the driving frequency. As a result, the terminal device P1 in Embodiment 1 can suppress the increase in the number of peaks caused by high-frequency components in each divided section, thereby suppressing the averaging of the spectra in each divided section during the approximation processing by the frequency characteristic bandwidth approximation unit 115.

[0135] Furthermore, in the first embodiment, the frequency characteristic conversion unit 112 in the terminal device P1 determines the sampling frequency and the number of sampling points in the conversion of vibration data. A predetermined bandwidth BW1 is determined based on the sampling frequency ÷ number of sampling points × N, where N is an integer of 1 or more. As a result, the terminal device P1 in the first embodiment can set a bandwidth BW1 that is a frequency band of one or more cycles of a periodically driven drive source. Here, one peak occurs for each cycle of the drive operation. Therefore, when the bandwidth BW1 is set to a frequency band of one or more cycles of the drive source, the terminal device P1 can reduce the discrepancy between the peak occurrence period (i.e., the drive period of the drive source) and the bandwidth BW1 used to divide the frequency characteristic data FCG1 in the approximation process.

[0136] Furthermore, the section maximum value calculation unit 114 in the terminal device P1 of Embodiment 1 extracts a predetermined frequency range from the frequency characteristic data FCG1 and divides the extracted frequency characteristic data FCG1 into a predetermined bandwidth BW1. As a result, the terminal device P1 of Embodiment 1 can omit approximation processing of unnecessary frequency ranges, thereby reducing the overhead required for approximation processing and shortening the processing time required for approximation processing.

[0137] Furthermore, the data analysis unit 118 in the terminal device P1 in Embodiment 1 classifies the object to be inspected 1 based on the mean square error between the frequency characteristic data AG1 and the frequency characteristic data FG1. As a result, the terminal device P1 in Embodiment 1 can limit the scope of calculation for the mean square error in determining whether an object is good or bad, thereby reducing the processing load in determining whether an object is good or bad, and shortening the processing time required for determining whether an object is good or bad.

[0138] As described above, the terminal device P1 (an example of a learning model generation device) in Embodiment 1 is a device that performs learning on an object to be inspected 1 (object) having a periodically driven drive source (motor, actuator, etc.), and comprises: a communication unit 10 (an example of an acquisition unit) that acquires time-axis waveform data (an example of time-axis waveform data) of the object to be inspected 1; a frequency characteristic conversion unit 112 (an example of a conversion unit) that converts the time-axis waveform data into frequency characteristic data FCG1 (an example of a first frequency characteristic data); a section maximum value calculation unit 114 (an example of a spectrum calculation unit) that divides the frequency characteristic data FCG1 into predetermined bandwidths (for example, 180 Hz) and calculates the maximum value of the spectrum for each division section; a frequency characteristic bandwidth approximation unit 115 (an example of an approximation processing unit) that approximates the frequency characteristic data FCG1 based on the maximum value of the spectrum for each division section and outputs frequency characteristic data AG1 (input data, approximation data) (an example of a second frequency characteristic data); and a machine learning model generation unit 117 (an example of a learning model generation unit) that trains a learning model with multiple frequency characteristic data AG1.

[0139] As a result, the terminal device P1 in Embodiment 1 can generate a learning model capable of classifying the object to be inspected 1 regardless of individual differences in frequency characteristic data acquired from the object to be inspected 1.

[0140] Although various embodiments have been described above with reference to the drawings, it goes without saying that this disclosure is not limited to such examples. It is clear to those skilled in the art that various modifications, alterations, substitutions, additions, deletions, and equivalents can be conceived within the scope of the claims, and these are also understood to fall within the technical scope of this disclosure. Furthermore, the components of the various embodiments described above can be combined arbitrarily without departing from the spirit of the invention. [Industrial applicability]

[0141] This disclosure is useful as a classification device, a learning model generation device, a classification method, and a learning model generation method that can perform good quality determination of inspected objects regardless of individual differences in the inspected objects. [Explanation of Symbols]

[0142] 1. Objects to be inspected 2. Vibration Sensor 3 Analog Filters 4 A / D converters 10 Communications Department 11 processors 12 memory 13 Display section 111 Vibration waveform measurement unit 112 Frequency Response Conversion Section 113 Drive frequency calculation unit 114 Section Maximum Value Calculation Unit 115 Frequency response bandwidth approximation section 116 Waveform prediction unit 117 Machine Learning Model Generation Unit 118 Data Analysis Department Frequency response data for AG1, FCG1, FCG11, FCG12, FCG21, FCG22, FG1, FG2 SC1, SC2, SC3, SC4, SC5, SC6 division section

Claims

1. A classification device for classifying objects having a periodically driven drive source, An acquisition unit that acquires time-axis waveform data of the aforementioned object, A conversion unit that converts the time-domain waveform data into first frequency characteristic data, A spectrum calculation unit divides the first frequency characteristic data into a plurality of division sections with a predetermined bandwidth and calculates the maximum value of the spectrum for each division section. An approximation processing unit that approximates the first frequency characteristic data and outputs second frequency characteristic data based on the maximum value of the spectrum for each of the division sections, A generation unit that generates a third frequency characteristic data from the second frequency characteristic data using a learning model, The system includes a classification unit that classifies the objects based on the second frequency characteristic data and the third frequency characteristic data, The learning model is a model that has learned training data approximated by at least the approximation processing unit. Classification device.

2. The system further includes a drive frequency calculation unit that calculates the drive frequency of the drive source based on the first frequency characteristic data, The classification device according to claim 1.

3. The system further includes a storage unit for storing the drive frequency of the drive source, The classification device according to claim 1.

4. The predetermined bandwidth is greater than the drive frequency. The classification device according to claim 2 or 3.

5. The predetermined bandwidth is less than twice the drive frequency. The classification device according to claim 2 or 3.

6. The conversion unit determines the sampling frequency and the number of sampling points in the conversion of the time-domain waveform data. The predetermined bandwidth is determined based on the sampling frequency ÷ the number of sampling points × N. Here, N is an integer greater than or equal to 1. The classification device according to claim 1.

7. The spectrum calculation unit extracts a predetermined frequency range from the first frequency characteristic data and divides the extracted predetermined frequency range into predetermined bandwidths. The classification device according to claim 1.

8. The classification unit classifies the objects based on the mean square error between the second frequency characteristic data and the third frequency characteristic data. The classification device according to claim 1.

9. A classification method performed by a classification device that classifies objects having a periodically driven drive source, The time-domain waveform data of the aforementioned object is acquired, The aforementioned time-domain waveform data is converted into first frequency characteristic data. The first frequency characteristic data is divided into a plurality of division intervals with a predetermined bandwidth, and the maximum value of the spectrum is calculated for each division interval. Based on the maximum value of the spectrum for each of the division sections, a second frequency characteristic data is output that approximates the first frequency characteristic data. Using the learning model, a third frequency characteristic data is generated from the second frequency characteristic data. Based on the second frequency characteristic data and the third frequency characteristic data, the object is classified, The learning model is a model that has learned training data approximated by at least the classification device. Classification method.

10. Based on the first frequency characteristic data, the drive frequency of the drive source is calculated. The classification method according to claim 9.

11. The drive frequency of the aforementioned drive source is stored. The classification method according to claim 9.

12. The predetermined bandwidth is greater than the drive frequency. The classification method according to claim 10 or 11.

13. The predetermined bandwidth is less than twice the drive frequency. The classification method according to claim 10 or 11.

14. In the conversion of the time-domain waveform data, the sampling frequency and the number of sampling points are determined. The predetermined bandwidth is determined based on the sampling frequency ÷ the number of sampling points × N. Here, N is an integer greater than or equal to 1. The classification method according to claim 9.

15. A predetermined frequency range is extracted from the first frequency characteristic data, The extracted predetermined frequency range is divided into predetermined bandwidths. The classification method according to claim 9.

16. The objects are classified based on the mean squared error between the second frequency characteristic data and the third frequency characteristic data. The classification method according to claim 9.

Images