Data augmentation method

The data expansion method addresses the bias in real-world vibration data by applying orthogonal forces and filters to a computer-modeled structure, effectively generating training data that improves the generalization performance of statistical machine learning models.

WO2024236693A1PCT designated stage expired Publication Date: 2024-11-21NT T INC
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2023/018118
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-15
Publication Date
2024-11-21

AI Technical Summary

Technical Problem

Vibration data from structures measured in real-world scenarios often exhibit significant bias and do not cover the entire potential data distribution, leading to poor generalization performance when used as training data for statistical machine learning methods.

Method used

A data expansion method that applies forces from three orthogonal directions to a computer-modeled rectangular parallelepiped structure, calculates frequency response spectra, and multiplies these by different weights or applies filters to generate pseudo vibration data that mimics actual measurement distributions, ensuring unbiased sampling for training data.

Benefits of technology

This approach efficiently generates vibration data with a distribution similar to actual measurements at low computational cost, improving the quality and accuracy of training data for statistical machine learning applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2023018118_21112024_PF_FP_ABST
    Figure JP2023018118_21112024_PF_FP_ABST
Patent Text Reader

Abstract

The present invention provides a data augmentation method for augmenting vibration data for a structure, wherein a data augmentation device applies forces from three directions orthogonal to a model structure of a rectangular parallelepiped modeled in a computer, and performs a first step (S2) of calculating a frequency response spectrum in each direction corresponding to the vibration due to each force and a second step (S3) of multiplying the frequency response spectrum in each direction by mutually different weights and adding up the multiplied values.
Need to check novelty before this filing date? Find Prior Art

Description

Data Augmentation Methods

[0001] The present disclosure relates to a data augmentation method.

[0002] There is a technology that extracts frequency components from time-series data that is continuously acquired of vibration data of a structure, and uses the frequency components to diagnose the condition of the structure. This technology is used in ultrasonic non-destructive testing, etc. As the strength of a structure changes due to fatigue or partial destruction, specific frequency components that correspond to the natural frequency specific to the structure shift to the higher or lower frequency side, and the condition of the structure is diagnosed based on the detection of this frequency shift. Patent Document 1 discloses a method for evaluating the degree of deterioration of a structure based on vibration data acquired by striking the structure.

[0003] Japanese Patent Application Laid-Open No. 2006-250682

[0004] We will consider a method for converting vibration data of a structure into frequency domain information and estimating the state of the structure based on the probability distribution of natural frequency peaks.When applying statistical machine learning techniques to this method, it is desirable that the vibration data used as training data be sampled randomly and without bias from the vibration data distribution that could potentially occur.

[0005] However, the vibration data of structures obtained through actual measurements has a large distribution bias within the latent vibration data distribution and does not cover the entire latent vibration data distribution. Therefore, when using only measured vibration data as training data for statistical machine learning methods, there is a problem in that high generalization performance cannot be expected.

[0006] The present disclosure has been made in consideration of the above circumstances, and an object of the present disclosure is to provide a technology for efficiently generating vibration data having a distribution close to that of vibration data that can be actually measured, from vibration data of a structure modeled in a computer.

[0007] A data expansion method according to one aspect of the present disclosure is a data expansion method for expanding vibration data of a structure, in which a data expansion device performs a first step of applying forces from three orthogonal directions to a rectangular parallelepiped model structure modeled in a computer and calculating a frequency response spectrum for each direction corresponding to the vibration caused by each force, and a second step of multiplying the frequency response spectrum for each direction by a different weight and adding up the multiplied values.

[0008] According to the present disclosure, it is possible to provide a technology for efficiently generating vibration data having a distribution close to that of vibration data that can be actually measured, from vibration data of a structure modeled in a computer.

[0009] Fig. 1 is a diagram showing the configuration of a data extension device. Fig. 2 is a diagram showing a data extension method. Fig. 3 is a diagram showing an example of a model structure. Fig. 4 is a diagram showing an example of a frequency response spectrum. Fig. 5 is a diagram showing data extension method 1. Fig. 6 is a diagram showing data extension method 2. Fig. 7 is a diagram showing data extension method 3. Fig. 8 is a diagram showing the hardware configuration of a data extension device.

[0010] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. In the description of the drawings, the same parts are designated by the same reference numerals and the description thereof will be omitted.

[0011] [Summary of the Disclosure] The present disclosure relates to a technology for estimating abnormalities or changes in a structure from vibration data of the structure.

[0012] In order to solve the above problem, a method of generating simulated vibration data under various conditions from vibration data of a structure modeled in a computer can be considered.

[0013] When creating training data for statistical machine learning methods from simulation data, which is vibration data of a structure, one possible method is to multiply each natural frequency specific to the structure by a random weight and combine them to generate a pseudo vibration spectrum. However, this method deviates from the sample distribution of vibration spectra that can actually be measured, and is therefore not suitable as training data for statistical machine learning methods.

[0014] It is generally understood that high-quality training data is important for improving the generalization accuracy of statistical machine learning. High-quality training data is data that has been randomly sampled without bias from the range of data distribution that may be obtained during estimation. To prepare such training data, the present disclosure focuses on the following two points.

[0015] (1) The measured intensity of the resonant frequency corresponding to the resonant mode of a structure depends on the direction and magnitude of the force actually applied to the structure and the frequency components contained in the applied force. In other words, when a force biased in a specific direction is applied, the resonant mode frequency corresponding to the vibration in that direction is measured strongly.

[0016] (2) Artificial vibrations tend to be dominated by high frequencies, while natural vibrations tend to be dominated by low frequencies. The frequency spectrum must reflect this.

[0017] The present disclosure focuses on the above two points and provides a data augmentation method that does not require computational costs.

[0018] [Configuration of the Data Expansion Device] Fig. 1 is a diagram showing the configuration of a data expansion device according to this embodiment. The data expansion device 1 is a device for expanding vibration data of a structure, and includes a first calculation unit 11, a second calculation unit 12, a third calculation unit 13, and a storage unit 14.

[0019] The first calculation unit 11 has a function of calculating information about a rectangular parallelepiped that encloses the general shape of a model structure modeled in a computer. Specifically, the first calculation unit 11 calculates the lengths of the three sides (length, width, and depth) of a rectangular parallelepiped model structure modeled by a modeling software program, as viewed from each of three orthogonal axes (X-axis, Y-axis, and Z-axis).

[0020] The second calculation unit 12 has a function of calculating vibration data of the structure. Specifically, the second calculation unit 12 applies forces to a rectangular parallelepiped model structure modeled in a computer from three orthogonal directions (X direction, Y direction, and Z direction), and calculates frequency response spectra in each direction corresponding to vibrations caused by each force.

[0021] The third calculation unit 13 has a function of expanding the vibration data of the structure.

[0022] In data extension method 1, the third calculation unit 13 multiplies the frequency response spectrum for each direction calculated by the second calculation unit 12 by different weights and adds up the multiplied values. The weight is, for example, a product of a weighting coefficient that multiplies the spectrum intensity by a constant and a random number that perturbs the spectrum.

[0023] As data extension method 2, the third calculation unit 13 performs a frequency domain modulation operation on the frequency response spectrum for each direction calculated by the second calculation unit 12, by multiplying the high frequency region by a large value and the low frequency region by a small value, or by multiplying the high frequency region by a small value and the low frequency region by a large value, and then adds up the calculated values.

[0024] As the data extension method 3, the third calculation unit 13 performs calculations by combining the data extension methods 1 and 2.

[0025] The storage unit 14 has a function of storing various data necessary for calculating vibration data.

[0026] [Data Extension Method] FIG. 2 is a diagram showing a data extension method.

[0027] First, the first calculation unit 11 calculates the lengths of the three sides (length, width, and depth) of a rectangular parallelepiped model structure modeled in a computer (step S1). Next, the second calculation unit 12 applies forces to the rectangular parallelepiped model structure in the X, Y, and Z directions, respectively, and calculates frequency response spectra for each direction corresponding to the vibrations caused by each force (step S2). Finally, the third calculation unit 13 extends the frequency response spectra for each direction calculated in step S2 by using one of data extension methods 1 to 3 (step S3).

[0028] [Data Augmentation Method 1] Data augmentation method 1 is a method in which the direction of the resonance mode is taken into consideration and different weights are applied to the frequency response spectrum in each direction.

[0029] When performing stress analysis on a model structure modeled in a computer, unless the model structure is a perfect sphere or tetrahedron, if it is properly modeled, the model structure will have a long axis and a short axis in an orthogonal XYZ coordinate system (Figure 3).

[0030] When natural vibration analysis (mode analysis) is performed on this model structure 100, a group of natural vibration modes having fluctuation components corresponding to the major axis and minor axis directions of the model structure 100 is generally derived.

[0031] The long axis and short axis directions of this model structure 100 are temporarily called the X direction and Y direction. When the model structure 100 is in a stationary position, for example, one eigenmode having a fluctuation component corresponding to the Y direction is expressed as a straight line marked with the letter M.

[0032] The second calculation unit 12 performs frequency response analysis for the application of forces in the X, Y, and Z directions, thereby calculating the frequency response spectra for each of the X, Y, and Z directions (Fig. 4). The eigenmode M in the Y direction is expressed as a curve marked M' when a force is applied in the Y direction.

[0033] The third calculation unit 13 prepares weighting coefficients Wx, Wy, and Wz (scalar quantities) and three sets of random number vectors Rndx, Rndy, and Rndz with randomness for the frequency response spectra Hx, Hy, and Hz (vector quantities) in the X, Y, and Z directions, respectively.

[0034] Then, the third calculation unit 13 calculates Hgen using equation (1) and outputs Hgen as the expanded frequency response spectrum (FIG. 5).

[0035] Hgen = Hx x Wx x Rndx + Hy x Wy x Rndy + Hz x Wz x Rndz (1) The values ​​of Wx, Wy, and Wz may be different from each other or the same. The values ​​of Rndx, Rndy, and Rndz may also be different from each other or the same. It is sufficient that "Wx x Rndx", "Wy x Rndy", and "Wz x Rndz" are different from each other.

[0036] [Data Extension Method 2] Data extension method 2 is a method of applying a high-pass filter or a low-pass filter with an appropriate threshold value to the frequency response analysis spectrum.

[0037] Artificial vibrations can be simulated by applying a high-pass filter to generate collision-like stimuli. Natural vibrations can be simulated by applying a low-pass filter to generate long-period vibrations. Data augmentation method 2 takes these points into consideration.

[0038] The third calculation unit 13 prepares a high-pass filter Hpass(cutoff) or a low-pass filter Lpass(cutoff) having randomness (Rndcutoff) and an appropriate cutoff frequency (cutoff) for the frequency response spectra Hx, Hy, Hz (vector quantities) in the X, Y, and Z directions of the model structure, respectively.

[0039] Then, the third calculation unit 13 calculates Hgen using equation (2) or equation (3) and outputs the calculated Hgen as the expanded frequency response spectrum (FIG. 6).

[0040] Hgen=Hx×Hpass(cutoff * Rndcutoff)…(2) Hgen=Hx×Lpass(cutoff * Rndcutoff)…(3)

[0041] [Data Extension Method 3] Data extension method 3 is a method in which data extension methods 1 and 2 are used in combination.

[0042] The third calculation unit 13 calculates Hgen using equation (4) or equation (5) and outputs Hgen as the expanded frequency response spectrum (FIG. 7).

[0043] Hgen=Hx×Hpass(cutoff * Rndcutoff)×Wx×Rndx+Hy×Hpass(cutoff * Rndcutoff)×Wy×Rndy+Hz×Hpass(cutoff * Rndcutoff)×Wz×Rndz…(4) Hgen=Hx×Lpass(cutoff * Rndcutoff)×Wx×Rndx+Hy×Lpass(cutoff * Rndcutoff) × Wy × Rndy + Hz × Lpass (cutoff * Rndcutoff) × Wz × Rndz … (5)

[0044] [Effects] Conventional methods require frequency response analysis from various angles and transient response analysis under various conditions to reproduce artificial phenomena, which must be performed as many times as the number of data points required, resulting in high computational costs.

[0045] On the other hand, according to this embodiment, data having a distribution close to that of data that can be actually measured can be generated at low calculation cost from vibration data of a model structure modeled in a computer.

[0046] In data expansion method 1, frequency response analysis is performed on a model structure modeled in a computer from three orthogonal directions, and the results are simply multiplied by different weights (specifically, random number vectors with different randomness) and linearly combined, thereby efficiently generating vibration data that is likely to occur in reality.

[0047] In data augmentation method 2, frequency response analysis is performed on a model structure modeled in a computer from three orthogonal directions, and a high-pass filter or low-pass filter with an appropriate threshold value is applied to the results, making it possible to efficiently generate vibration data that is likely to occur in reality due to artificial or natural phenomena.

[0048] Data extension method 3 provides the combined effects of data extension methods 1 and 2.

[0049] [Others] The present disclosure is not limited to the above-described embodiments, and various modifications are possible within the scope of the gist of the present disclosure.

[0050] The data expansion device 1 of the present embodiment described above can be realized, for example, by using a general-purpose computer system including a CPU 901, a memory 902, a storage 903, a communication device 904, an input device 905, and an output device 906, as shown in Fig. 8. The memory 902 and the storage 903 are storage devices. In this computer system, the CPU 901 executes a predetermined program loaded on the memory 902, thereby realizing each function of the data expansion device 1.

[0051] The data expansion device 1 may be implemented by one computer, or by multiple computers, or may be a virtual machine implemented on a computer.

[0052] The program for the data extension device 1 can be stored in a computer-readable recording medium such as a HDD, SSD, USB memory, CD, or DVD. The computer-readable recording medium is, for example, a non-transitory recording medium. The program for the data extension device 1 can also be distributed via a communication network.

[0053] REFERENCE SIGNS LIST 1 Data expansion device 11 First calculation unit 12 Second calculation unit 13 Third calculation unit 14 Storage unit 100 Model structure 901 CPU 902 Memory 903 Storage 904 Communication device 905 Input device 906 Output device

Claims

1. A data expansion method for expanding vibration data of a structure, comprising: a first step in which a data expansion device applies forces from three orthogonal directions to a rectangular parallelepiped model structure modeled in a computer, and calculates a frequency response spectrum for each direction corresponding to the vibration caused by each force; and a second step in which the frequency response spectrum for each direction is multiplied by a different weight and each multiplied value is added together.

2. The data extension method of claim 1, wherein in the second step, a frequency domain modulation operation is further performed on the frequency response spectrum in each direction, in which a large value is multiplied in the high frequency region and a small value is multiplied in the low frequency region, or a frequency domain modulation operation is further performed on the frequency response spectrum in each direction, in which a small value is multiplied in the high frequency region and a large value is multiplied in the low frequency region.

3. The data extension method according to claim 1, wherein the weight is a product of a weighting coefficient that multiplies the spectral intensity by a constant and a perturbing random number.

4. A data expansion method for expanding vibration data of a structure, comprising: a first step in which a data expansion device applies forces from three orthogonal directions to a rectangular parallelepiped model structure modeled in a computer, and calculates a frequency response spectrum in each direction corresponding to the vibration caused by each force; and a second step in which a frequency domain modulation operation is performed on the frequency response spectrum in each direction, multiplying a large value in the high frequency region and a small value in the low frequency region, or multiplying a small value in the high frequency region and a large value in the low frequency region, and adding up the calculated values.

Citation Information

Patent Citations

  • Social network feature dynamic extraction method based on vector compression and reconstruction

    CN112052940A

  • Method and device for learning neural network

    JP1993298277A

  • Internal state estimation system

    JP2020085631A

  • Information processor, control method of information processor and program

    JP2020106340A

  • Image processing device, image processing method, learning device, learning method, and program

    JP2022039989A