CN114692831A - Method of calculating variable convolution kernel for variable resolution, storage medium - Google Patents

Method of calculating variable convolution kernel for variable resolution, storage medium Download PDF

Info

Publication number
CN114692831A
CN114692831A CN202210304788.XA CN202210304788A CN114692831A CN 114692831 A CN114692831 A CN 114692831A CN 202210304788 A CN202210304788 A CN 202210304788A CN 114692831 A CN114692831 A CN 114692831A
Authority
CN
China
Prior art keywords
time
frequency
window
domain width
variable
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210304788.XA
Other languages
Chinese (zh)
Inventor
丁煦
王杨
郑航
翟华
赵冲
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hefei University of Technology
Original Assignee
Hefei University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hefei University of Technology filed Critical Hefei University of Technology
Priority to CN202210304788.XA priority Critical patent/CN114692831A/en
Publication of CN114692831A publication Critical patent/CN114692831A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
    • G06F17/141—Discrete Fourier transforms
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
    • G06F17/148—Wavelet transforms
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Algebra (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Databases & Information Systems (AREA)
  • Discrete Mathematics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Complex Calculations (AREA)

Abstract

The invention discloses a method for calculating a variable convolution kernel aiming at variable resolution, wherein the time resolution and the frequency resolution of each window function in wavelet transformation are different, the time domain width and the frequency domain width of each window in a time-frequency image of wavelet transformation are different, and the size of the convolution kernel convolved with each window in the time-frequency image is respectively determined according to the time domain width and the frequency domain width of each window in the time-frequency image. The method designs the corresponding convolution kernel size according to the time domain width and the frequency domain width of each window in the time-frequency image, solves the problem that objects with different scales or deformations at different positions are difficult to be convolved by the convolution kernels with the same size, and can diagnose and classify the bearing faults in a shorter time, thereby obtaining higher precision.

Description

Method of calculating variable convolution kernel for variable resolution, storage medium
Technical Field
The invention relates to the technical field of convolutional neural networks, in particular to a method and a storage medium for calculating a variable convolution kernel according to variable resolution.
Background
The bearing is one of the core components of the mechanical system and is important for the efficient, stable and reliable operation of the mechanical system. Many methods for classifying bearing faults have been developed, and the conventional methods for diagnosing and classifying bearing faults can be divided into three major categories: based on signal processing, machine learning, and deep learning. The signal processing based method requires more prior knowledge and cannot realize a higher precision and automatic diagnosis process. Machine learning-based bearing fault classification methods require more correlation techniques to extract sensitive features. The bearing fault classification based on deep learning can automatically identify representative characteristics of bearing faults from original data, reduces the dependence on technologies such as fault characteristic extraction and the like, and becomes a mainstream technology for bearing fault diagnosis.
Meanwhile, the bearing fault diagnosis classification method based on deep learning can also be divided into three categories: respectively, a time domain based analysis method, a frequency domain based analysis method, and a time-frequency domain based analysis method. The time domain based analysis method is completely localized in the time domain range and does not contain frequency domain information. The frequency domain based analysis method is fully localized in the frequency domain range and does not contain time domain information. The time-frequency signals are analyzed based on a time-frequency domain analysis method, the time-frequency signals have both time domain information and frequency domain information, and the signals subjected to wavelet transformation have the requirement of automatically adapting to time-frequency analysis, so that the selection of the time-frequency signals subjected to wavelet transformation as convolutional neural network input becomes a trend.
However, wavelet transformation is combined with a convolutional neural network, and the processed time-frequency image is input into a model of the convolutional neural network. If the receptive field size of the activation cells in the same convolutional neural network layer is the same (the receptive field refers to the area in the input space that affects a particular cell of the network), it is not preferable for the higher convolutional neural network layer to be in spatial position. It is difficult to convolve with convolution kernels of the same size, since there are objects of different scales or deformations at different locations.
Disclosure of Invention
In order to overcome the defects in the prior art, the invention provides a method for calculating a variable convolution kernel aiming at variable resolution, wherein the variable convolution kernel is designed according to the time domain width and the frequency domain width of each window in a time-frequency image, and the method is used for solving the problem that objects with different scales or deformations at different positions are difficult to be convolved by the convolution kernel with the same size.
In order to achieve the purpose, the invention adopts the following technical scheme that:
a method of computing a variable convolution kernel for variable resolution, comprising the steps of;
s1, performing wavelet transformation on the signal to obtain a time-frequency image of the wavelet transformation; wherein, the time resolution and the frequency resolution of each window function in the wavelet transformation are different, and the time domain width delta (psi) of each window in the time-frequency image(a,b)) And frequency domain width
Figure BDA0003560249810000021
All are different;
s2, calculating the time domain width delta (psi) of each window in the time-frequency image(a,b)) And frequency domain width
Figure BDA0003560249810000022
S3, according to the time domain width delta phi (psi) of each window in the time frequency image(a,b)) And frequency domain width
Figure BDA0003560249810000023
Respectively determining the sizes of convolution kernels convolved with windows in the time-frequency image;
the convolution kernel width for performing convolution with a certain window in the time-frequency image is as follows: the time domain width delta (/) of the window(a,b)) Rounding to the rounded value;
the height of a convolution kernel convolved with a certain window in a time-frequency image is as follows:the frequency domain width of the window
Figure BDA0003560249810000024
And rounding to the rounded value.
In step S2, the temporal width Δ (ψ) of the window in the time-frequency image(a,b)) And frequency domain width
Figure BDA0003560249810000025
The calculation of (c) is as follows:
Δ(ψ(a,b))＝|a|Δ(ψ)
Figure BDA0003560249810000026
wherein, a is the scale of wavelet transformation, and b is the displacement of wavelet transformation; a is not equal to 0, and b is any real number;
ψ represents a wavelet transform function when a is 1 and b is 0;
Figure BDA0003560249810000027
in the form of psi after fourier transformation; the time domain width of the wavelet transform with a being 1 and b being 0 is delta (psi);
Figure BDA0003560249810000028
the frequency domain width of the wavelet transform with a-1 and b-0;
ψ(a,b)a wavelet transformation function with the scale a and the displacement b is represented;
Figure BDA0003560249810000029
is psi(a,b)A form after Fourier transform; delta (psi)(a,b)) The time domain width of the wavelet transform with the scale a and the displacement b is obtained;
Figure BDA00035602498100000210
the width of the frequency domain of the wavelet transform with the scale a and the displacement b;
due to the delta (psi) of the window in the time-frequency image(a,b)) And frequency domain width
Figure BDA00035602498100000211
The product of (a) is a constant value, i.e.
Figure BDA0003560249810000031
Figure BDA0003560249810000032
The calculation mode of the scale a of the wavelet transformation is as follows:
Figure BDA0003560249810000033
wherein, FcIs the wavelet center frequency, TsTo sample time, FaIs the actual frequency;
and Δ (ψ) and
Figure BDA0003560249810000034
the calculation method is as follows:
Figure BDA0003560249810000035
Figure BDA0003560249810000036
wherein, w0Gamma is a constant and is a positive value; t is a time variable; w is a frequency domain variable; i represents an imaginary symbol;
thus, the time domain width Δ (ψ) of each window of the time-frequency image is found(a,b)) And frequency domain width
Figure BDA0003560249810000037
The value of (c).
In step S3, in the deep layer of the convolutional neural network, the tensor convolved by the convolution kernel is convolved with the ResNet module.
The signal is a bearing fault signal, and the bearing fault signal is subjected to wavelet transformation to obtain a time-frequency image of the wavelet transformation of the bearing fault signal; carrying out characteristic identification on the time-frequency image of the wavelet transform of the bearing fault signal by utilizing a convolutional neural network so as to classify the bearing fault; the convolution kernel in the convolution neural network is a variable convolution kernel, namely the convolution kernel has different sizes with each window in the time-frequency image; and the convolution kernel size convolved with each window in the time-frequency image is solved by adopting the method of the steps S1-S3.
The invention also provides a storage medium for calculating a variable convolution kernel for variable resolution, the storage medium storing a computer program comprising program instructions which, when executed by a processor, cause the processor to carry out the method of claim 1 or 2 or 3.
The invention has the advantages that:
(1) the algorithm can design the corresponding convolution kernel size according to the time domain width and the frequency domain width of each window in the time-frequency image, solves the problem that objects with different scales or deformations at different positions are difficult to be convolved by the convolution kernels with the same size, and enables bearing fault classification diagnosis to be carried out in a shorter time to obtain higher precision.
(2) In order to solve the problem that objects with different scales or deformations at different positions are difficult to be convolved by convolution kernels with the same size, the prior art mainly improves the design of convolution shapes and weights and the like. However, since the conventional design of convolution shape and weight is based on image, the variable convolution kernel is not designed in consideration of time-frequency image of wavelet transform. However, the variable convolution kernel is designed in the aspect of time-frequency images of wavelet transformation, and particularly, the variable convolution kernel is designed according to the time-domain width and the frequency-domain width of each window in the time-frequency images, so that bearing fault diagnosis and classification can be carried out in a shorter time, and higher precision is obtained.
(3) According to the invention, a ResNet related module is embedded in a designed variable convolution kernel method to extract texture and detail information of bearing faults, a larger variable convolution kernel is used in a shallow layer to extract feature change of most original texture details, and then a ResNet module is used in a deep layer to convolve tensors of the bearing faults after being convolved by the variable convolution kernel, so that the phenomenon that feature correlation in a larger local range is lost is avoided, and satisfactory bearing fault diagnosis classification prediction performance is obtained.
Drawings
FIG. 1 is a flow chart of a method for designing a variable convolution kernel according to the present invention.
Fig. 2 is a schematic diagram of a time-frequency image window of wavelet transform.
FIG. 3 is a diagram of variable convolution kernels of different sizes designed for a time-frequency image.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
In this embodiment, the UT6818 mechanical vibration analysis and fault simulation experiment table is used to collect bearing fault diagnosis classification data. The test bed mainly comprises an engine, a belt, a bearing seat, an acceleration sensor and a bearing part. The acceleration sensor is installed on the bearing seat, and the speed of the bearing is controlled by the three-phase motor through the elastic coupling.
Data set collection was performed using UT6818 mechanical vibration analysis and fault simulation bench. The bearing vibration signals are collected by connecting an acceleration sensor with a PC end, and the vibration signals are collected at the sampling frequency of 51200 Hz. Seven conditions of bearing failure were collected, including one bearing healthy no failure condition and six bearing crack failure conditions.
The method for calculating the variable convolution kernel for variable resolution is specifically as follows:
and S1, extracting the signals of the bearing vibration signals, and performing wavelet transformation on the extracted bearing fault signals to obtain a time-frequency image of the wavelet transformation of the bearing fault signals. In this embodiment, Gabor transform is performed on the bearing fault time domain signal to obtain a time-frequency image.
As shown in fig. 2, in a time-frequency image, the vertical axis is frequency, the horizontal axis is time, as the frequency increases, the low frequency corresponds to a larger time window, the high frequency corresponds to a larger frequency window, and the width and height of the variable convolution kernel correspond to the time resolution and the frequency resolution of the window function in the wavelet transform, so that when calculating the variable convolution kernel for the variable resolution, the time domain width Δ (ψ) of each window in the time-frequency image needs to be calculated(a,b)) And frequency domain width
Figure BDA0003560249810000051
S2, calculating the time domain width delta (psi) of each window in the time-frequency image(a,b)) And frequency domain width
Figure BDA0003560249810000052
Δ(ψ(a,b))＝|a|Δ(ψ)
Figure BDA0003560249810000053
Wherein, a is the scale of wavelet transformation, and b is the displacement of wavelet transformation; a is not equal to 0, and b is any real number;
psi represents the wavelet transform function when a is 1 and b is 0;
Figure BDA0003560249810000054
in the form of psi after fourier transformation; the time domain width of the wavelet transform with a being 1 and b being 0 is delta (psi);
Figure BDA0003560249810000055
the width of the frequency domain of the wavelet transform with a 1 and b 0;
ψ(a,b)the expression scale is a,A wavelet transform function with a displacement of b;
Figure BDA0003560249810000056
is psi(a,b)A form after Fourier transform; delta (psi)(a,b)) The time domain width of the wavelet transform with the scale a and the displacement b is obtained;
Figure BDA0003560249810000057
the width of the frequency domain of the wavelet transform with the scale a and the displacement b;
due to the delta (psi) of each window in the time-frequency image(a,b)) And frequency domain width
Figure BDA0003560249810000058
The product of (a) is a constant value, i.e.
Figure BDA0003560249810000059
Figure BDA00035602498100000510
In this embodiment, the Gabor transform indicates that:
Figure BDA0003560249810000061
the calculation formula of the scale a of the wavelet transformation is as follows:
Figure BDA0003560249810000062
wherein, FcIs the wavelet center frequency, TsTo sample time, FaIs the actual frequency;
and Δ (ψ) and
Figure BDA0003560249810000063
the calculation method is as follows:
Figure BDA0003560249810000064
Figure BDA0003560249810000065
wherein, w0Gamma is a constant and is a positive value; the Gabor function is a complex-valued sinusoidal function windowed by a gaussian function centered at the origin, with the fourier transform being w-w0A central gaussian function which approximately satisfies the condition if γ is large enough, although the Gabor function does not satisfy the tolerable condition in a strict sense;
t is a time variable; w is a frequency domain variable; i represents an imaginary symbol;
thus, the time domain width Δ (ψ) of each window of the time-frequency image can be found(a,b)) And frequency domain width
Figure BDA0003560249810000066
The value of (c).
S3, according to the time domain width delta (psi) of each window in the time frequency image(a,b)) And frequency domain width
Figure BDA0003560249810000067
Respectively determining the sizes of convolution kernels convolved with windows in the time-frequency image;
the convolution kernel width convolved with a certain window in the time-frequency image is as follows: the time domain width delta (/) of the window(a,b)) Rounding to the rounded value;
the height of a convolution kernel convolved with a certain window in a time-frequency image is as follows: the frequency domain width of the window
Figure BDA0003560249810000068
And rounding to the rounded value.
In this embodiment, the bearing fault time domain signal is made into a time-frequency image of 128 × 128 pixels by continuous Gabor conversion, and the time-frequency image is obtained by calculationTo the temporal width delta phi (psi) of each window in the time-frequency image(a,b)) And frequency domain width
Figure BDA0003560249810000069
Time-domain width delta (psi) of window in time-frequency image(a,b)) The value of (c) varies in the interval of (9, + ∞), and the frequency domain width of the window in the time-frequency image
Figure BDA00035602498100000610
Is taken to be (0, 6 x 10)-2) The interval of (2) varies. If according to the temporal width delta (psi) of the window in the time-frequency image(a,b)) And frequency domain width
Figure BDA0003560249810000071
It is not practical to set the size of the convolution kernel by 1:1, so the invention selects the time-domain width Δ (ψ) of the window in the time-frequency image(a,b)) Rounding the rounded value as the convolution kernel width for convolution with the window; selecting window frequency domain width in time-frequency image
Figure BDA0003560249810000072
The rounded value is rounded again as the convolution kernel height for convolution with the window.
As shown in fig. 3, for each window in the time-frequency image shown in fig. 2, several variable convolution kernels with different sizes are designed, and the variable convolution kernels with different sizes are convolved with the image at different frequency values of the time-frequency image.
Determining these several different sized variable convolution kernels is critical to the overall model. The method comprises the steps of determining the type number of variable convolution kernels, enabling the bearing fault classification effect to be better and time to be shorter, carrying out three groups of experiment comparison in the embodiment, and enabling the type number of the variable convolution kernels of each group of experiment to be different.
In this embodiment, based on that the time-frequency image is 128 × 128 pixels, the number of the types of the variable convolution kernels selected by the three sets of experiments is 3, 7, and 15, respectively, in the first set of experiments, as shown in fig. 3, 3 kinds of variable convolution kernels with different sizes are selected, the time-frequency image is divided into 3 blocks, the size of each block of image is 64 × 128, wherein the height is 64, the width is 128, and the overlapped pixels are 32. In the second set of experiments, 7 different sizes of variable convolution kernels were selected, the time-frequency image was divided into 7 blocks, each block was 32 × 128 in image size, and the overlapping pixels were 16. In the third group of experimental species, 15 variable convolution kernels with different sizes are selected, the time-frequency image is divided into 15 blocks, the size of each block of image is 16 multiplied by 128, and the overlapped pixels are 8.
Experiments are respectively carried out according to the selection, 150 epochs are defined, the training rate is 0.01, and Adam is selected as an optimizer. Dropout is used and set to 0.5 for reducing overfitting, and the loss function chooses the cross entropy loss. The experimental results, classification accuracy and model computation time for these three different sets of variable convolution kernels are shown in table 1. From the experimental data in table 1, the corresponding accuracy in the first group is highest and the time taken for model calculation is least. This is because when the time-frequency image is divided into more blocks by pixels, not only the model complexity but also the calculation time are increased. And in order to avoid losing the edge information of the blocks, the image is taken according to overlapped pixels, but some data redundancy is caused at the same time. Therefore, we take the first group of 3 variable convolution kernels with different sizes as the variable convolution kernel selection of the experiment.
TABLE 1
Accuracy of classification Calculating time(s)
First group 99.89±0.03 2292.14
Second group 99.30±0.09 8832.70
Third group 99.15±0.21 8994.31
In the invention, a large variable convolution kernel is used in a shallow layer to extract the most original characteristic change of texture details, and then a ResNet module is used in a deep layer to convolve tensors convolved by the convolution kernel respectively, so that the characteristic correlation in a large local range is prevented from being lost. For example, if the total number of layers is five, then the first two layers can be defined as the shallow layers and the second two layers as the deep layers.
The bearing fault time domain signal is subjected to continuous Gabor wavelet transform to obtain a time-frequency image, the time-frequency image is divided into a plurality of blocks according to pixel values, and in order to avoid the situation that partial information can be lost at the edges of the divided image blocks, pixel block parts need to be repeatedly divided, namely pixel points are overlapped. The width and the height of a plurality of divided block images are calculated according to the method of the invention, and then convolution kernels with different sizes and different image blocks are convoluted respectively. Padding (padding) of different sizes is then used according to the variable convolution kernel size. Then, the tensors after being convolved by the variable convolution kernels are convolved by a ResNet module in a deep layer, and after the operation of convolution is carried out by a plurality of small convolution kernels, a plurality of three-dimensional tensors are obtained. The method comprises the steps of developing a three-dimensional tensor to obtain a two-dimensional tensor with the same value in a certain dimension, splicing the two-dimensional tensors in the same dimension to obtain a two-dimensional tensor map, and finally outputting the two-dimensional tensor map obtained through splicing as a required category through two full-connection layers.
In order to evaluate the convolutional neural network based on the present invention, the collected seven failure category data sets are tested, the average performance of five consecutive tests is used as the test result, and the classification accuracy and the calculation time of the convolutional neural network of the present invention are compared with the following four prior arts:
depth Residual Shrinkage Network (DRSN): the DRSN uses a residual error network, a self-attention network and soft thresholding to respectively reduce the training difficulty of a convolutional neural network, weight the channels of the feature map and reduce the noise of the signal.
Dynamic convolution (DynamicConv): the DynamicConv dynamically aggregates multiple parallel convolution kernels according to the associated attention, increasing model complexity without increasing the depth and width of the network.
Deep residual learning (ResNet 18): ResNet18 changes the functional relationship to be learned by the network layer into learning the residual function about the layer input, making training of deeper convolutional neural networks possible.
Visual geometry group (VGG-19): VGG-19 uses the smallest 3 x 3 convolution kernel size and the smallest interval to train and test pictures on the whole picture and multi-scale, improving the classification accuracy.
The comparative results are shown in table 2 below:
TABLE 2
Figure BDA0003560249810000091
The invention designs a new variable convolution kernel, determines the size and the type of the convolution kernel, and is used for classifying the bearing faults. The method can determine the number of the variable convolution kernels so as to realize a better solution of bearing fault classification. And combining the variable convolution kernel with a ResNet related module in the convolutional neural network to extract texture and detail information of the bearing fault. Therefore, satisfactory bearing failure classification prediction performance can be obtained.
In addition to the above-described methods, embodiments of the invention may also be a computer program product comprising computer program instructions which, when executed by a processor, cause the processor to perform the steps in the decision-making behavior decision method according to various embodiments of the invention described in this specification.
The computer program product may write program code for carrying out operations for embodiments of the present invention in any combination of one or more programming languages, including an object oriented programming language such as Java, C + + or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's computing device and partly on a remote computing device, or entirely on the remote computing device or server.
Furthermore, embodiments of the present invention may also be a computer-readable storage medium having stored thereon computer program instructions which, when executed by a processor, cause the processor to perform the steps in a decision-making behavior decision method according to various embodiments of the present invention described in the methods of the present specification.
The computer-readable storage medium may take any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may include, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or a combination of any of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
The invention is not to be considered as limited to the specific embodiments shown and described, but is to be understood to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention as defined by the appended claims.

Claims (5)

1. A method of computing a variable convolution kernel for variable resolution, comprising the steps of;
s1, performing wavelet transformation on the signals to obtain a time-frequency image of the wavelet transformation; wherein, the time resolution and the frequency resolution of each window function in the wavelet transformation are different, and the time domain width delta (psi) of each window in the time-frequency image(a，b)) And frequency domain width
Figure FDA0003560249800000011
All are different;
s2, calculating the time domain width delta (psi) of each window in the time-frequency image(a，b)) And frequency domain width
Figure FDA0003560249800000012
S3, according to the time domain width delta (psi) of each window in the time frequency image(a，b)) And frequency domain width
Figure FDA0003560249800000013
Respectively determining the sizes of convolution kernels convolved with windows in the time-frequency image;
the convolution kernel width convolved with a certain window in the time-frequency image is as follows: the time domain width delta (/) of the window(a，b)) Rounding to the rounded value;
the height of a convolution kernel convolved with a certain window in a time-frequency image is as follows: the frequency domain width of the window
Figure FDA0003560249800000014
And rounding to the rounded value.
2. The method of claim 1, wherein in step S2, the time-domain width Δ (ψ) of the window in the time-frequency image(a，b)) And frequency domain width
Figure FDA0003560249800000015
The calculation of (c) is as follows:
Δ(ψ(a，b))＝|a|Δ(ψ)
Figure FDA0003560249800000016
wherein, a is the scale of wavelet transformation, and b is the displacement of wavelet transformation; a is not equal to 0, and b is any real number;
psi represents the wavelet transform function when a is 1 and b is 0;
Figure FDA0003560249800000017
in the form of psi after fourier transformation; the time domain width of the wavelet transform with a being 1 and b being 0 is delta (psi);
Figure FDA0003560249800000018
the frequency domain width of the wavelet transform with a-1 and b-0;
ψ(a，b)a wavelet transformation function with a scale a and a displacement b is represented;
Figure FDA0003560249800000019
is psi(a，b)A form after Fourier transform; delta (psi)(a，b)) The time domain width of the wavelet transform with the scale a and the displacement b is obtained;
Figure FDA00035602498000000110
the width of the frequency domain of the wavelet transform with the scale a and the displacement b;
due to the delta (psi) of the window in the time-frequency image(a，b)) And frequency domain width
Figure FDA00035602498000000111
The product of (a) is a constant value, i.e.
Figure FDA00035602498000000112
Figure FDA00035602498000000113
The calculation mode of the scale a of the wavelet transformation is as follows:
Figure FDA0003560249800000021
wherein, FcIs the wavelet center frequency, TsTo sample time, FaIs the actual frequency;
and Δ (ψ) and
Figure FDA0003560249800000022
the calculation method is as follows:
Figure FDA0003560249800000023
Figure FDA0003560249800000024
wherein, w0Gamma is a constant and is a positive value; t is a time variable; w is a frequency domain variable; i represents an imaginary symbol;
thus, the time domain width Δ (ψ) of each window of the time-frequency image is found(a，b)) And frequency domain width
Figure FDA0003560249800000025
The value of (c).
3. The method of claim 1, wherein in step S3, in the deep layer of the convolutional neural network, the tensor convolved by the convolutional kernel is convolved by using a ResNet module.
4. The method for calculating the variable convolution kernel for variable resolution according to any one of claims 1 to 3, characterized in that the signal is a bearing fault signal, and the wavelet transform is performed on the bearing fault signal to obtain a time-frequency image of the wavelet transform of the bearing fault signal; carrying out characteristic identification on the time-frequency image of the wavelet transform of the bearing fault signal by utilizing a convolutional neural network so as to classify the bearing fault; the convolution kernel in the convolution neural network is a variable convolution kernel, namely the convolution kernel has different sizes with each window in the time-frequency image; and the convolution kernel size convolved with each window in the time-frequency image is solved by adopting the method of the steps S1-S3.
5. Storage medium for calculating a variable convolution kernel for variable resolution, characterized in that the storage medium stores a computer program comprising program instructions which, when executed by a processor, cause the processor to carry out the method according to claim 1 or 2 or 3.
CN202210304788.XA 2022-03-23 2022-03-23 Method of calculating variable convolution kernel for variable resolution, storage medium Pending CN114692831A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210304788.XA CN114692831A (en) 2022-03-23 2022-03-23 Method of calculating variable convolution kernel for variable resolution, storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210304788.XA CN114692831A (en) 2022-03-23 2022-03-23 Method of calculating variable convolution kernel for variable resolution, storage medium

Publications (1)

Publication Number Publication Date
CN114692831A true CN114692831A (en) 2022-07-01

Family

ID=82139252

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210304788.XA Pending CN114692831A (en) 2022-03-23 2022-03-23 Method of calculating variable convolution kernel for variable resolution, storage medium

Country Status (1)

Country Link
CN (1) CN114692831A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115774830A (en) * 2022-10-28 2023-03-10 贵州大学 Rapid mapping method for sparse interference array

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115774830A (en) * 2022-10-28 2023-03-10 贵州大学 Rapid mapping method for sparse interference array
CN115774830B (en) * 2022-10-28 2023-09-29 贵州大学 Rapid imaging method for sparse interference array

Similar Documents

Publication Publication Date Title
CN106682688B (en) Particle swarm optimization-based stacked noise reduction self-coding network bearing fault diagnosis method
Yu et al. A bearing fault diagnosis technique based on singular values of EEMD spatial condition matrix and Gath-Geva clustering
Chen et al. Semi-supervised multiresolution classification using adaptive graph filtering with application to indirect bridge structural health monitoring
CN108109140A (en) Low Grade Gliomas citric dehydrogenase non-destructive prediction method and system based on deep learning
CN109498037B (en) Brain cognition measurement method based on deep learning extraction features and multiple dimension reduction algorithm
CN112419203B (en) Diffusion weighted image compressed sensing recovery method and device based on countermeasure network
CN114758137B (en) Ultrasonic image segmentation method and device and computer readable storage medium
CN111415353A (en) Detection structure and detection method for fastener burr defects based on ResNet58 network
CN111680755A (en) Medical image recognition model construction method, medical image recognition device, medical image recognition medium and medical image recognition terminal
CN112861915A (en) Anchor-frame-free non-cooperative target detection method based on high-level semantic features
CN115830436A (en) Marine organism intelligent detection method based on deep learning
CN111144425B (en) Method and device for detecting shot screen picture, electronic equipment and storage medium
CN114692831A (en) Method of calculating variable convolution kernel for variable resolution, storage medium
CN115526847A (en) Mainboard surface defect detection method based on semi-supervised learning
CN114048787A (en) Method and system for intelligently diagnosing bearing fault in real time based on Attention CNN model
CN116958131B (en) Image processing method, device, equipment and storage medium
CN116805387A (en) Model training method, quality inspection method and related equipment based on knowledge distillation
CN116563251A (en) Intelligent processing method and system for tie rod ball head for automobile
CN115147348B (en) Tire defect detection method and system based on improved YOLOv3
CN113496228B (en) Human body semantic segmentation method based on Res2Net, transUNet and cooperative attention
CN113780335B (en) Small sample commodity image classification method, device, equipment and storage medium
CN113409213B (en) Method and system for enhancing noise reduction of time-frequency diagram of fault signal of plunger pump
CN115631526A (en) Shielded facial expression recognition method based on self-supervision learning technology and application
CN112346126B (en) Method, device, equipment and readable storage medium for identifying low-order faults
CN111612021B (en) Error sample identification method, device and terminal

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination