Time series data processing method, learning method, processing system and learning system

Abstract

To provide a time series data processing method or the like capable of improving the speed of learning or inference and improving inference accuracy in time series data analysis using machine learning.SOLUTION: An encoder SB2a is a neural network for analyzing time series data and having no recursive structure, and uses an attention mechanism. A learned model 130 learns so as to output a feature amount SB2b from the encoder SB2a on the basis of the time series data. A time series data processing method includes a step for acquiring a plurality of time series data DT1, DT2 in a plurality of windows WD1, WD2, and performing phase adjustment processing for keeping an initial phase of the time series data in respective start positions TS1, TS2 of each window constant. The time series data processing method includes a step for acquiring the feature amount SB2b from the encoder SB2a by inputting the time series data subjected to the phase adjustment processing to the learned model 130.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] The present invention relates to a time-series data processing method, a learning method, a processing system, a learning system, and the like.

Background Art

[0002] In Non-Patent Document 1, a Transformer has been proposed as an architecture of a neural network. The Transformer has an encoder-decoder structure, encodes input data to extract feature amounts, and performs decoding processing based on the feature amounts to generate an output. Each of the encoder and the decoder captures the relationship between elements by parallel processing using an attention mechanism, and can shorten the learning or inference time as compared with a recurrent neural network or the like that sequentially processes input data.

[0003] The Transformer as described above can be used for analysis of time-series data. The feature amounts output by the encoder indicate the features of the time-series data input to the Transformer, and by using the feature amounts, the time-series data can be analyzed. For example, when the time-series data is vibration data, it is possible to perform abnormality detection or the like of the object on which the vibration data is measured.

Prior Art Documents

Non-Patent Documents

[0004]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, since the phase of the frequency components included in the time-series data changes according to the start time of the time-series data, when attempting to extract features by parallel processing using an attention mechanism, the features learned or inferred by the phase at the start time change, making it difficult to analyze based on those features.

[0006] Also, not limited to transformers, when extracting features by parallel processing using an attention mechanism, there were the same problems as above.

Means for Solving the Problems

[0007] One aspect of the present disclosure relates to a time-series data processing method that analyzes time-series data and uses a neural network without a recursive structure, includes an encoder using an attention mechanism, and uses a trained model trained to output the features from the encoder based on the time-series data. The method includes steps of obtaining a plurality of time-series data in a plurality of windows and performing a phase adjustment process to make the initial phase of the time-series data at the start position of each window constant, and inputting the time-series data after the phase adjustment process into the trained model to obtain the features from the encoder.

[0008] Another aspect of the present disclosure relates to a learning method that analyzes time-series data and uses a neural network without a recursive structure, includes an encoder and a decoder using an attention mechanism, trains a learning model to infer time-series data outside the time range of the time-series data to be analyzed, obtains a plurality of time-series data in a plurality of windows, performs a phase adjustment process to make the initial phase of the time-series data at the start position of each window constant, and inputs the time-series data after the phase adjustment process into the learning model and performs the training so that the learning model infers the time-series data outside the time range.

[0009] Further, still another aspect of the present disclosure is a neural network that analyzes time-series data and does not have a recursive structure, includes an encoder using an attention mechanism, and stores a learned model trained to output the feature amount from the encoder based on the time-series data, and a processing unit that performs processing using the learned model, wherein the processing unit acquires a plurality of time-series data in a plurality of windows, performs a phase adjustment process for making the initial phase of the time-series data at the start position of each window constant, and acquires the feature amount from the encoder by inputting the time-series data after the phase adjustment process to the learned model, and relates to a processing system.

[0010] Further, still another aspect of the present disclosure is a neural network that analyzes time-series data and does not have a recursive structure, includes an encoder and a decoder using an attention mechanism, and stores a storage unit that stores a learning model for inferring time-series data outside the time range of the time-series data to be analyzed, and a processing unit that performs learning of the learning model, wherein the processing unit acquires a plurality of time-series data in a plurality of windows, performs a phase adjustment process for making the initial phase of the time-series data at the start position of each window constant, inputs the time-series data after the phase adjustment process to the learning model, and relates to a learning system that performs the learning so that the learning model infers the time-series data outside the time range.

Brief Description of Drawings

[0011]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Mode for Carrying Out the Invention

[0012] Hereinafter, preferred embodiments of the present disclosure will be described in detail. It should be noted that the embodiments described below do not unduly limit the content described in the claims, and not all of the configurations described in the embodiments are essential constituent elements.

[0013] 1. Processing System and Learning System Hereinafter, an example in which the time series data to be analyzed is vibration data measured by a vibration sensor will be described. Also, hereinafter, an example in which a transformer is used as a model using an attention mechanism will be described. However, as will be described later, the application targets of the method of the present embodiment are not limited to these.

[0014] FIG. 1 is an explanatory diagram of a sensor for detecting the vibration of an object. The object 10 includes, for example, a vibration source 11 that generates vibration by mechanical operation. The sensor 140 detects the vibration of the object 10 generated by the vibration source 11. The vibration source 11 is, as an example, a motor, an engine, a turbine, or the like. Further, the object 10 may not include the vibration source 11, vibration may be applied from the outside of the object 10, and the sensor 140 may detect the vibration.

[0015] The object 10 is, for example, the vibration source 11 itself, that is, a motor, an engine, a turbine, or the like. Alternatively, the object 10 is a machine, device, or apparatus including the vibration source 11, such as a home appliance or industrial equipment like a printer, an air conditioner, a robot, a pump, a belt conveyor, or a processing device, a moving body like an automobile or an airplane, or an industrial facility like a generator or a manufacturing plant. Alternatively, the object 10 may be a structure that vibrates due to an external force, such as a building, a road, or a bridge.

[0016] The sensor 140 detects acceleration, velocity, displacement, angular acceleration, angular velocity, or angle, and outputs a signal indicating the detected physical quantity as vibration information. The sensor 140 may be a sensor that detects one type of physical quantity, or a sensor that detects a plurality of types of physical quantities. Further, the sensor 140 may be a sensor that detects a physical quantity on one axis, or a sensor that detects a physical quantity on two or more axes. The sensor 140 outputs vibration information from one or more channels. One channel means a channel that outputs a signal of one type of physical quantity on one axis.

[0017] The sensor 140 is an acceleration sensor or a gyro sensor using a crystal oscillator as a detection element, or an acceleration sensor or a gyro sensor using MEMS as a detection element, etc. Further, the sensor 140 may be an IMU in which an acceleration sensor and a gyro sensor are combined and unitized. The sensor 140 may detect speed or displacement by integrating the acceleration detected by the detection element, or may use a detection element for detecting speed or the like. The sensor 140 may detect angular acceleration or angle by differentiating or integrating the angular velocity detected by the detection element, or may use a detection element for detecting angular acceleration or the like. An example of an acceleration sensor is a sensor that utilizes the change in the vibration frequency according to the stress applied to the crystal oscillator and detects the acceleration by measuring the vibration frequency. An example of a gyro sensor is a sensor that detects the angular velocity by detecting the Coriolis force applied to the crystal oscillator. Another example of an acceleration sensor or a gyro sensor is a sensor that detects acceleration or angular velocity by detecting the capacitance between electrodes that changes according to the inertial force applied to the mass part, which is constituted by MEMS and electrodes.

[0018] Although the sensor 140 is assumed to be attached so as to contact the object 10, it is not limited thereto, and it is sufficient that vibration is transmitted from the object 10 to the sensor 140. One sensor unit may be attached to the object 10, or a plurality of sensor units may be attached to the object 10. The plurality of sensor units may be attached to different positions of the object 10, or may be attached to the same position of the object 10.

[0019] FIG. 2 is a configuration example of the processing system. The processing system 100 infers the state of the object by analyzing vibration data, which is time-series data. Detailed examples of the state will be described later. The processing system 100 includes a processing unit 110, a storage unit 120, a sensor 140, and a presentation unit 150. Note that the sensor 140 may be provided outside the processing system 100 and connected to the processing system 100 via a cable, a network, or the like.

[0020] The processing unit 110 includes an acquisition unit 111, a calculation unit 112, and an inference unit 113.

[0021] The acquisition unit 111 acquires vibration data of the object from the sensor 140 by receiving the signal of the physical quantity output by the sensor 140. The sensor 140 may output either an analog signal or digital data. When receiving an analog signal, the acquisition unit 111 may include an A / D converter that performs A / D conversion of the analog signal into digital data. The acquisition unit 111 outputs the vibration data, which is the vibration information of the object, to the calculation unit 112. The vibration data is time-series data of acceleration, velocity, displacement, angular acceleration, angular velocity, or angle detected by the sensor 140.

[0022] Before inputting the vibration data into the learned model 130, the calculation unit 112 performs preprocessing on the vibration data. Although details will be described later, the acquisition unit 111 acquires a plurality of vibration data with different start times. The phases at the start times of this plurality of vibration data are different from each other. The phase at the start time of the vibration data is referred to as the initial phase. The calculation unit 112 performs phase adjustment processing on each vibration data so that the initial phases of the plurality of vibration data are aligned. Specifically, the calculation unit 112 performs phase adjustment processing on the components of one or more target frequencies included in each vibration data so that the initial phases of the components are aligned.

[0023] The inference unit 113 receives the time-series data after the phase adjustment process. The inference unit 113 uses the learned model 130 to infer the vibration state from the time-series data after the phase adjustment process. The vibration state is a state related to at least one of the state monitoring of the object, quality control, and predictive maintenance. Details of these states will be described later. The inference unit 113 inputs the time-series data after the phase adjustment process to the learned model 130 and causes the learned model 130 to output an inference result. As will be described later, the learned model 130 includes a transformer encoder and a vibration analysis unit that infers the vibration state based on the feature amounts output by the encoder. The vibration analysis unit is, for example, a classifier that classifies the vibration state of the object into a plurality of vibration states, or a detector that detects an abnormality or failure of the object or the like. The inference result of the classifier is, for example, the probability of each vibration state, or a flag indicating which vibration state it is. The inference result of the detector is a flag or the like indicating whether an abnormality or failure has been detected.

[0024] The processing unit 110 outputs presentation information based on the inference result to the presentation unit 150. The presentation unit 150 presents the presentation information to the user. The presentation unit 150 is, for example, a display, a speaker, a lamp, or a vibrator. The presentation information is a numerical value, a character, a color, an image, a sound, a light, or a vibration. Note that the processing unit 110 may store the classification result of the vibration state, the detection result of the vibration state, or the presentation information generated therefrom in a memory or a storage. The memory or the storage may be common to the storage unit 120 described below.

[0025] The storage unit 120 stores the learned model 130. The learned model 130 is generated by the learning system training the learning model in advance. The generated learned model 130 is stored in the storage unit 120. As the learning method, various methods such as supervised learning or unsupervised learning may be adopted, and may be appropriately selected according to the architecture of the model or the content to be inferred by the model.

[0026] As the hardware configuration of the processing system 100, various configurations may be adopted. The processing system 100 is, for example, a computer, or a cloud system in which a plurality of computers are connected by a network or the like. The computer is not limited to a general-purpose one such as a personal computer, and may be a dedicated one for performing the vibration analysis of the present embodiment, or one incorporated in a specific device or the like.

[0027] The processing unit 110 is, as an example, a processor. The processor includes, for example, one or more of a CPU, a GPU, a microcomputer, a DSP, an ASIC, or an FPGA. The CPU is an abbreviation for Central Processing Unit. The GPU is an abbreviation for Graphics Processing Unit. The DSP is an abbreviation for Digital Signal Processor. The ASIC is an abbreviation for Application Specific Integrated Circuit. The FPGA is an abbreviation for Field Programmable Gate Array. The storage unit 120 stores a program in which the functions of each part of the processing unit 110 are described. The processor realizes the functions of each part of the processing unit 110 as processing by executing the program.

[0028] The processing unit 110 is not limited to the software processing as described above, and may be a circuit in which the functions of each part are implemented in hardware. In that case, the storage unit 120 may not store a program.

[0029] The storage unit 120 is a memory or a register. The memory is a volatile memory such as a RAM, or a non-volatile memory such as an OTP memory or an EEPROM. The RAM is an abbreviation for Random Access Memory. The OTP is an abbreviation for One Time Programmable. The EEPROM is an abbreviation for Electrically Erasable Programmable Read Only Memory.

[0030] A non-transitory information storage medium, which is a computer-readable medium, may store the learned model 130 and the above program. The information storage medium is, for example, an optical disk, a memory card, a hard disk drive, or a non-volatile semiconductor memory.

[0031] Figure 3 is a configuration example of a learning system. The learning system 200 generates a learned model 130 by training a learning model 230. The learning system 200 includes a processing unit 210, a storage unit 220, an operation unit 260, and a display unit 270.

[0032] The storage unit 220 stores the learning model 230 and the learning data 225. The architectures of the learning model 230 and the learned model 130 are basically the same. However, the learned model 130 may be obtained by omitting the parts of the components of the learning model 230 that are not used in the inference stage. The learning data 225 is data used for training the learning model 230. The learning data 225 is a large number of vibration data in unsupervised learning, and is a large number of vibration data and correct labels corresponding to each vibration data in supervised learning.

[0033] The processing unit 210 includes an acquisition unit 211, an arithmetic unit 212, and a learning processing unit 214. The acquisition unit 211 reads the learning data 225 from the storage unit 220. The arithmetic unit 212 performs the above-described phase adjustment processing on the vibration data included in the learning data 225 as preprocessing. The learning processing unit 214 trains the learning model 230 using the preprocessed learning data 225.

[0034] The operation unit 260 and the display unit 270 are a so-called user interface. The operation unit 260 is a device for a user to operate the learning system 200, and is, for example, a keyboard, a mouse, or a touch panel. The display unit 270 is a display for displaying information, and is, for example, a liquid crystal display device.

[0035] FIG. 4 is an explanatory diagram of the processing performed by the learning system. The learning model 230 includes a transformer SA2 and a vibration analysis unit SA3. The transformer SA2 includes an encoder SA2a and a decoder SA2c.

[0036] The learning process includes a first learning process for the transformer SA2 and a second learning process for the vibration analysis unit SA3.

[0037] The first learning process will be described. The acquisition unit 211 inputs the vibration data included in the learning data 225 to the calculation unit 212. The calculation unit 212 performs a phase adjustment process SA1 on the vibration data. The learning processing unit 214 inputs the vibration data after the phase adjustment process to the encoder SA2a. The encoder SA2a outputs a feature amount SA2b, and the decoder SA2c predicts the future value of the vibration data using the feature amount SA2b. The learning processing unit 214 updates the internal parameters of the transformer SA2 based on the predicted value to train the transformer SA2.

[0038] The first learning process is, for example, supervised learning, and the correct label is the future value of the vibration data. For example, when the sequence length of the vibration data included in the learning data 225 is 1000, among them, the vibration data with a sequence length of 800 may be input to the transformer SA2, and the remaining vibration data with a sequence length of 200 may be used as the correct label. The target of prediction is not limited to the future value, and a value in the past compared to the input vibration data or a value at the time between two input vibration data may be predicted.

[0039] The second learning process will be described. The operations of the acquisition unit 211 and the calculation unit 212 are the same as those in the first learning process. The learning processing unit 214 inputs the vibration data after the phase adjustment process to the encoder SA2a of the transformer SA2 obtained by the first learning process, and inputs the feature amount SA2b output by the encoder SA2a to the vibration analysis unit SA3. The vibration analysis unit SA3 infers the vibration state based on the feature amount SA2b. The learning processing unit 214 updates the internal parameters of the vibration analysis unit SA3 based on the inference result to train the vibration analysis unit SA3.

[0040] The vibration analysis unit SA3 may be a model of various algorithms using machine learning. As an example, the vibration analysis unit SA3 may be a model of a non-neural network such as a support vector machine or the k-means method, or may be a model using a deep learning neural network. The second learning process may be either unsupervised learning or supervised learning. In the case of unsupervised learning, for example, the learning processing unit 214 trains the vibration analysis unit SA3 so that the vibration analysis unit SA3 performs detection or classification based on clustering. In the case of supervised learning, the learning processing unit 214 trains the vibration analysis unit SA3 so that the inference result of the vibration analysis unit SA3 approaches the correct label. When the vibration analysis unit SA3 is a classifier, the correct label is a flag indicating which of the plurality of vibration states it corresponds to. When the vibration analysis unit SA3 is a detector, the correct label is a flag indicating whether the object is abnormal or faulty.

[0041] Figure 5 is an explanatory diagram of the process performed by the processing system. The learned model 130 includes a transformer SB2 and a vibration analysis unit SB3. The transformer SB2 includes an encoder SB2a which is the learned encoder SA2a. The vibration analysis unit SB3 is the learned vibration analysis unit SB3. Note that the transformer SB2 may further include a learned decoder SA2c.

[0042] The acquisition unit 111 inputs the vibration data acquired from the sensor 140 to the calculation unit 112. The calculation unit 112 performs a phase adjustment process SB1 on the vibration data. The inference unit 113 inputs the vibration data after the phase adjustment process to the encoder SB2a, and inputs the feature amount SB2b output by the encoder SB2a to the vibration analysis unit SB3. The vibration analysis unit SB3 infers the vibration state based on the feature amount SB2b.

[0043] Regarding the display of the inference results, several examples will be given. The first example will be described. The vibration analysis unit SB3 generates, for example, a three-dimensional support vector from the feature amount by a support vector machine. The processing unit 110 plots the three-dimensional support vector on a 3D graph and displays it on the display. Each plot may be colored with a color corresponding to normal or each failure mode. Alternatively, the boundary between normal or each failure mode may be shown on the 3D graph.

[0044] The second example will be described. The vibration analysis unit SB3 classifies, for example, into the first to sixth vibration states based on the feature amount, and acquires probability information of each vibration state. The processing unit 110 creates a bar chart, a radar chart, etc. from the probability information of the first to sixth vibration states and displays it on the display.

[0045] The third example will be described. The vibration analysis unit SB3 infers the deterioration progress state of each part of the object based on the feature amount. The processing unit 110 creates a bar chart, a radar chart, etc. from the deterioration progress state of each part and displays it on the display. For example, when the object is a motor, the vibration analysis unit SB3 infers the deterioration progress state of the shaft, bearing, rotor core, frame, bearing cover, stator, etc.

[0046] The fourth example will be described. The processing unit 110 arranges the charts of the third example in time series and displays them so that the change in the deterioration progress state can be monitored.

[0047] 2. Phase adjustment process The following describes the details of the phase adjustment process. First, the transformer and the initial phase of the vibration data will be described, and it will be explained that the inference accuracy can be improved by applying the phase adjustment process to the transformer.

[0048] Since the details of the transformer are described in Non-Patent Document 1 etc. mentioned above, here, the outline of the encoder of the transformer and the fact that the attention mechanism performs parallel processing will be explained.

[0049] Although not shown in FIGS. 4 and 5, the encoder of the transformer performs embedding processing on the input vibration data and processing for adding position encoding information. For example, when using a 3-axis sensor and setting the sequence length to 64, the vibration data input to the transformer is 2D data of 64×3. When the internal state dimension number is 512, the embedding processing converts the 2D data of 64×3 into 2D data of 64×512. In the processing of the attention mechanism, the encoder generates matrices Q, K, and V, each of which is 64×512, based on the 2D data after position encoding, and generates 2D data of 64×512 by softmax(QK T )V. The encoder performs normalization processing, feed-forward processing, etc. on the 2D data generated by the attention mechanism, and outputs a feature amount that is 2D data of 64×512.

[0050] FIG. 6 is a diagram for explaining the initial phase of the vibration data. Signals FA to FC indicate signals of each frequency included in the vibration data, and can be obtained, for example, by performing Fourier transform or wavelet transform on the vibration data. In FIG. 6, the amplitudes of signals FA to FC are constant, but the amplitudes may change with time.

[0051] The vibration data DT1 and DT2 are obtained by cutting out a part of the vibration data output by the sensor 140 in time series with the input sequence length of the transformer. The cut-out areas are referred to as windows WD1 and WD2. TS1 and TS2 are the start times of the windows WD1 and WD2. TS1 ≠ TS2. Here, only two windows are shown, but actually, a large number of windows are set. Also, there may be an overlapping part between the windows.

[0052] The initial phases of the signals FA, FB, and FC at the start time TS1 of the window WD1 are PA1, PB1, and PC1. The initial phases of the signals FA, FB, and FC at the start time TS2 of the window WD2 are PA2, PB2, and PC2. Note that PA1, etc. are symbols indicating the phase, but for the convenience of illustration, instead of the phase itself, it is substituted by indicating the signal value at that phase. The windows are set at regular intervals, for example. However, since the initial phases of signals that are not at the frequency corresponding to that interval are generally random, PA1 ≠ PA2, PB1 ≠ PB2, and PC1 ≠ PC2. Therefore, when the vibration data DT1 and DT2 are used as they are, vibration data with unaligned initial phases will be input to the transformer.

[0053] In the analysis of conventional time-series data, a model that sequentially processes time-series data, such as LSTM (Long Short Term Memory), has been used. In such a model, since information is recursively extracted, feature quantities having a relationship in the time direction are calculated. On the other hand, in the attention mechanism of the transformer, matrix operations are performed, so the calculation speed of learning and inference can be improved compared to LSTMs that perform recursive operations. For example, if the calculation time is the same, it is expected that the transformer can improve the accuracy by processing more data.

[0054] However, in the attention mechanism of the transformer, since the vibration data is processed in parallel as a whole to calculate the feature amounts, the temporal relationship in the feature amounts may be weaker compared to the method of recursively performing calculations. For this reason, there is a possibility that the accuracy of vibration analysis may decrease due to being affected by the initial phase of the vibration data. That is, even if the vibration data is actually in the same vibration state, if the initial phases are different, the feature amounts corresponding to the actual vibration state may not be obtained, and it may not be possible to infer the correct vibration state from the feature amounts of the vibration data.

[0055] Therefore, in the present embodiment, by performing learning and inference with the initial phases of the vibration data aligned through phase adjustment processing, an improvement in inference accuracy can be expected in the analysis of vibration data using a transformer.

[0056] FIG. 7 is a diagram for explaining a first method of phase adjustment processing. Hereinafter, it will be described assuming that the calculation unit 112 in FIG. 2 performs phase adjustment processing, but the phase adjustment processing performed by the calculation unit 212 in FIG. 3 is the same.

[0057] The calculation unit 112 performs phase adjustment processing so that the initial phase of the signal at the target frequency is constant in the vibration data of each window. FIG. 7 shows an example in which the frequency of signal FA is set as the target frequency. The calculation unit 112 may use a predetermined frequency as the target frequency. For example, the target frequency may be the vibration frequency of the vibration source 11 in FIG. 1. Alternatively, the calculation unit 112 may obtain the amplitude spectrum or power spectrum of the vibration data and use the frequency having a peak in the spectrum as the target frequency.

[0058] The calculation unit 112 shifts the phase of the signal of each frequency included in the vibration data DT1 of the window WD1 so that the initial phase of the signal FA of the target frequency becomes PA1’ = 0. That is, the phase shift amount is PA1. Let the angular frequencies of the signals FA, FB, and FC be ωa, ωb, and ωc. If this phase shift is converted into a time shift amount, then Δt1 = PA1 / ωa. After the phase adjustment, the initial phases of the signals FB and FC are PB1’ = PB1 - ωb×Δt1 and PC1’ = PC1 - ωc×Δt1. Note that the “time shift” is introduced for the explanation of the phase shift. Actually, the window position is not shifted, and only the phase of the vibration data is shifted while the window position remains the same. That is, only the phase is shifted while the amplitude information at each time within the window is maintained.

[0059] Similarly, the calculation unit 112 shifts the phase of the signal of each frequency included in the vibration data DT2 of the window WD2 so that the initial phase of the signal FA of the target frequency becomes PA2’ = 0. This phase shift corresponds to a time shift of Δt2 = PA2 / ωa. After the phase adjustment, the initial phases of the signals FB and FC are PB2’ = PB2 - ωb×Δt2 and PC2’ = PC2 - ωc×Δt2. Here, an example where PA2’ = PA1’ = 0 is shown, but PA2’ = PA1’ = a constant value is sufficient, and the constant value does not have to be zero.

[0060] By performing such a phase shift, the phase of the target frequency at the start positions TS1 and TS2 of the vibration data DT1 and DT2 becomes a constant value. For frequencies other than the target frequency, relative phase information with respect to the phase of the target frequency can be obtained. Thereby, even when a transformer that extracts feature amounts by parallel processing is used, the influence of the initial phase is reduced, and an improvement in the accuracy of vibration analysis is expected.

[0061] The calculation unit 112 converts the vibration data DT1 and DT2 into spectral data, shifts the phase of the spectral data, and inverse-transforms the spectral data after the phase shift to obtain the vibration data after phase adjustment, and inputs the vibration data after the phase adjustment into the transformer. The conversion is, for example, a Fourier transform or a wavelet transform.

[0062] FIG. 8 is a diagram for explaining a second method of the phase adjustment process. The calculation unit 112 performs a phase adjustment process so that the initial phases of the signals at a plurality of target frequencies are constant in the vibration data of each window. FIG. 8 shows an example in which the frequencies of signals FA to FC are used as a plurality of target frequencies.

[0063] The calculation unit 112 determines the time shifts Δt1 and Δt2 such that TS1 - Δt1 = TS2 - Δt2 = TS0 with reference to TS0. The relationship between the time shift and the phase shift, and the phase shift using the Fourier transform are the same as those in the first method. Also, the same as the first method is that the window position is not shifted and only the phase of the vibration data is shifted. The initial phases of the signals FA, FB, and FC after the phase adjustment are PA1', PB1', and PC1' in all windows and are always constant.

[0064] By performing such a phase shift, the phases of a plurality of target frequencies at the start positions TS1 and TS2 of the vibration data DT1 and DT2 become constant values. As a result, even when using a transformer that extracts feature amounts by parallel processing, the influence of the initial phase is reduced, and an improvement in the accuracy of vibration analysis is expected. Here, FIG. 8 has been described as an example in which the frequencies of signals FA to FC are used as a plurality of target frequencies, but generally, the phase shift by the second method corresponds to performing phase adjustment on all signal components of the vibration data. Ideally, all the initial phases of the stationary signal components included in the vibration data always become constant after the phase adjustment.

[0065] A third method of phase adjustment processing will be described. The arithmetic unit 112 sets a plurality of windows so that the initial phases at one or more target frequencies are constant. For example, when the frequency of the signal FA is set as the target frequency, a plurality of windows are set with the period of the signal FA or an integral multiple of the period. Thereby, the initial phase of the signal FA in each window becomes constant. Also by the third method, the same effect as the first method or the second method can be obtained.

[0066] In the present embodiment, the time-series data processing method is a processing method using a learned model 130 including an encoder SB2a. The encoder SB2a is a neural network that analyzes time-series data and does not have a recursive structure, and uses an attention mechanism. The learned model 130 is learned to output a feature amount SB2b from the encoder SB2a based on time-series data. The time-series data processing method includes a step of performing a phase adjustment process for making the initial phases of the time-series data constant at the start positions TS1, TS2 of each window, by acquiring a plurality of time-series data DT1, DT2 in a plurality of windows WD1, WD2. The time-series data processing method includes a step of acquiring the feature amount SB2b from the encoder SB2a by inputting the time-series data after performing the phase adjustment process into the learned model 130.

[0067] According to this embodiment, since the encoder SB2a does not have a recursive structure and uses an attention mechanism, the encoder SB2a generates feature quantities from time-series data by parallel processing. Therefore, compared with a model having a recursive structure, feature quantities can be calculated at high speed by parallel processing, and the calculation time for learning or inference can be shortened. On the other hand, the feature quantities output by the encoder SB2a have less information related to the time direction compared with the feature quantities output by a model having a recursive structure. For this reason, the feature quantities are likely to be affected by the initial phase of the time-series data, and there is a possibility that the accuracy in the analysis of the time-series data using the feature quantities will decrease. According to this embodiment, by adjusting so that the phases of the time-series data at the start positions TS1 and TS2 of the window become constant, the influence of the initial phase can be reduced, and an improvement in accuracy can be expected in the analysis of the time-series data using the feature quantities.

[0068] In addition, in the above-described embodiment, an example of inferring the vibration state based on the vibration data measured by the vibration sensor has been described. However, the method of this embodiment can be applied when performing analysis regarding the time-series data based on various time-series data. The time-series data may be, for example, time-series data measured by a sensor other than the vibration sensor. Sensors other than the vibration sensor are, for example, a temperature sensor, a microphone, a voltmeter, an ammeter, or the like. Alternatively, the time-series data may be data related to weather such as air temperature or atmospheric pressure, or data related to marketing such as sales, price, or inventory.

[0069] Also, in the above-described embodiment, an example of using a transformer as the model has been described, but the model may be a neural network that analyzes time-series data and does not have a recursive structure, as long as it uses an attention mechanism. As an example, there are various derivative models based on the transformer, and those may be adopted. The derivative models are, for example, Autoformer, Informer, PatchTST, NSTransformer, or Crossformer.

[0070] Also, as described with reference to FIG. 7, the time-series data processing method may perform a phase adjustment process so that the initial phases of the signals FA at the target frequencies included in the plurality of time-series data DT1 and DT2 become a constant value PA1’ = PA2’.

[0071] According to the present embodiment, the phases of the target frequencies at the start positions TS1 and TS2 of the vibration data DT1 and DT2 become constant values. Also, for frequencies other than the target frequency, information on the relative phase with reference to the phase of the target frequency can be obtained. Thereby, even when an encoder SB2a that extracts feature amounts by parallel processing is used, the influence of the initial phase on the feature amounts is reduced.

[0072] Also, as described with reference to FIG. 8, the time-series data processing method may perform a phase adjustment process so that the initial phases of the signals FA to FC at the plurality of target frequencies included in the plurality of time-series data DT1 and DT2 become constant values PA1’, PB1’, and PC1’ for each target frequency.

[0073] According to the present embodiment, the phases of the plurality of target frequencies at the start positions TS1 and TS2 of the vibration data DT1 and DT2 become constant values. Thereby, even when an encoder SB2a that extracts feature amounts by parallel processing is used, the influence of the initial phase on the feature amounts is reduced.

[0074] Also, as described with reference to FIG. 8, the time-series data processing method may perform a phase adjustment process so that the initial phases of all the stationary signal components included in the plurality of time-series data DT1 and DT2 become constant values for the corresponding frequencies.

[0075] According to the present embodiment, the phases of all the stationary signal components at the start positions TS1 and TS2 of the vibration data DT1 and DT2 become constant values. Thereby, even when an encoder SB2a that extracts feature amounts by parallel processing is used, the influence of the initial phase on the feature amounts is reduced.

[0076] Also, in this embodiment, the step of performing phase adjustment processing in the time-series data processing method may include a step of converting each time-series data of a plurality of time-series data DT1 and DT2 into spectral data, and a step of obtaining a time shift amount for shifting the phase at the frequency of interest in the spectral data to a constant value. The step of performing phase adjustment processing may include a step of shifting the phase at each frequency of the spectral data by a phase shift amount corresponding to the time shift amount, and a step of inverse-converting the spectral data with the shifted phase into time-series data. As described with reference to FIG. 7, the phase at the frequency of interest is shifted from, for example, PA1 to zero. This phase shift amount is PA1. When this is converted into a time shift amount, Δt1 = PA1 / ωa. The phase shift amounts of signals FB and FC corresponding to this time shift amount are ωb×Δt1 and ωc×Δt1. Note that the time shift amount is a quantity used in calculation, and it is the phase that is actually shifted, not the time.

[0077] According to this embodiment, the initial phase can be set to a constant value in the signal FA at the frequency of interest included in the plurality of time-series data DT1 and DT2. Also, by shifting the phase by the phase shift amount corresponding to the same time shift amount, the phase can be shifted without disrupting the phase relationship between the frequency of interest and other frequencies.

[0078] Also, as described in the third method of phase adjustment processing, in the time-series data processing method, the step of performing phase adjustment processing may include a step of setting a plurality of windows in which the initial phase in the signal at the frequency of interest is a constant value.

[0079] According to this embodiment, by obtaining the time-series data in the plurality of set windows, a plurality of time-series data with a constant initial phase can be obtained.

[0080] In this embodiment, the plurality of time-series data DT1 and DT2 may be measurement data of physical quantities related to vibrations measured from an object. The time-series data processing method may include a step of inferring the vibration state related to at least one of state monitoring, quality control, and predictive maintenance of the object 10 based on the feature quantity SB2b obtained from the encoder SB2a.

[0081] According to this embodiment, the vibration state of the object 10 can be inferred based on the feature quantity output by the encoder based on the time-series data. By adjusting the initial phase of the time-series data to a constant value, the influence of the initial phase on the feature quantity is reduced, and an improvement in the accuracy of inferring the vibration state can be expected.

[0082] Also in this embodiment, the learned model 130 may include a vibration analysis unit SB3 learned to infer the vibration state from the feature quantity SB2b. The feature quantity SB2b may be two-dimensional data. In the time-series data processing method, the inferring step may include a step of inputting all elements of the two-dimensional data into the vibration analysis unit SB3.

[0083] Alternatively, in this embodiment, in the time-series data processing method, the inferring step may include a step of inputting the two-dimensional data into the vibration analysis unit SB3 one row or one column at a time.

[0084] According to this embodiment, it becomes possible to use the feature quantity in an appropriate manner according to the nature of the feature quantity or according to the content to be inferred. For example, when the row vectors of the feature quantity are orthogonalized, each row vector contains independent information. Therefore, by inputting the two-dimensional data into the vibration analysis unit SB3 one row at a time, the information of each row vector can be effectively utilized.

[0085] Furthermore, a processing system 100 that executes the above time-series data processing method may be configured. That is, the processing system 100 includes a storage unit 120 that stores a learned model 130 including an encoder SB2a, and a processing unit 110 that performs processing using the learned model 130. The encoder SB2a is a neural network that analyzes time-series data and has no recursive structure, and uses an attention mechanism. The learned model 130 is learned to output a feature amount SB2b from the encoder SB2a based on time-series data. The processing unit 110 acquires a plurality of time-series data DT1, DT2 in a plurality of windows WD1, WD2, and performs a phase adjustment process to make the initial phases of the time-series data at the start positions TS1, TS2 of each window constant. The processing unit 110 inputs the time-series data after the phase adjustment process into the learned model 130 to obtain the feature amount SB2b from the encoder SB2a.

[0086] Also, in this embodiment, the learning method performs learning of the learning model 230 so as to infer time-series data outside the time range of the time-series data to be analyzed. The learning model 230 includes an encoder SA2a and a decoder SA2c. The encoder SA2a and the decoder SA2c are neural networks that analyze time-series data and have no recursive structure, and use an attention mechanism. The learning method includes a step of acquiring a plurality of time-series data DT1, DT2 in a plurality of windows WD1, WD2, and performing a phase adjustment process to make the initial phases of the time-series data at the start positions TS1, TS2 of each window constant. The learning method includes a step of inputting the time-series data after the phase adjustment process into the learning model 230 and performing learning so that the learning model 230 infers time-series data outside the time range.

[0087] Also, in this embodiment, the learning method may perform a phase adjustment process so that the initial phases in the signal FA of the frequency of interest included in the plurality of time-series data DT1, DT2 become a constant value PA1' = PA2'.

[0088] Also, in this embodiment, for the signals FA to FC of a plurality of target frequencies included in the plurality of time-series data DT1 and DT2, phase adjustment processing may be performed so that the initial phase becomes a constant value PA1', PB1', PC1' for each target frequency.

[0089] Also, in this embodiment, the step of performing phase adjustment processing in the learning method may include a step of converting each time-series data of the plurality of time-series data DT1 and DT2 into spectral data, and a step of obtaining a time shift amount for shifting the phase at the target frequency of the spectral data to a constant value. The step of performing phase adjustment processing may include a step of shifting the phase at each frequency of the spectral data by a phase shift amount corresponding to the time shift amount, and a step of inverse-converting the spectral data with the shifted phase into time-series data.

[0090] Also, in this embodiment, the step of performing phase adjustment processing in the learning method may include a step of setting a plurality of windows in which the initial phase in the signal of the target frequency is a constant value.

[0091] Also, in this embodiment, the plurality of time-series data DT1 and DT2 may be measurement data of a physical quantity related to vibration measured from an object. The learning model 230 may include a vibration analysis unit SA3 that infers the state of vibration from the feature quantity SA2b. The learning method may include a step of inputting the feature quantity SA2b output by the encoder SA2a to the vibration analysis unit SA3 and performing learning of the vibration analysis unit SA3 so that the vibration analysis unit SA3 infers the state of vibration. The state of vibration is a state related to at least one of state monitoring, quality control, and predictive maintenance of the object 10.

[0092] Note that the effects achieved by the above learning method are the same as those of the time-series data processing method described above, so the description thereof is omitted.

[0093] Furthermore, a learning system 200 that executes the above learning method may be configured. The learning system 200 includes a storage unit 220 that stores a learning model 230 including an encoder SA2a and a decoder SA2c, and a processing unit 210 that performs learning of the learning model 230. The encoder SA2a and the decoder SA2c are neural networks that analyze time-series data and do not have a recursive structure, and use an attention mechanism. The learning model 230 infers time-series data outside the time range of the time-series data to be analyzed. The processing unit 210 acquires a plurality of time-series data DT1, DT2 in a plurality of windows WD1, WD2, and performs a phase adjustment process to make the initial phases of the time-series data at the start positions TS1, TS2 of each window constant. After performing the phase adjustment process, the processing unit 210 inputs the time-series data to the learning model 230 and performs learning so that the learning model 230 infers time-series data outside the time range.

[0094] 3. Orthogonalization of Feature Quantities As described below, the transformer may be learned so as to orthogonalize the feature quantities. This orthogonalization may be combined with the above-described phase adjustment process.

[0095] FIG. 9 is a diagram for explaining the orthogonalization of feature quantities. The feature quantities output by the encoder are two-dimensional data of the internal state dimension number × sequence length. Hereinafter, the two-dimensional data is also referred to as a matrix. In FIG. 9, one circle represents one element of the matrix, and h1, h2, h3, ···, h m is the row vectors of each row of the matrix. m is an integer of 2 or more, and here it is the internal state dimension number.

[0096] In the learning of the transformer, the learning processing unit 214 uses the loss function L shown in FIG. 9. The first term of the loss function L is the MSE loss indicating the error between the inference result and the correct label, and the second term is the orthogonalization loss indicating the correlation between the vectors. MSE is an abbreviation for Mean Squared Error. The x of the MSE loss i is the data output by the transformer, and y iis the correct data. n is an integer of 2 or more and is the output sequence length of the transformer. h of the orthogonality loss i , h j is a row vector obtained from the feature amount. i and j are integers from 1 to m. Cov[h i , h j is the covariance of h i and h j , and Ver[h i is the variance of h i . λ is a hyperparameter. Note that the first term of the loss function L is not limited to the MSE loss, and any function that can evaluate the error between xi and yi may be used.

[0097] The row vectors h1, h2, h3, ···, h of the feature amount m become more orthogonal to each other, the orthogonality loss becomes smaller. By performing the learning of the transformer using the loss function L including such an orthogonality loss, the row vectors h1, h2, h3, ···, h of the feature amount output by the encoder m approach a state where they are orthogonal to each other.

[0098] Fig. 10 is a diagram for explaining the separation of information by orthogonality. The left diagram of Fig. 10 conceptually shows the information included in the feature amount when not orthogonalized. The row vectors h1, h2, h3, ···, h of the feature amount m are not linearly independent, so various information is not separated and is mixed and included in each row vector. For example, the amplitude information and the phase information are mixed and included in each row vector.

[0099] The right diagram of Fig. 10 conceptually shows the information included in the feature amount when orthogonalized. The row vectors h1, h2, h3, ···, h of the feature amount m have a small correlation with each other, that is, are close to linearly independent, so various information is separated and included in each row vector. That is, one row vector can be considered to correspond to a certain piece of information. For example, h1 corresponds to the first amplitude information, h2 corresponds to the first phase information, h3 corresponds to the second amplitude information, h mcorresponds to other information. Here, although amplitude information and phase information are taken as examples, the information included in the feature amount only needs to be analyzable by the vibration analysis unit, and it does not need to be information that can be understood by humans in terms of meaning.

[0100] Using a feature amount in which each row vector has independent information as shown in the right figure of FIG. 10 rather than a feature amount in which information is mixed in each row vector as shown in the left figure of FIG. 10 can be expected to improve the accuracy of vibration analysis. As an example, assume that a feature appears in the first amplitude information in a certain vibration state. If the information is mixed, the first amplitude information will be dispersed among many row vectors, making it difficult to grasp the features of the first amplitude information. On the other hand, by separating the first amplitude information into the row vector h1 through orthogonalization, the features of the first amplitude information become easier to grasp, and the vibration state becomes easier to discriminate.

[0101] FIG. 11 is a diagram for explaining another example of orthogonalization. The learning processing unit 214 executes a plurality of batches in one learning. As a result, the feature amount of the internal state dimension number × sequence length shown in the left figure of FIG. 11 can be obtained for the number of batches. As shown in the middle figure of FIG. 11, the learning processing unit 214 vertically multiplies the columns of the feature amounts obtained in one batch to form a column of (internal state dimension number × sequence length), and arranges it horizontally for the number of batches to form a matrix of (internal state dimension number × sequence length) × batch size length. The batch size length is the number of batches. The learning processing unit 214 calculates the orthogonalization loss using the row vectors h1, h2, h3, ···, h m In this example, m is (internal state dimension number × sequence length).

[0102] FIGS. 12 and 13 are diagrams for explaining the relationship between orthogonalization and phase adjustment.

[0103] FIG. 12 illustrates orthogonalization when using an LSTM. Here, an example is shown where the input dimension number is 3 and the sequence length is t. Since the LSTM extracts information recursively, there is a relationship between the feature amounts at each time. The row vectors h1, h2, h3, ···, h mSince it extracts the elements of the feature amount in the time direction, each row vector contains the time-direction relationship of the recursively extracted information. By orthogonalizing such row vectors, it is considered that row vectors containing information corresponding to the vibration state can be obtained.

[0104] FIG. 13 illustrates orthogonalization when using a transformer. In a transformer, since feature amounts are extracted by parallel processing, it is considered that there is no relationship between the feature amounts at each time. Therefore, even if orthogonalization is performed as it is, each row vector does not contain the time-direction relationship of the information. For this reason, if the initial phase of the vibration data is different, feature amounts may be extracted as data representing different vibration states, and appropriate information may not appear in the row vectors. In the present embodiment, by performing phase adjustment processing, the initial phase is constant for the target frequency. As a result, since the influence of the initial phase on the feature amount is reduced, it can be expected that feature amounts appropriately representing the vibration state are extracted and appropriate information appears in the row vectors.

[0105] In addition, although an example of orthogonalizing the row vectors of the feature amounts has been described above, the column vectors of the feature amounts may be orthogonalized. The fact that information is more easily extracted by aligning the initial phases by phase adjustment is the same for column vectors.

[0106] In the present embodiment, as described with reference to FIG. 9, the learning method may include a step of calculating an orthogonalization loss indicating the correlation between the row vectors or column vectors of the two-dimensional data from the two-dimensional data of the feature amounts output by the encoder, and performing learning of the learning model 230 using a loss function L including the orthogonalization loss.

[0107] Also, in the present embodiment, when the i-th row or the i-th column of the two-dimensional data is h i the orthogonalization loss may be calculated by the second term on the right side of the equation shown in FIG. 9.

[0108] According to this embodiment, since the transformer is trained using an orthogonality loss that indicates the correlation between row vectors or column vectors of two-dimensional data, it is trained so that the correlation between row vectors or column vectors of the feature quantities output by the encoder becomes small. As described with reference to FIG. 10, since one row vector can be considered to correspond to a certain piece of information, an improvement in the accuracy of inference based on time-series data can be expected.

[0109] Also, as described with reference to FIG. 11, the two-dimensional data may be a matrix of (sequence length × internal state dimension number) × batch size length.

[0110] According to this embodiment, the feature quantity includes not only the information in one sequence but also the information of a plurality of batches. It is expected that there is some correlation between batches, and by orthogonalizing the feature quantity including the information of such a plurality of batches, a feature quantity that more accurately represents the state information can be obtained. It is expected that there is some correlation between batches. As shown in the middle and right figures of FIG. 11, for a plurality of batches defined by the batch size length, let the row vectors h1, h2, h3, ···, hm be the arrangement of the respective elements corresponding to the product of the internal state dimension number and the sequence length. In this case, by learning so that the row vectors h1, h2, h3, ···, hm are in an orthogonal state, it can be expected to obtain a feature quantity in which information is considered and organized not only between internal state dimensions but also between sequences, that is, with respect to time evolution.

[0111] 4. Examples of Vibration Analysis for Condition Monitoring, Quality Control, and Predictive Maintenance Hereinafter, as an example to which the processing system 100 of this embodiment can be applied, examples of vibration analysis related to condition monitoring, quality control, and predictive maintenance are shown.

[0112] Condition Monitoring refers to the technology of continuously or periodically monitoring the current operating state or performance of machinery or equipment. The main purpose of condition monitoring is to monitor specific parameters such as vibration, sound, temperature, or pressure, and evaluate the soundness of machinery or equipment. Examples of systems for performing condition monitoring include the following (a) to (f).

[0113] (a) A system that attaches sensors to industrial motors and issues an alarm when abnormal vibration or a temperature rise is detected.

[0114] (b) A system that installs vibration sensors in each part of a building and monitors the vibration during an earthquake or strong wind to monitor the soundness of the building's structure.

[0115] (c) A system that has vibration sensors attached to monitor the performance of a gas turbine and warns the operator when abnormal vibration is detected.

[0116] (d) A system that has vibration sensors attached to monitor the soundness of heavy machinery in a mine or quarry and helps detect early wear or failure.

[0117] (e) A system that has vibration sensors attached to monitor the soundness of each machine on a factory production line and issues an alarm when abnormal vibration is detected.

[0118] (f) A system that uses vibration sensors attached to the blades or gearboxes of wind power turbines to detect whether abnormal wear or damage has occurred.

[0119] Quality Control refers to the process of confirming whether a product or service meets the defined quality standards or requirements. Examples of systems for performing quality control include the following (g) to (k).

[0120] (g) A system that uses cameras or sensors on the manufacturing line to check the quality of products such as dimensions, colors, or shapes in real time and automatically eliminates products that deviate from the standards.

[0121] (h) A system that checks the vibration pattern when a product is operated using a vibration sensor after the assembly process of a product where multiple parts are combined. If abnormal vibration is detected, defects in assembly or malfunctions of parts are suspected.

[0122] (i) A system that operates a motor or generator after manufacturing and measures the vibration during its operation using a vibration sensor. If vibration exceeding specific criteria is detected, internal imbalance or damage is considered.

[0123] (j) A system that installs an electronic device on a vibration test bench and monitors the impact of vibration using a vibration sensor to verify whether a newly manufactured electronic device has the set vibration resistance.

[0124] (k) A system that monitors the vibration during operation using a vibration sensor to check whether a new type of railway vehicle or aircraft has the designed operating performance. If abnormal vibration is detected, identify the cause and use it to improve quality.

[0125] Predictive Maintenance refers to an approach that collects and analyzes operation data or state data of equipment or machinery. The purpose of predictive maintenance is to predict the risk of future failures or performance degradation. Examples of systems for predictive maintenance include the following (l) - (q).

[0126] (l) A system that attaches sensors to industrial robots, analyzes the sensor data when the industrial robots are operating and the sensor data when they failed in the past, detects signs indicating precursors of specific parts failing, and schedules the replacement of those parts predictively.

[0127] (m) A system that uses sensors attached to the wheels of a railway vehicle to monitor the degree of wear and predict the replacement time of the wheels based on data indicating excessive wear.

[0128] (n) A system that monitors the condition of pipes or valves in an oil plant using sensors and predicts the possibility of future leaks.

[0129] (o) A system that analyzes sensor data when an elevator or escalator is operating and predicts the risk of failure before parts need to be replaced.

[0130] (p) A system that monitors the efficiency of industrial cooling devices or air conditioners based on sensor data and predicts part wear or failure.

[0131] (q) A system that collects and analyzes sensor data when agricultural machinery is operating, predicts part wear or failure, and supports scheduling appropriate maintenance activities.

[0132] Although the present embodiment has been described in detail as above, those skilled in the art will easily understand that many modifications can be made without substantially departing from the novel matters and effects of the present disclosure. Therefore, all such modifications are intended to be included within the scope of the present disclosure. For example, in the specification or drawings, terms that have been described at least once together with broader or synonymous different terms can be replaced with those different terms anywhere in the specification or drawings. Also, all combinations of the present embodiment and modification examples are included within the scope of the present disclosure. Further, the configurations and operations of the processing system, learning system, learned model, learning model, sensor, object, etc. are not limited to those described in the present embodiment, and various modified implementations are possible.

Description of Reference Numerals

[0133] 10... Object, 11... Vibration source, 100... Processing system, 110... Processing unit, 111... Acquisition unit, 112... Arithmetic unit, 113... Inference unit, 120... Memory unit, 130... Learned model, 140... Sensor, 150... Presentation unit, 200... Learning system, 210... Processing unit, 211... Acquisition unit, 212... Arithmetic unit, 214... Learning processing unit, 220... Memory unit, 225... Learning data, 230... Learning model, 260... Operation unit, 270... Display unit, DT1, DT2... Vibration data, FA, FB, FC... Signals, SA1, SB1... Phase adjustment processing, SA2, SB2... Transformers, SA2a, SB2a... Encoders, SA2b, SB2b... Feature quantities, SA2c... Decoder, SA3, SB3... Vibration analysis unit, TS1, TS2... Initial positions, h1~h m ... Row vector, Δt1, Δt2... Time shift amounts

Claims

1. A neural network that analyzes time-series data and has no recursive structure, includes an encoder using an attention mechanism, and uses a trained model learned to output the feature amounts from the encoder based on the time-series data, obtaining a plurality of time-series data in a plurality of windows, and performing a phase adjustment process for making the initial phases of the time-series data at the start positions of the respective windows constant; obtaining the feature amounts from the encoder by inputting the time-series data after the phase adjustment process to the trained model; A time-series data processing method characterized by including these steps.

2. In Claim 1, the phase adjustment process is performed such that the initial phases in the signals of the target frequencies included in the plurality of time-series data become constant values. A time-series data processing method characterized by this.

3. In Claim 1, the phase adjustment process is performed for the signals of a plurality of target frequencies included in the plurality of time-series data such that the initial phase becomes a constant value for each target frequency. A time-series data processing method characterized by this.

4. In Claim 1, the phase adjustment process is performed for all the stationary signal components included in the plurality of time-series data such that the initial phases of the corresponding frequencies become constant values. A time-series data processing method characterized by this.

5. In Claim 2, the step of performing the phase adjustment process includes the step of converting each time-series data of the plurality of time-series data into spectral data; the step of obtaining a time shift amount for shifting the phase at the target frequency of the spectral data to the constant value; the step of shifting the phases at the respective frequencies of the spectral data by a phase shift amount corresponding to the time shift amount; and the step of inversely converting the spectral data with the shifted phase into time-series data. A time-series data processing method characterized by including these steps.

6. In Claim 2, the step of performing the phase adjustment process includes the step of setting a plurality of windows in which the initial phase in the signal of the target frequency is the constant value. A time-series data processing method characterized by this.

7. In Claim 1, the plurality of time-series data are measurement data of physical quantities related to vibrations measured from an object, A time-series data processing method, comprising the step of inferring a vibration state related to at least one of state monitoring, quality control, and predictive maintenance of the object based on the feature amount obtained from the encoder.

8. In claim 7, The learned model includes a vibration analysis unit learned to infer the vibration state from the feature amount, The feature amount is two-dimensional data, The step of inferring includes the step of inputting all elements of the two-dimensional data into the vibration analysis unit, and is a time-series data processing method characterized in that.

9. In claim 7, The learned model includes a vibration analysis unit learned to infer the vibration state from the feature amount, The feature amount is two-dimensional data, The step of inferring includes the step of inputting the two-dimensional data into the vibration analysis unit one row or one column at a time, and is a time-series data processing method characterized in that.

10. A neural network that analyzes time-series data and has no recursive structure, includes an encoder and a decoder using an attention mechanism, and learns a learning model to infer time-series data outside the time range of the time-series data to be analyzed. The steps of acquiring a plurality of time-series data in a plurality of windows and performing a phase adjustment process to make the initial phase of the time-series data at the start position of each window constant. After performing the phase adjustment process, input the time-series data into the learning model, and perform the learning so that the learning model infers the time-series data outside the time range. A learning method characterized by including.

11. In claim 10, The learning method is characterized in that the phase adjustment process is performed so that the initial phase in the signal of the target frequency included in the plurality of time-series data becomes a constant value.

12. In claim 10, The learning method is characterized in that, for signals of a plurality of target frequencies included in the plurality of time-series data, the phase adjustment process is performed so that the initial phase becomes a constant value for each target frequency.

13. In claim 11, The step of performing the phase adjustment process is The step of converting each time-series data of the plurality of time-series data into spectrum data. The step of obtaining a time shift amount for shifting the phase at the target frequency of the spectrum data to the constant value. shifting the phase at each frequency of the spectral data by a phase shift amount corresponding to the time shift amount; inverse-transforming the spectral data with the shifted phase into time-series data; A learning method characterized by including the above steps.

14. In claim 11, the step of performing the phase adjustment process includes a step of setting a plurality of windows in which the initial phase in the signal of the target frequency is the constant value. A learning method characterized by this.

15. In claim 10, the plurality of time-series data are measurement data of physical quantities related to vibrations measured from an object, the learning model includes a vibration analysis unit learned to infer the state of the vibration from the feature amounts, inputting the feature amounts output by the encoder to the vibration analysis unit, and learning the vibration analysis unit so that the vibration analysis unit infers the state of the vibration related to at least one of state monitoring, quality control, and predictive maintenance of the object. A learning method characterized by including the step of performing.

16. In claim 10, calculating an orthogonality loss indicating the correlation between row vectors or column vectors of the two-dimensional data from the two-dimensional data of the feature amounts output by the encoder, and using the loss function including the orthogonality loss to perform the learning of the learning model. A learning method characterized by including the step of performing.

17. In claim 16, Let h be the i-th row or the i-th column of the two-dimensional data ｉ when 【Number 1】 characterized by calculating the orthogonality loss thereby. A learning method.

18. In claim 17, the two-dimensional data is a matrix of (sequence length × internal state dimension number) × batch size length. A learning method characterized by this.

19. a storage unit that stores a learned model that is a neural network that analyzes time-series data and has no recursive structure, includes an encoder using an attention mechanism, and is learned to output the feature amounts from the encoder based on the time-series data; a processing unit that performs processing using the learned model; including the processing unit acquires a plurality of time-series data in a plurality of windows, performs a phase adjustment process to make the initial phase of the time-series data at the start position of each window constant, acquires the feature amounts from the encoder by inputting the time-series data after performing the phase adjustment process to the learned model; A processing system characterized by this.

20. A neural network that analyzes time-series data and has no recursive structure, including an encoder and a decoder using an attention mechanism, and a storage unit that stores a learning model for inferring time-series data outside the time range of the time-series data to be analyzed. A processing unit that performs learning of the learning model. Including The processing unit Obtains a plurality of time-series data in a plurality of windows, and performs a phase adjustment process to make the initial phase of the time-series data at the start position of each window constant. Inputs the time-series data after the phase adjustment process into the learning model, and performs the learning so that the learning model infers the time-series data outside the time range. A learning system characterized by the above.

Images