Information processing device, information processing method, and program

The information processing device uses a machine learning model to generate local waveforms for continuous wavelet transform, addressing precision issues in wavelet transforms and enhancing time-frequency analysis for equipment anomaly detection and service life estimation.

JP2024157752AActive Publication Date: 2024-11-08KK TOSHIBA
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
JP2023072296
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-04-26
Publication Date
2024-11-08
Estimated Expiration
2043-04-26

AI Technical Summary

Technical Problem

Existing wavelet transform methods, such as discrete and continuous wavelet transforms, struggle to achieve high precision in time-frequency analysis due to limitations in time-frequency resolution and complexity in configuring task-specific functions using neural networks.

Method used

An information processing device that utilizes a machine learning model to generate a local waveform for continuous wavelet transform, learned through unsupervised or supervised learning, to enhance the precision of signal analysis.

Benefits of technology

Enables high-precision signal analysis by configuring task-specific wavelets, improving the accuracy of time-frequency representation and state estimation in devices like bearings.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024157752000001_ABST
    Figure 2024157752000001_ABST
Patent Text Reader

Abstract

To provide an information processing device, an information processing method, and a program capable of further accurately executing analysis of a signal by using wavelet transformation.SOLUTION: An information processing device includes a processing unit. The processing unit generates a generation function which is a function for generating a local waveform used for consecutive wavelet transformation and at least a part of which is represented by a machine learning model. The processing unit learns the generation function by using learning data containing a first input signal.SELECTED DRAWING: Figure 1
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] An embodiment of the present invention relates to an information processing device, an information processing method, and a program. [Background technology]

[0002] Features extracted from time series signals are used for failure prediction and maintenance, such as detecting anomalies in equipment and estimating useful life. For equipment that operates periodically, such as bearings, features related to the frequency components of the time series and time information are used.

[0003] A technology has been proposed that applies wavelet transform (wavelet analysis) to obtain the features of time series signals. There are two types of wavelet transform: discrete wavelet transform (discrete transform), which allows for high-speed analysis, and continuous wavelet transform (continuous transform), which allows for high-resolution time-frequency analysis by introducing redundancy into the analysis results. Wavelet transform provides excellent features for estimating the state of a time series. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 7042315 [Patent Document 2] JP 2004-288144 A [Patent Document 3] JP 2004-168286 A [Non-patent literature]

[0005] [Non-Patent Document 1] Gabriel Michaua et al., “Fully learnable deep wavelet transform for unsupervised monitoring of high-frequency time series”, PNAS 2022 Vol. 119 No. 8 e2106598119 Summary of the Invention [Problem to be solved by the invention]

[0006] An object of the present invention is to provide an information processing device, an information processing method, and a program that enable signal analysis using wavelet transform to be performed with higher accuracy. [Means for solving the problem]

[0007] According to an embodiment, an information processing apparatus includes a processor that generates a generating function for generating a local waveform to be used in a continuous wavelet transform, the generating function being at least partially represented by a machine learning model, and that learns the generating function using training data including a first input signal. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 is a block diagram of an information processing apparatus according to a first embodiment. [Diagram 2] FIG. 4 is a diagram showing an example of a signal stored in a storage unit. [Diagram 3] FIG. 13 is a diagram showing an example of the configuration of a generation function. [Figure 4] 13 is a flowchart of a learning process. [Diagram 5] 13 is a flowchart of a process for calculating a wavelet representation. [Figure 6] 13 is a flowchart of an estimation process. [Figure 7] FIG. 4 is a diagram showing an example of a display screen for an estimation result according to the first embodiment. [Figure 8] FIG. 4 is a diagram showing an example of a display screen for an estimation result according to the first embodiment. [Figure 9]FIG. 11 is a block diagram of an information processing apparatus according to a second embodiment. [Figure 10] FIG. 11 is a diagram showing an example of a display screen for an estimation result according to the second embodiment. [Figure 11] FIG. 2 is a hardware configuration diagram of an information processing apparatus according to the first or second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of an information processing apparatus according to the present invention will be described in detail with reference to the accompanying drawings.

[0010] One of the techniques for frequency analysis is the Fast Fourier transform (FFT). Since the FFT loses all time information contained in the signal, it is not suitable for analysis of sudden phenomena. One technique that overcomes the drawback of the FFT, which loses time information, is the Short-Term Fourier transform (STFT). The STFT can capture time-frequency information by performing a Fourier transform on the signal within a compact window function while shifting it over time. However, due to the uncertainty principle, there is a trade-off between time precision and frequency precision, so it is not possible to perform high-precision analysis covering multiple frequency domains using the STFT, which uses a fixed window width.

[0011] Wavelet analysis is a method that enables time-frequency analysis by examining the time correlation of a single local waveform called an analyzing wavelet while changing the scale. Unlike STFT, wavelet analysis can handle various frequencies by zooming in on high-frequency components with a smaller scale and zooming in on low-frequency components with a larger scale.

[0012] The results of the local waveforms used in wavelet analysis may differ depending on the function form (the form of the function used to generate the local waveforms). For discrete wavelet transform, a technology has been proposed that uses a neural network trained to match the generated local waveforms to the data, enabling more accurate analysis.

[0013] On the other hand, the discrete wavelet transform has low time-frequency resolution and cannot detect slight frequency differences. Therefore, even in a configuration using a neural network, state estimation that requires detailed frequency information remains difficult.

[0014] The continuous wavelet transform uses existing functions such as the Mexican hat, but because the wavelet construction conditions are complex, it has been difficult to realize task-specific functions using neural networks.

[0015] In the following embodiment, a more versatile wavelet is used to realize a continuous wavelet transform, making it possible to perform signal analysis using the wavelet transform with higher accuracy.

[0016] (First embodiment) The information processing device of the first embodiment represents a generating function for generating a local waveform (analyzing wavelet) used in a continuous wavelet transform by a machine learning model, and learns the generating function by unsupervised learning.

[0017] Fig. 1 is a block diagram showing an example of a configuration of an information processing device 100 according to a first embodiment. As shown in Fig. 1, the information processing device 100 includes a storage unit 151, a display unit 152, a reception unit 101, an output control unit 102, a generation unit 111, a learning unit 112, a calculation unit 121, and an estimation unit 122.

[0018] The storage unit 151 stores various information used in the information processing device. For example, the storage unit 151 stores a signal used for learning and a signal to be analyzed.

[0019] 2 is a diagram showing an example of a signal stored in the storage unit 151. In the example of FIG. 2, the storage unit 151 stores an id, which is identification information for identifying a signal, and 1 ~t 6 2 shows an example in which signal values ​​at six times are stored as one unit. For example, the storage unit 151 may store, as a unit, a signal sampled periodically or a signal sampled at a specified time among the time-series signals that are constantly output. This makes it possible to suppress an increase in the storage area for storing signals. The signals in FIG. 2 are merely an example and are not limited to this. For example, signal values ​​at a number of times other than six may be stored as one unit.

[0020] The signal may be of any type, but for example, the following signals may be used: As shown below, the signal is not limited to a time series signal (time series data), and may be of any type (format) of data as long as the wavelet transform can be applied. -Time series data representing vibration data, voice data, operation processes, etc. Image data (still images, video images)

[0021] This embodiment can be applied to, for example, estimation of the state of an object as follows. - Estimation of equipment status (such as abnormalities) using time-series signals detected by sensors (detection devices) from equipment such as bearings

[0022] The signal may be pre-processed, for example by independent component analysis that decomposes the signal into independent components, and by transformation that changes at least one of the mean and variance of the signal. For example, the pre-processing may be a process that shifts the values ​​of the signal so that the mean becomes zero.

[0023] The storage unit 151 can be configured from any commonly used storage medium, such as a flash memory, a memory card, a RAM (Random Access Memory), a HDD (Hard Disk Drive), or an optical disk.

[0024] The display unit 152 is a display device for displaying various information used in the information processing device, and is realized by, for example, a liquid crystal display.

[0025] The reception unit 101 receives input of various information used in the information processing device. For example, the reception unit 101 receives input of a signal. The signal may be transmitted, for example, from a device that is a target of state estimation. The signal may be input sequentially (in real time) or may be input every time a certain period of time elapses. A plurality of signals may be input collectively when designated by a user or the like.

[0026] The output control unit 102 controls the output of various information used in the information processing device. For example, the output control unit 102 controls a process of displaying the estimation result by the estimation unit 122 on the display unit 152.

[0027] The generating unit 111 generates a generating function for generating a local waveform used in a continuous wavelet transform. The local waveform is, for example, an analyzing wavelet (also called a mother wavelet). The generating function is a function in the Fourier domain (frequency domain) at least a part of which is represented by a machine learning model. The machine learning model is, for example, a neural network model (hereinafter, sometimes simply referred to as a neural network) defined in the Fourier domain. In the neural network, parameters such as weights and biases are updated by learning.

[0028] The machine learning model is not limited to a neural network, and may be any model that is constructed by adjusting (updating) parameters. For example, the machine learning model may be a model that integrates multiple wavelets based on parameters.

[0029] FIG. 3 is a diagram showing a configuration example of a generating function. As shown in FIG. 3, the generating function includes a neural network 310 and a function 320. The neural network 310 includes an input layer 311, fully connected layers 312 to 314 as intermediate layers, and an output layer 315. The input layer 311 inputs, for example, a frequency component k expressed in one dimension. The output layer 315 outputs, for example, a one-dimensional output value. Note that FIG. 3 shows an example in which the generating function is a real function. The generating function may be a complex function. In the case where the generating function is a complex function, the neural network 310 may input a frequency component k expressed in one-dimensional real numbers and output a two-dimensional output value including a real part and an imaginary part.

[0030] A function obtained by multiplying the output value of the output layer of the neural network 310 by the function 320 is generated as the generating function 330. The generating function 330 satisfies the permissible condition that should be satisfied as an analyzing wavelet. The permissible condition is, for example, that for a function ψ, a function ψ^ (a function with a hat symbol added above ψ) expressed in the Fourier domain of the function ψ satisfies the following formula (1).

number

[0031] An example of a generating function including a neural network that satisfies this tolerance condition is the generating function expressed by the following equation (2).

number

[0032] g(k) represents a neural network that takes a one-dimensional frequency component k as input and outputs a one-dimensional output value. The function kexp(-k 22. By multiplying by function 320, the generating function becomes 0 at the origin and converges to 0 at a sufficient speed in the distance. In other words, the generating function outputs 0 when the input k is 0, and the output converges to 0 at a speed equal to or greater than a predetermined threshold as the input k approaches infinity. Function 320 can also be interpreted as a damping term that damps the output of the generating function as the input approaches infinity.

[0033] The neural network g(k) may be a multi-layer perceptron including an affine transformation and an activation function, or a neural network including a mechanism using skip connections.

[0034] When the activation function is a linear unit or a rectified linear unit (ReLU), a generating function expressed by the following equation (3) may be used.

number

[0035] When the activation function is a logistic sigmoid function or a hyperbolic tangent function (tanh function), a generating function expressed by the following equation (4) may be used.

number

[0036] In equations (3) and (4), α is a value that satisfies α > 0. Equations (3) and (4) can be interpreted as an example of weakening the damping term from exponential damping to polynomial damping.

[0037] The reason for generating the generating function by multiplying a function containing k as in equations (2) to (4) is to introduce a term that is 0 at the origin. As a method for making the origin 0, it is also possible to remove all biases from the neural network parameters (for example, set all biases to 0) and use the ReLU or tanh function as the activation function.

[0038] In order to restrict the support of the generating function in the Fourier domain to be positive, the generating function expressed by the following equation (5) may be used. The generating function in equation (5) is a function that outputs 0 when the input is negative.

number

[0039] Since it is possible to apply any of several types of generating functions, as shown in equations (2) to (5), it is possible to construct a wavelet that is more specialized (suitable) for the task (target of estimation).

[0040] The learning unit 112 learns the generating function generated by the generating unit 111. For example, the learning unit 112 learns the generating function using learning data including an input signal (first input signal) to be analyzed.

[0041] The learning unit 112 first obtains the frequency components of the input signal f(x). The method of calculating the frequency components may be any conventional method. For example, the fast Fourier transform, the Blackman-Turkey method, the maximum entropy method, etc. may be applied. In addition, when extracting the frequency components by the fast Fourier transform, the learning unit 112 may apply a window function used in spectrum analysis, such as a Hanning window, a Hamming window, or a Gaussian window, to the signal in advance so that an unnecessary spectrum is not included.

[0042] The learning unit 112 uses the frequency component k and the generating function to generate a continuous wavelet representation T wav Calculate f.

[0043] The learning unit 112 calculates the product (multiplication) of the frequency component k in the Fourier domain and the generating function as shown in the following equation (6), and converts the calculated product into an inverse Fourier transform signal (first output signal) into a continuous wavelet representation T wav For example, F[f(x)](k) corresponds to the frequency component k in the Fourier domain, and the term with a bar in ψ^(ak) corresponds to the generating function in the Fourier domain. The bar indicates the complex conjugate. F and F -1 represents the Fourier transform and the inverse Fourier transform. a represents the scale and b represents the time.

number

[0044] The learning unit 112 learns the continuous wavelet representation T wav The generating function is learned by optimizing the value of a loss function based on f. In this embodiment, the learning unit 112 learns the generating function by unsupervised learning.

[0045] First, the learning unit 112 calculates a continuous wavelet representation T wav Calculate the feature from f. The feature is calculated as a real number so that it is differentiable when updating the parameters. The feature is expressed as a continuous wavelet representation T wav It is calculated by a functional with f as an input. The functional is, for example, the following function. Continuous wavelet representation T wav A function that outputs the entropy of f as the output value. Continuous wavelet representation T wav The input signal is estimated from f and the continuous wavelet representation T wav A function that outputs the reconstruction error, which indicates the difference from the input signal used to calculate f, as its output value.

[0046] The functional can be determined according to, for example, the target (task) of estimation, which makes it possible to perform signal analysis using wavelet transform with higher accuracy.

[0047] As described above, the feature amount is, for example, the entropy or the reconstruction error. The entropy is calculated, for example, by the following equation (7). p(a,b) in equation (7) is the absolute value of the continuous wavelet representation |T wav f(a,b)| is a normalized value, and is expressed, for example, by the following equation (8).

number

number

[0048] The reconstruction error is calculated, for example, by the following equation (9).

number

[0049] |||| represents the norm. The norm may be any norm, such as the L1 norm and the L2 norm. rec represents the reconstructed signal and is calculated as follows: Continuous wavelet representation T ψ1 wav The wavelet ψ used to obtain f(a,b) 1 In addition to the above, there is a function ψ defined in the time domain (rather than the Fourier domain) that satisfies the condition expressed by the following equation (10). 2 In addition, the function ψ 2 can be an existing function or can be constructed with a neural network.

number

number

[0050] The learning unit 112 obtains a loss function including the calculated feature amount and a regularization term. The regularization term is, for example, a regularization term of a generating function. When the generating function is expressed by a function ψ, the regularization term is calculated, for example, by the following formulas (12) to (14). Formulas (12), (13), and (14) are the L1 norm of the function ψ, the L2 norm of the function ψ, and the L∞ norm of the function ψ, respectively.

number

number

number

[0051] The function ψ is 1 ,···,x N ) is required. Also, Δx n represents the width between each point.

[0052] The learning unit 112 updates the parameters of the generating function so as to reduce the value of the loss function. The updating method may be any conventionally used method, for example, the backpropagation method. In the backpropagation method, the gradient with respect to the parameters of the loss function is obtained, and the parameter value is updated in the gradient direction. The optimization method used when updating the parameters may be any conventionally used optimization method such as the stochastic gradient descent method, RMSProp, and ADAM.

[0053] The learned generating function can be used for estimation using a newly input input signal (second input signal) to estimate the state of the target. The estimation is mainly performed by the calculation unit 121 and the estimation unit 122 described below. In addition, setting values ​​and the like used in the estimation may be specified by the user or the like.

[0054] For example, the receiving unit 101 may receive input of information indicating the scale range and resolution for performing the continuous wavelet transform in addition to the input signal used for estimation. The input signal used for estimation may differ from the input signal used during learning in at least one of the size (time series length) and the sampling rate. Also, for example, when the input signal is a long-term time series signal, it may be divided into short time series signals for use.

[0055] The calculation unit 121 inputs an input signal to the learned generating function, and calculates an output signal (second output signal) by performing a continuous wavelet transform on the input signal. Similarly to the learning unit 112, the calculation unit 121 calculates a continuous wavelet representation T wav Calculate f.

[0056] When the information indicating the scale range and resolution is received by the receiving unit 101, the calculation unit 121 determines the scale a according to the information and calculates the continuous wavelet representation T wav The calculation unit 121 may determine the scale range and resolution based on the magnitude of the frequency component of the applied signal. For example, the calculation unit 121 may increase the resolution as the frequency increases.

[0057] The calculation unit 121, like the learning unit 112, calculates the continuous wavelet representation T wav If the value of f(a,b) is equal to or less than a threshold, the value may be changed to 0. This makes the estimation robust to slight fluctuations in the signal.

[0058] The estimation unit 122 estimates the state of the object by using the output signal (continuous wavelet representation) calculated by the calculation unit 121. For example, the estimation unit 122 calculates the conformance of the input signal, and estimates the state of the object based on a comparison result between the conformance and a threshold (threshold for state determination).

[0059] The degree of compatibility indicates, for example, the degree of compatibility between an input signal and a local waveform (analyzing wavelet) generated by a generating function. The degree of compatibility can be calculated, for example, by the continuous wavelet representation T wav The entropy of f is calculated as Exp(-H), where H is the entropy of f. The fitness may be normalized to a value between 0 and 1.

[0060] For example, when the degree of conformance is smaller than a threshold, the estimation unit 122 estimates that the object is in an abnormal state, and when the degree of conformance is equal to or greater than the threshold, the estimation unit 122 estimates that the object is in a normal state.

[0061] As described above, the output control unit 102 displays the estimation result by the estimation unit 122 on the display unit 152. An example of output of the estimation result by the output control unit 102 will be described later.

[0062] At least a part of each of the above units (reception unit 101, output control unit 102, generation unit 111, learning unit 112, calculation unit 121, and estimation unit 122) may be realized by one processing unit. Each of the above units is realized by, for example, one or more processors. For example, each of the above units may be realized by having a processor such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit) execute a program, that is, by software. Each of the above units may be realized by a processor such as a dedicated IC (Integrated Circuit), that is, by hardware. Each of the above units may be realized by using both software and hardware. When multiple processors are used, each processor may realize one of the units, or two or more of the units.

[0063] Furthermore, the information processing device 100 may be physically configured by one device, or may be physically configured by multiple devices. For example, the information processing device 100 may be constructed on a cloud environment. Furthermore, each unit in the information processing device 100 may be provided in a distributed manner in multiple devices. For example, the information processing device 100 (information processing system) may be configured to include a device (for example, a learning device) having functions required for learning (such as a generation unit 111 and a learning unit 112) and a device (for example, an estimation device) having functions required for estimation (such as a calculation unit 121 and an estimation unit 122).

[0064] Next, a description will be given of a generation function learning process performed by the information processing device 100 according to the first embodiment. Fig. 4 is a flowchart showing an example of the learning process according to the first embodiment.

[0065] The learning unit 112 acquires an input signal for learning and a generating function (step S101). The input signal is, for example, an input signal received by the receiving unit 101 and stored in the storage unit 151. The generating function is generated by the generating unit 111.

[0066] The learning unit 112 calculates the frequency components of the input signal (step S102). For example, the learning unit 112 calculates the frequency components of the input signal by using the fast Fourier transform, the Blackman-Turkey method, the maximum entropy method, or the like.

[0067] The learning unit 112 uses the calculated frequency components and the acquired generating function to generate a continuous wavelet representation T wav f is calculated (step S103). The wavelet expression calculation process in step S103 will be described in detail later.

[0068] The learning unit 112 calculates the continuous wavelet representation T wav Using f, the learning unit 112 calculates the value of the loss function (step S104). For example, the learning unit 112 calculates the loss function value of the continuous wavelet representation T wavThe value of a loss function including a feature such as entropy or reconstruction error calculated using f and a regularization term such as the norm of a generating function is calculated.

[0069] The learning unit 112 updates the parameters of the generating function so as to optimize the value of the loss function (step S105). For example, the learning unit 112 updates the parameters of the generating function so as to reduce the value of the loss function.

[0070] The learning unit 112 determines whether or not to end the learning (step S106). Any method may be used to determine whether the learning is ended, but for example, a method may be applied in which the learning is ended when the change in the value of the loss function becomes equal to or smaller than a threshold (a threshold for the change in the value).

[0071] If it is determined that the learning is not to be ended (step S106: No), the process returns to step S103 and is repeated. In the repetition, the same input signal may be used, or a different input signal may be used.

[0072] If it is determined that the learning is to be ended (step S106: Yes), the learning unit 112 outputs the learned generating function (step S107) and ends the learning process.

[0073] Next, the details of the calculation process of the wavelet representation in step S103 will be described below. Fig. 5 is a flowchart showing an example of the calculation process of the wavelet representation.

[0074] The learning unit 112 determines a scale (step S201). For example, the learning unit 112 determines one of a plurality of predetermined scales a.

[0075] The learning unit 112 converts the generating function with the determined scale a (step S202). For example, the learning unit 112 converts the generating function so that an input signal multiplied by the scale a is used as an input.

[0076] The learning unit 112 calculates the product of the scale-converted generating function and the frequency component of the input signal for each frequency component of the input signal (step S203). The learning unit 112 performs an inverse Fourier transform on the calculated product (step S204). As a result, the continuous wavelet representation T wav f is calculated.

[0077] The learning unit 112 determines whether all scales have been processed (step S205). If all scales have not been processed (step S205: No), the learning unit 112 returns to step S201 to determine the next scale and repeats the process.

[0078] If all scales have been processed (step S205: Yes), the learning unit 112 outputs the continuous wavelet transform calculated for all specified scales (step S206), and ends the calculation process of the wavelet representation.

[0079] Next, a description will be given of the estimation process performed by the information processing device 100 according to the first embodiment. Fig. 6 is a flowchart showing an example of the estimation process according to the first embodiment.

[0080] The calculation unit 121 acquires an input signal for estimation, a list of scales, and a generating function (step S301). The input signal is received, for example, by the reception unit 101. The generating function is a generating function learned by a learning process. The list of scales is information specifying a plurality of scales to be used for estimation. The list of scales may be specified by a user or the like and received by the reception unit 101. As described above, the calculation unit 121 may determine the range and resolution of the scales based on the magnitude of the frequency components of the signal to be applied, and create the list of scales.

[0081] The calculation unit 121 calculates the frequency components of the input signal (step S202). For example, the calculation unit 121 calculates the frequency components of the input signal by using a fast Fourier transform, the Blackman-Turkey method, the maximum entropy method, or the like.

[0082] The calculation unit 121 uses the calculated frequency components and the obtained generating function to generate a continuous wavelet representation T wav The calculation unit 121 calculates f by the same procedure as in FIG. wav It is possible to calculate f. Note that the function of executing the calculation process of the wavelet representation may be configured to be shared between the learning unit 112 and the calculation unit 121.

[0083] The estimation unit 122 estimates the state of the object by using the output signal (continuous wavelet representation) calculated by the calculation unit 121 (step S304). For example, the estimation unit 122 compares the degree of conformance calculated from the continuous wavelet representation with a threshold, and estimates the state of the object based on the comparison result.

[0084] The output control unit 102 displays the estimation result by the estimation unit 122 on, for example, the display unit 152 (step S305), and ends the estimation process.

[0085] 7 and 8 are diagrams showing examples of a display screen of the estimation result displayed on the display unit 152. Fig. 7 is an example of a display screen 700 including the estimation result when it is determined to be normal. As shown in Fig. 7, the display screen 700 includes a wavelet representation 701 and a compatibility display field 702.

[0086] The wavelet representation 701 is expressed by two-dimensional coordinates with the horizontal axis representing time and the vertical axis representing frequency, respectively. Note that the wavelet representation 701 is, for example, a continuous wavelet representation T wav It is a real value of the feature amount calculated from f (a value between 0 and 0.8 in FIG. 7).

[0087] The input signal in the example of Fig. 7 includes a wavelet representation corresponding to a low frequency and a wavelet representation corresponding to a high frequency when the state is normal, for example. Wavelet representations 711 and 712 in Fig. 7 correspond to such low-frequency and high-frequency wavelet representations. That is, Fig. 7 shows an example in which a wavelet representation more suitable for the wavelet representation when the state is normal is obtained.

[0088] When the threshold value for determining the state is set to 0.5, in the example of FIG. 7, the degree of conformance of the target input signal (target data) is 0.85, so that the estimation unit 122 estimates that the state is normal.

[0089] Fig. 8 is an example of a display screen 800 including an estimation result when it is determined that there is an abnormality. As shown in Fig. 8, the display screen 800 includes a wavelet representation 801 and a compatibility display field 802. In the example of Fig. 8, the wavelet representation 801 is different from the wavelet representation 701 of Fig. 7 and includes wavelet representations corresponding to various frequencies. That is, Fig. 8 shows an example in which a wavelet representation that does not match the wavelet representation when the state is normal is obtained.

[0090] When the threshold value for determining the state is set to 0.5, in the example of FIG. 8, the degree of conformance of the target input signal (target data) is 0.4, so that the estimation unit 122 estimates that the state is abnormal.

[0091] The output control unit 102 may display only the compatibility display fields 702 (FIG. 7) and 802 (FIG. 8) corresponding to the estimation results, or may also display the wavelet representations 701 and 801 as in FIG. 7 and FIG. 8. By displaying the wavelet representations, the basis of the estimation results can be visualized.

[0092] The obtained wavelet representation may be used as an input for an AI (Artificial Intelligence) model such as a SVM (Support Vector Machine).The obtained wavelet representation may also be utilized for denoising processing by restoring the original signal.

[0093] In this way, in the information processing device of the first embodiment, a generating function for generating a local waveform used in a continuous wavelet transform is represented by a machine learning model, and the generating function is learned by unsupervised learning. Therefore, for example, a wavelet specialized for a task can be configured without using an existing wavelet. This makes it possible to perform signal analysis using a wavelet transform with higher accuracy.

[0094] Second Embodiment The information processing device of the second embodiment learns a generating function by supervised learning. In the following, an example of learning a generating function that can be used for a classification task of classifying an input signal to which of a plurality of labels (classes) it corresponds will be described. Applicable tasks are not limited to this, and the present invention can also be applied to, for example, a regression task.

[0095] Fig. 9 is a block diagram showing an example of the configuration of an information processing device 100-2 according to the second embodiment. As shown in Fig. 9, the information processing device 100-2 includes a storage unit 151-2, a display unit 152, a reception unit 101, an output control unit 102, a generation unit 111, a learning unit 112-2, a calculation unit 121, and an estimation unit 122-2.

[0096] In the second embodiment, the functions of the storage unit 151-2, the learning unit 112-2, and the estimation unit 122-2 are different from those in the first embodiment. The other configurations and functions are the same as those in the block diagram of the information processing device 100 in the first embodiment shown in FIG. 1, so they are denoted by the same reference numerals and will not be described here.

[0097] The storage unit 151-2 differs from the storage unit 151 of the first embodiment in that it stores learning data including a label indicating a correct answer for use in supervised learning. The learning data is, for example, data in which an input signal is associated with a label, of multiple labels, into which the input signal is classified.

[0098] The learning unit 112-2 learns the generating function by supervised learning using learning data including a label indicating a correct answer. The learning unit 112-2 uses a functional for converting a continuous wavelet representation into a predicted value according to a task. The learning unit 112-2 prepares a functional Ψ expressed by, for example, the following formula (15).

number

[0099] The functional Ψ in equation (15) is a function that converts a continuous wavelet representation into a real number, and includes a term that corresponds to the inner product of the continuous wavelet representation and the aggregation function Φ. The aggregation function Φ is predetermined as a function that aggregates the continuous wavelet representation into a value corresponding to one of a plurality of labels. A plurality of aggregation functions Φ are determined corresponding to each of a plurality of labels to be classified. The aggregation function Φ may be a predetermined function, or may be configured as a machine learning model such as a neural network and constructed by machine learning in advance or simultaneously with the generating function.

[0100] In the case of a classification task, the learning unit 112-2 performs learning using, for example, the softmax cross entropy expressed by the following formula (16) as a loss function. In the following, the signals and labels included in the learning data are respectively represented by f i and y i (i is an integer between 1 and N, and N is an integer between 2 and N).

number

[0101] If the number of labels to be classified in a classification task is L (L is an integer greater than or equal to 2), then the functional Ψ 1, Ψ L Each of the labels corresponds to one of the L labels. Each label is also one-hot vectorized y i In equation (16), p i For example, it is expressed by the following equation (17): The softmax in the equation (17) is expressed by the following equation (18).

number

number

[0102] Note that the loss function that can be used in the classification task is not limited to the above-mentioned softmax cross entropy, and any conventionally used loss function may be used.

[0103] In the case of a regression task, the learning unit 112-2 performs learning using, for example, a squared error expressed by the following equation (19) as a loss function.

number

[0104] The learning unit 112-2 may learn both the generating function and the aggregation function together, or may learn one at a time using an unsupervised learning technique.

[0105] The learning unit 112-2 may perform supervised learning by contrastive learning. In contrastive learning, learning is performed to distance continuous wavelet representations of signals belonging to different labels. In contrastive learning, for example, a contrastive loss E(f i , f j ), or the contrast loss E(f, f + , f - ) is used as the loss function.

number

number

[0106] m is a positive hyperparameter. f in (21) is the reference signal, f + is a signal that belongs to the same class as f, and f - represents a signal that belongs to a different class than f.

[0107] The estimation unit 122-2 differs from the estimation unit 122 of the first embodiment in that it estimates a state using a functional including an aggregation function. For example, the estimation unit 122-2 calculates the value of a functional corresponding to each label including an aggregation function corresponding to each label, and calculates the probability of each label (classification probability) as an estimation result using the calculated value. The probability represents, for example, the reliability that the input signal corresponds to each label. The probability may be calculated so that the sum of the probabilities for each label is 1.

[0108] The overall flow of the learning process (FIG. 4), the wavelet representation calculation process (FIG. 5), and the estimation process (FIG. 6) is the same as that of the first embodiment, and therefore the description will be omitted. In this embodiment, the specific processes of, for example, updating the parameters using the loss function (FIG. 4, step S105) and the estimation using the wavelet representation (FIG. 6, step S304) are different from those of the first embodiment.

[0109] Fig. 10 is a diagram showing an example of a display screen 1000 of this embodiment. The display screen 1000 of this embodiment includes, for example, a wavelet representation 1001 and a probability display field 1002 showing an estimation result. The wavelet representation 1001 may be displayed separately for each frequency corresponding to each label. Fig. 10 shows three aggregation functions 1011-1013 corresponding to three labels 1-3 and three wavelet representations 1021-1023 as examples.

[0110] In this way, the output control unit 102 may further output information indicating the aggregation function. By displaying the aggregation function, the basis of the estimation result can be visualized.

[0111] For example, when a label is specified on the display screen 1000, the output control unit 102 may display an aggregation function and a wavelet representation corresponding to the specified label. In the example of Fig. 10, when label 1 is specified on the display screen 1000, the output control unit 102 may display an aggregation function 1011 and a wavelet representation 1021.

[0112] The information corresponding to the specified labels may be displayed along with the display screen 1000 or may be displayed as a separate screen from the display screen. The output control unit 102 may display a display screen including the aggregation functions and wavelet representations for all the labels.

[0113] In this way, in the information processing device of the second embodiment, even when a generating function is learned by supervised learning, a wavelet specialized for a task can be configured, as in the first embodiment.

[0114] As described above, according to the first and second embodiments, it is possible to perform signal analysis using wavelet transform with higher accuracy.

[0115] Next, a hardware configuration of the information processing device according to the first or second embodiment will be described with reference to Fig. 11. Fig. 11 is an explanatory diagram showing an example of the hardware configuration of the information processing device according to the first or second embodiment.

[0116] The information processing device of the first or second embodiment includes a control device such as a CPU 51, a storage device such as a ROM (Read Only Memory) 52 or a RAM 53, a communication I / F 54 that connects to a network and communicates, and a bus 61 that connects each part.

[0117] The programs executed by the information processing device of the first or second embodiment are provided in a state that they are pre-installed in the ROM 52 or the like.

[0118] The program executed by the information processing device of the first or second embodiment may be configured to be provided as a computer program product by being recorded in an installable or executable format on a computer-readable recording medium such as a CD-ROM (Compact Disk Read Only Memory), a flexible disk (FD), a CD-R (Compact Disk Recordable), or a DVD (Digital Versatile Disk).

[0119] Furthermore, the program executed by the information processing device of the first or second embodiment may be configured to be stored on a computer connected to a network such as the Internet and provided by being downloaded via the network. Also, the program executed by the information processing device of the first or second embodiment may be configured to be provided or distributed via a network such as the Internet.

[0120] The program executed by the information processing device of the first or second embodiment can cause the computer to function as each part of the information processing device described above. In this computer, the CPU 51 can read the program from a computer-readable storage medium onto the main storage device and execute the program.

[0121] A configuration example of the embodiment will be described below. (Configuration example 1) A generating function for generating a local waveform to be used in a continuous wavelet transform is generated, the generating function being at least partially represented by a machine learning model; learning the generating function using training data including a first input signal; Processing section An information processing device comprising: (Configuration example 2) The machine learning model is a neural network model defined in the frequency domain. The information processing device according to configuration example 1. (Configuration example 3) The generating function outputs 0 when the input is 0, and as the input approaches infinity, the output converges to 0 at a rate equal to or greater than a predetermined threshold. The information processing device according to configuration example 1 or 2. (Configuration Example 4) The generating function outputs 0 when the input is negative. The information processing device according to any one of configuration examples 1 to 3. (Configuration Example 5) The processing unit includes: a first output signal obtained by performing a continuous wavelet transform on the first input signal by multiplying the first input signal by an output of the generating function when the first input signal is input in a frequency domain and performing an inverse Fourier transform on the resultant value; and learning the generating function by optimizing a value of a loss function based on the first output signal. The information processing device according to any one of configuration examples 1 to 4. (Configuration Example 6) The processing unit includes: optimizing a value of the loss function including an output value of a functional having the first output signal as an input; The information processing device according to configuration example 5. (Configuration Example 7) The processing unit includes: The generating function is learned by unsupervised learning. The information processing device according to configuration example 6. (Configuration Example 8) The functional outputs the entropy of the first output signal as the output value. The information processing device according to configuration example 7. (Configuration Example 9) The functional outputs, as the output value, a reconstruction error indicating a difference between an input signal estimated from the first output signal and the first input signal. The information processing device according to configuration example 7. (Configuration Example 10) The processing unit includes: The generating function is learned by supervised learning using the learning data including a label indicating a correct answer. The information processing device according to configuration example 6. (Configuration Example 11) The supervised learning is contrastive learning. The information processing device according to configuration example 10. (Configuration Example 12) the functional includes an aggregation function that aggregates the first output signal into a value corresponding to the label; The information processing device according to configuration example 10. (Configuration Example 13) The processing unit includes: calculating a frequency domain value of the first input signal by a fast Fourier transform; performing an inverse Fourier transform on a value obtained by multiplying the first input signal in the frequency domain by an output of the generating function when the first input signal is input; The information processing device according to configuration example 5. (Configuration Example 14) The processing unit includes: performing the fast Fourier transform on the first input signal using a window function including a Hanning window, a Hamming window, and a Gaussian window; The information processing device according to configuration example 13. (Configuration Example 15) The processing unit includes: changing the first output signal, which is less than or equal to a threshold value, to zero; The information processing device according to configuration example 5. (Configuration Example 16) The processing unit includes: A second input signal for estimating a state of an object is input to the learned generating function, and a second output signal is calculated by subjecting the second input signal to a continuous wavelet transform; estimating the state using the second output signal; 16. The information processing device according to any one of configuration examples 1 to 15. (Configuration Example 17) The second input signal differs from the first input signal in at least one of size and sampling rate. The information processing device according to configuration example 16. (Configuration Example 18) The processing unit includes: determining a scale to be used in a continuous wavelet transform based on the frequency components of the second input signal; The information processing device according to configuration example 16. (Configuration Example 19) The processing unit includes: changing the second output signal, which is less than or equal to a threshold value, to zero; The information processing device according to configuration example 16. (Configuration Example 20) The processing unit includes: Output the estimation results, The information processing device according to configuration example 16. (Configuration Example 21) The processing unit includes: estimating the state using a functional including an aggregation function that aggregates the second output signals into a value corresponding to the state; further outputting information indicating the aggregation function; The information processing device according to configuration example 20. (Configuration Example 22) The aggregation function is constructed by machine learning. The information processing device according to configuration example 21. (Configuration Example 23) The processing unit includes: A generating unit that generates the generating function; A learning unit that learns the generating function. The information processing device according to any one of configuration examples 1 to 22. (Configuration Example 24) An information processing method executed by an information processing device, A generating step of generating a generating function for generating a local waveform to be used in a continuous wavelet transform, the generating function being at least partially represented by a machine learning model; a learning step of learning the generating function using learning data including a first input signal; An information processing method comprising: (Configuration Example 25) On the computer, A generating step of generating a generating function for generating a local waveform to be used in a continuous wavelet transform, the generating function being at least partially represented by a machine learning model; a learning step of learning the generating function using learning data including a first input signal; A program for executing the above.

[0122] Although some embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope and spirit of the invention, and are included in the scope of the invention and its equivalents described in the claims. [Explanation of symbols]

[0123] 100, 100-2 Information processing device 101 Reception 102 Output control section 111 Generation part 112, 112-2 Learning Department 121 Calculation section 122, 122-2 Estimation part 151, 151-2 Storage section 152 Display section 310 Neural Networks 320 Functions 330 Generating Functions

Claims

1. A generating function for generating a local waveform to be used in a continuous wavelet transform is generated, the generating function being at least partially represented by a machine learning model; training the generating function using training data including a first input signal; Processing section An information processing device comprising:

2. The machine learning model is a neural network model defined in the frequency domain. The information processing device according to claim 1 .

3. The generating function outputs 0 when the input is 0, and as the input approaches infinity, the output converges to 0 at a rate equal to or greater than a predetermined threshold. The information processing device according to claim 1 .

4. The generating function outputs 0 when the input is negative. The information processing device according to claim 1 .

5. The processing unit includes: a first output signal obtained by performing a continuous wavelet transform on the first input signal by performing an inverse Fourier transform on a value obtained by multiplying the first input signal by an output of the generating function when the first input signal is input in a frequency domain, and learning the generating function by optimizing a value of a loss function based on the first output signal; The information processing device according to claim 1 .

6. The processing unit includes: optimizing a value of the loss function including an output value of a functional having the first output signal as an input; The information processing device according to claim 5 .

7. The processing unit includes: The generating function is learned by unsupervised learning. The information processing device according to claim 6.

8. The functional outputs the entropy of the first output signal as the output value. The information processing device according to claim 7.

9. The functional outputs, as the output value, a reconstruction error indicating a difference between an input signal estimated from the first output signal and the first input signal. The information processing device according to claim 7.

10. The processing unit includes: The generating function is learned by supervised learning using the learning data including a label indicating a correct answer. The information processing device according to claim 6.

11. The supervised learning is contrastive learning. The information processing device according to claim 10.

12. the functional includes an aggregation function that aggregates the first output signal into a value corresponding to the label; The information processing device according to claim 10.

13. The processing unit includes: calculating a frequency domain value of the first input signal by a fast Fourier transform; performing an inverse Fourier transform on a value obtained by multiplying the first input signal in the frequency domain by an output of the generating function when the first input signal is input; The information processing device according to claim 5 .

14. The processing unit includes: performing the fast Fourier transform on the first input signal using a window function including a Hanning window, a Hamming window, and a Gaussian window; The information processing device according to claim 13.

15. The processing unit includes: changing the first output signal, which is less than or equal to a threshold value, to zero; The information processing device according to claim 5 .

16. The processing unit includes: A second input signal for estimating a state of an object is input to the learned generating function, and a second output signal is calculated by subjecting the second input signal to a continuous wavelet transform; estimating the state using the second output signal; The information processing device according to claim 1 .

17. The second input signal differs from the first input signal in at least one of size and sampling rate. The information processing device according to claim 16.

18. The processing unit includes: determining a scale to be used in a continuous wavelet transform based on the frequency components of the second input signal; The information processing device according to claim 16.

19. The processing unit includes: changing the second output signal, which is less than or equal to a threshold value, to zero; The information processing device according to claim 16.

20. The processing unit includes: Output the estimation results, The information processing device according to claim 16.

21. The processing unit includes: estimating the state using a functional including an aggregation function that aggregates the second output signals to a value corresponding to the state; further outputting information indicating the aggregation function; The information processing device according to claim 20.

22. The aggregation function is constructed by machine learning. The information processing device according to claim 21.

23. The processing unit includes: A generating unit that generates the generating function; A learning unit that learns the generating function. The information processing device according to claim 1 .

24. An information processing method executed by an information processing device, A generating step of generating a generating function for generating a local waveform to be used in a continuous wavelet transform, the generating function being at least partially represented by a machine learning model; a learning step of learning the generating function using learning data including a first input signal; An information processing method comprising:

25. On the computer, A generating step of generating a generating function for generating a local waveform to be used in a continuous wavelet transform, the generating function being at least partially represented by a machine learning model; a learning step of learning the generating function using learning data including a first input signal; A program for executing.

Citation Information

Patent Citations

  • System and method for obtaining optimal mother wavelets for facilitating machine learning tasks

    JP2019121376A

  • Method for determining an optimally weighted wavelet transform based on supervised training for detection of microcalcifications in digital mammograms

    US6075878A

  • Road surface state estimator, road surface friction state estimator, road surface state physical quantity arithmetic unit and road surface state annunciator

    JP2004168286A

  • Apparatus and method for analyzing operation result of manufacturing process, and computer-readable storage medium

    JP2004288144A

  • How to perform predictive maintenance on mobile equipment

    JP7042315B2