A Convolutional Neural Network Optimization Method and System Based on Adaptive Mode Decomposition

By combining adaptive mode decomposition and convolutional neural networks, the problem of improper selection of mode decomposition parameters in rotating machinery fault diagnosis is solved, achieving high-precision multi-scale feature extraction and fault diagnosis, and improving the accuracy and efficiency of rotating machinery fault diagnosis.

CN123115648APending Publication Date: 2026-11-17湛江科技学院
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611245338.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the diagnosis of rotating machinery faults, traditional time-frequency analysis and machine learning methods rely on manual feature selection, which leads to the learning of noise and irrelevant frequency components. Furthermore, improper selection of adaptive mode decomposition parameters can easily lead to mode aliasing or over-decomposition, affecting the accuracy of diagnosis.

Method used

An adaptive mode decomposition convolutional neural network method is adopted. The VMD parameters are selected by jointly evaluating three indicators: relative energy, kurtosis and correlation coefficient. Combined with dual-threshold mode screening and multi-scale one-dimensional convolutional branch, multi-scale features of rotating machinery are extracted, and the fault-related features are enhanced by channel attention mechanism.

Benefits of technology

It achieves high-precision adaptive diagnosis of rotating machinery faults, improves the quality of reconstructed signals and diagnostic accuracy, reduces model complexity, and can effectively capture multi-scale fault features and suppress invalid channels.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115648A_ABST
    Figure CN123115648A_ABST
Patent Text Reader

Abstract

This invention provides a convolutional neural network optimization method and system based on adaptive mode decomposition (VMD), relating to the field of rotating machinery fault diagnosis technology. The method includes: acquiring vibration signals and preprocessing them to obtain samples; decomposing the samples using candidate parameter combinations to obtain modal components; calculating the relative energy, kurtosis, and correlation coefficient of each modal component, weighting them to obtain a comprehensive evaluation value, and determining the target parameter combination; filtering effective modal components through dual thresholds and superimposing them to obtain a reconstructed signal; extracting features through multi-scale one-dimensional convolutional branches and generating channel attention weights; and using the channel attention weights to weight the features and determine the fault category. This invention achieves adaptive VMD parameter selection through joint evaluation of three indicators, retains fault feature modes through dual threshold screening, and enhances key features through multi-scale convolution and channel attention, thus achieving high-precision and lightweight fault diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of rotating machinery fault diagnosis technology, and in particular to a convolutional neural network optimization method and system based on adaptive mode decomposition. Background Technology

[0002] Components such as bearings and gears in rotating machinery are subjected to alternating loads for extended periods, and their fault vibration signals typically exhibit characteristics such as non-stationarity, strong noise, and weak impact components. Traditional time-frequency analysis and machine learning methods rely on manual feature selection. While convolutional neural networks can automatically extract local features, they tend to learn noise and irrelevant frequency components along with the raw signal when directly processing it.

[0003] Chinese patent document CN116186590A discloses a bearing fault diagnosis method based on data layer feature fusion and convolutional neural networks. It employs VMD decomposition, uses kurtosis thresholds to filter modes and reconstruct signals, then constructs a multidimensional composite feature matrix, which is input into a CNN after kernel principal component analysis, achieving a combination of signal decomposition and deep classification. However, it uses a fixed number of modes, and a single kurtosis criterion cannot adequately consider modal energy and its correlation with the original signal. Furthermore, the manually constructed feature matrix adds processing steps. In addition, existing approximation schemes by scholars such as Li Ke et al. and Yang Sen generally suffer from the common drawback of VMD decomposition relying on manually set parameters; improper parameter selection can easily lead to mode aliasing or over-decomposition, and directly feeding all modes into the network can cause a large number of noisy modes to be mistakenly treated as effective classification features by the neural network, thus exacerbating the training load of non-stationary signal processing.

[0004] To enhance the network's focus on key features, Chinese patent document CN115034267A further discloses a one-dimensional convolutional neural network method for rolling bearing fault diagnosis that integrates channel attention. This method adjusts channel weights through cross-channel information interaction to improve the accuracy of bearing fault detection. However, its improvements mainly focus on network feature weighting and do not address the issues of adaptive selection of non-stationary signal decomposition parameters, noise mode filtering, and collaborative extraction of impact features at different time scales.

[0005] In view of this, there is an urgent need for convolutional neural network optimization methods and systems based on adaptive mode decomposition to at least address the aforementioned shortcomings. Summary of the Invention

[0006] One of the objectives of this invention is to provide a method and system for optimizing convolutional neural networks based on adaptive mode decomposition, so as to solve the problems pointed out in the background art.

[0007] The convolutional neural network optimization method based on adaptive mode decomposition provided in this invention includes: Step S1: Collect the raw vibration signal of the rotating machinery using a vibration sensor and preprocess it to obtain a vibration signal sample; Step S2: Decompose the vibration signal sample using each candidate parameter combination consisting of candidate mode number and candidate penalty factor to obtain the corresponding modal components; Step S3: Calculate the relative energy, kurtosis and correlation coefficient of each modal component, and weight the mean values ​​of each index corresponding to the same candidate parameter combination to obtain the comprehensive evaluation value. The candidate parameter combination with the largest comprehensive evaluation value is determined as the target parameter combination. Step S4: From the modal components corresponding to the target parameter combination, determine the modal components with relative energy not less than the energy threshold and correlation coefficient not less than the correlation coefficient threshold as effective modal components, and superimpose them to obtain the reconstructed vibration signal; Step S5: Extract scale branch features of the reconstructed vibration signal through multiple one-dimensional convolutional branches with different kernel lengths, align and splice the scale branch features into multi-scale fusion features, perform global max pooling on each feature channel along the time dimension, and generate channel attention weights based on the pooling results. Step S6: Use channel attention weights to weight the multi-scale fusion features and determine the state category of rotating machinery based on the obtained channel enhancement features.

[0008] Preferably, step S1 includes: Calculate the mean and standard deviation of the original vibration signal, and then standardize it with zero mean and unit variance based on the mean and standard deviation to obtain the standardized vibration signal; The standardized vibration signal is sliced ​​according to the preset sample length to obtain multiple vibration signal samples.

[0009] Preferably, step S2 includes: Read the pre-stored set of candidate modal numbers and candidate penalty factors; The candidate modes in the candidate mode number set are combined with the candidate penalty factors in the candidate penalty factor set to obtain the candidate parameter combinations; The candidate parameter combinations are traversed, and variational mode decomposition is performed on the same vibration signal sample using each candidate parameter combination to obtain the modal components corresponding to each candidate parameter combination.

[0010] Preferably, step S3 includes: The first The ratio of the sum of squares of the amplitudes of the first modal component to the sum of squares of the amplitudes of the vibration signal samples is determined as the ratio of the sum of squares of the amplitudes of the second modal component to the sum of squares of the amplitudes of the third modal component. The relative energy of each modal component; The first The ratio of the fourth-order central moment of each modal component to its fourth standard deviation is determined as the... Kurtosis of each modal component; According to the The covariance and standard deviation between each modal component and the vibration signal sample are used to determine the first modal component. The correlation coefficients between each modal component and the vibration signal sample, where... This represents the modal component number.

[0011] Preferably, step S3 further includes: Calculate the arithmetic mean of relative energy, arithmetic mean of kurtosis, and arithmetic mean of correlation coefficient for each modal component corresponding to the same candidate parameter combination; Read the three pre-stored evaluation weights and calculate the comprehensive evaluation value according to the following formula: ; in, The number of candidate modes, As a candidate penalty factor, , and In order, they are the arithmetic mean of relative energy, the arithmetic mean of kurtosis, and the arithmetic mean of correlation coefficient. , and For evaluation weights; The candidate parameter combination with the largest comprehensive evaluation value is determined as the target parameter combination, and the number of candidate modes and the candidate penalty factor in it are determined as the target number of modes and the target penalty factor, respectively.

[0012] Preferably, step S4 includes: Read the pre-stored energy threshold and correlation coefficient threshold; The relative energy of each modal component corresponding to the target parameter combination is compared with the energy threshold, and the correlation coefficient of each modal component is compared with the correlation coefficient threshold. Modal components with relative energy not less than the energy threshold and correlation coefficient not less than the correlation coefficient threshold are defined as effective modal components.

[0013] Preferably, step S4 further includes: Modal filtering flags are set according to whether each modal component is a valid modal component. The modal filtering flag for valid modal components is 1, and the modal filtering flag for other modal components is 0. Each modal component is multiplied by its corresponding modal screening marker, and the multiplication results are superimposed according to the sampling point positions to obtain the reconstructed vibration signal.

[0014] Preferably, step S5 includes: The reconstructed vibration signal is input in parallel into the first-scale one-dimensional convolution branch, the second-scale one-dimensional convolution branch, and the third-scale one-dimensional convolution branch, with the kernel length of the three one-dimensional convolution branches increasing sequentially. The reconstructed vibration signal is convolved and nonlinearly activated by three one-dimensional convolutional branches to obtain the corresponding scale branch features. The features of the three scale branches are aligned along the time dimension and then spliced ​​along the channel dimension to obtain multi-scale fused features.

[0015] Preferably, steps S5 and S6 include: Global max pooling is performed along the time dimension on each feature channel of the multi-scale fused feature to obtain the corresponding channel descriptor; The channel attention weights for each feature channel are generated based on the channel descriptors and the channel attention parameters obtained during training. The attention weights of each channel are multiplied by their corresponding feature channels to obtain the channel-enhanced features; The channel enhancement features are input into the Softmax classifier to obtain the state category probability of each preset rotating machinery state category, and the rotating machinery state category corresponding to the maximum state category probability is determined as the diagnostic result. The network parameters of the one-dimensional convolutional branch, the channel attention parameters, and the classification parameters of the Softmax classifier are obtained by training with vibration signal training samples with state category labels and the cross-entropy loss function.

[0016] The convolutional neural network optimization system based on adaptive mode decomposition provided in this invention includes: Vibration sensors are installed on rotating machinery to collect raw vibration signals during the operation of the rotating machinery. The memory is used to store the computer program, the set of candidate modalities, the set of candidate penalty factors, evaluation weights, energy thresholds, correlation coefficient thresholds, and convolutional neural network parameters; and The processor, connected to the vibration sensor and memory, is used to execute a computer program to perform any of the aforementioned rotating machinery fault diagnosis methods on the raw vibration signal and output the rotating machinery condition category.

[0017] The beneficial effects of this invention are as follows: This invention achieves adaptive selection of VMD parameters through a joint evaluation of relative energy, kurtosis, and correlation coefficient, avoiding parameter selection bias caused by a single indicator. It removes noise and redundant modes through a dual-threshold mode selection mechanism, retaining fault characteristic modes and improving the quality of the reconstructed signal. It extracts features from different temporal receptive fields through multi-scale one-dimensional convolutional branches, comprehensively capturing the multi-scale features of rotating machinery faults. Finally, it adaptively enhances fault-related feature channels and suppresses invalid channels through a channel attention mechanism, improving diagnostic accuracy. Overall, it forms a complete technical chain from signal preprocessing, adaptive parameter selection, intelligent mode selection, multi-scale feature extraction to attention-weighted classification, achieving high-precision adaptive diagnosis of rotating machinery faults while maintaining low model complexity.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the convolutional neural network optimization method based on adaptive mode decomposition in an embodiment of the present invention; Figure 2 This is a schematic diagram of a convolutional neural network optimization system based on adaptive mode decomposition in an embodiment of the present invention; Figure 3 This is a comparison chart of the diagnostic results of different fault diagnosis models in the embodiments of the present invention; Figure 4 The diagram shows the ablation experiment results of each functional module in the embodiments of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] Example 1: This example provides a convolutional neural network optimization method based on adaptive mode decomposition, such as... Figure 1 As shown, it includes: Step S1: Collect the raw vibration signal of the rotating machinery using a vibration sensor and preprocess it to obtain a vibration signal sample; Step S2: The vibration signal samples are decomposed using various candidate parameter combinations consisting of candidate mode numbers and candidate penalty factors to obtain the corresponding modal components. Step S3: Calculate the relative energy, kurtosis and correlation coefficient of each modal component, and weight the mean values ​​of each index corresponding to the same candidate parameter combination to obtain the comprehensive evaluation value. The candidate parameter combination with the largest comprehensive evaluation value is determined as the target parameter combination. Step S4: From the modal components corresponding to the target parameter combination, determine the modal components with relative energy not less than the energy threshold and correlation coefficient not less than the correlation coefficient threshold as effective modal components, and superimpose them to obtain the reconstructed vibration signal. Step S5: Extract scale branch features of the reconstructed vibration signal through multiple one-dimensional convolutional branches with different kernel lengths, align and splice the scale branch features into multi-scale fusion features, perform global max pooling on each feature channel along the time dimension, and generate channel attention weights based on the pooling results. Step S6: Use channel attention weights to weight the multi-scale fusion features, and determine the rotating machinery state category based on the obtained channel enhancement features.

[0023] In this embodiment, the vibration sensor is a PCB 352C33 type accelerometer, which is installed on the bearing housing at the position closest to the outer ring of the bearing, and the sampling frequency is set to 12800Hz; the original vibration signal is an acceleration time domain signal, and the acquisition time is not less than 10 seconds.

[0024] In this embodiment, preprocessing includes standardization and sample slicing: standardization adopts the zero-mean, unit variance standardization method, that is, subtracting the mean of each signal segment and dividing by its standard deviation to eliminate amplitude differences and DC bias between different operating conditions and different sensors; sample slicing adopts a sliding window method, with each sample containing 2048 sampling points and the window step size set to 1024 sampling points (i.e., the window overlap rate is 50%).

[0025] In this embodiment, the candidate mode number set is {3,4,5,6,7,8}, and the candidate penalty factor set is {500,1000,2000,3000,5000}. The selection of the candidate mode number is based on the fact that rotating machinery vibration signals typically contain 3 to 8 main frequency components (fundamental frequency, harmonics, and fault characteristic frequencies). The selection range of the candidate penalty factor refers to the recommended value range for vibration signals in the VMD algorithm literature.

[0026] In this embodiment, Variational Mode Decomposition (VMD) is a non-recursive signal decomposition method with explicit variational constraints, proposed by Dragomiretskiy and Zosso in 2014. VMD decomposes the input signal into K modal components with finite bandwidth by solving a constrained variational problem. Each modal component has a specific center frequency. The decomposition process is solved iteratively using the Alternating Direction Multiplier (ADMM) method until the modal components converge.

[0027] In this embodiment, the energy threshold is set to 0.05, meaning that modal components with an energy percentage of not less than 5% are retained. This threshold is determined through statistical analysis of 100 sets of bearing vibration signals. Modes with an energy percentage of less than 5% are usually noise modes or redundant modes. The correlation coefficient threshold is set to 0.3, meaning that modal components with an absolute correlation coefficient with the original signal of not less than 0.3 are retained. Modes with a correlation coefficient of less than 0.3 have a weak correlation with the original signal and are difficult to represent the characteristics of the original signal.

[0028] In this embodiment, the reconstructed vibration signal represents the noise-reduced signal obtained by retaining only the effective modal components and superimposing them according to the time sampling point positions. It has the same length as the original sample with 2048 sampling points. Through modal filtering and reconstruction, the signal-to-noise ratio can be improved by 5 to 10 dB.

[0029] In this embodiment, the rotating machinery condition categories include four preset categories: normal state, inner ring fault, outer ring fault, and rolling element fault. The inner ring fault, outer ring fault, and rolling element fault correspond to defects such as pitting and spalling at different locations of the bearing.

[0030] The working principle and beneficial effects of the above technical solution: This invention achieves adaptive selection of VMD parameters through a three-index joint evaluation: relative energy (from the perspective of energy distribution), kurtosis (from the perspective of impact characteristics), and correlation coefficient (from the perspective of signal consistency) comprehensively evaluate the VMD decomposition quality, avoiding the instability of decomposition quality caused by manually setting parameters based on experience. A dual-threshold modal screening mechanism removes noise and redundant modes while retaining fault characteristic modes, improving the quality of the reconstructed signal. Multi-scale one-dimensional convolutional branches extract features from receptive fields at different times, comprehensively capturing the multi-scale features of rotating machinery faults. A channel attention mechanism adaptively enhances fault-related feature channels and suppresses invalid channels, improving diagnostic accuracy. Overall, this forms a complete technical chain from signal preprocessing, adaptive parameter selection, intelligent modal screening, multi-scale feature extraction to attention-weighted classification, achieving high-precision adaptive diagnosis of rotating machinery faults.

[0031] Example 2: Based on Example 1, the convolutional neural network optimization method based on adaptive mode decomposition includes step S1 as follows: Calculate the mean and standard deviation of the original vibration signal, and then standardize it with zero mean and unit variance based on the mean and standard deviation to obtain the standardized vibration signal; The standardized vibration signal is sliced ​​according to the preset sample length to obtain multiple vibration signal samples.

[0032] In this embodiment, zero-mean, unit variance standardization is a commonly used data preprocessing method. Specifically, for each sampling point in the original vibration signal, the mean of the signal segment is subtracted, and then divided by the standard deviation of the signal segment, so that the mean of the standardized signal is 0 and the standard deviation is 1. This standardization method eliminates amplitude differences and DC bias between different working conditions and different sensors, so that the vibration signal samples have uniform dimensions and distribution characteristics.

[0033] In this embodiment, the preset sample length is 2048 sampling points, and a sliding window method is used for slicing. The window step size is set to 1024 sampling points (i.e., the window overlap rate is 50%). By overlapping slices, the number of training samples can be increased and fault features at the window boundary can be captured.

[0034] The working principle and beneficial effects of the above technical solution: This invention makes vibration signals collected by different working conditions and different sensors comparable by standardizing with zero mean and unit variance, avoiding the impact of amplitude differences on VMD decomposition and neural network training, and improving the generalization ability of the algorithm. By using fixed-length slicing and sliding window strategies, continuous vibration signals are converted into a sample set with a uniform format, which not only meets the consistency requirement of input dimension of convolutional neural networks, but also increases the number of samples by overlapping slices, thereby improving the network training effect.

[0035] Example 3: Based on Example 1, the optimization method for convolutional neural networks based on adaptive mode decomposition includes step S2 as follows: Read the pre-stored set of candidate modal numbers and candidate penalty factors; The candidate modes in the candidate mode number set are combined with the candidate penalty factors in the candidate penalty factor set to obtain the candidate parameter combinations; The candidate parameter combinations are traversed, and variational mode decomposition is performed on the same vibration signal sample using each candidate parameter combination to obtain the modal components corresponding to each candidate parameter combination.

[0036] In this embodiment, variational mode decomposition achieves adaptive signal decomposition by solving a constrained variational problem, decomposing the input signal into K modal components, each with a center frequency and a finite bandwidth. The variational problem is solved numerically using the alternating direction multiplier method (ADMM), iteratively updating the modal components, center frequency, and Lagrange multipliers until convergence. This achieves adaptive and fine separation of different modal components and avoids redundant mode aliasing. The convergence criterion is set to the change in modal components between two adjacent iterations being less than [a certain value]. .

[0037] In this embodiment, the penalty factor Controlling the bandwidth of modal components, A larger modal bandwidth results in a narrower modal bandwidth and higher modal separation, but an excessively large one... This may lead to over-decomposition; mode number The decision was made to decompose the signal into several modal components. Too small a value can lead to modal aliasing. Excessive values ​​can lead to spurious modes, so it is necessary to select the optimal combination of parameters through comprehensive evaluation.

[0038] The working principle and beneficial effects of the above technical solution: This invention performs multiple VMD decompositions on the same vibration signal sample by traversing candidate parameter combinations to obtain decomposition results under different parameters, providing a complete candidate set for subsequent comprehensive evaluation and parameter optimization. Compared with empirical mode decomposition, variational mode decomposition has better anti-mode aliasing ability and a rigorous mathematical theoretical basis. Through variational framework and frequency domain constraints, it can effectively separate the modes of different frequency components in rotating machinery vibration signals, providing high-quality modal components for subsequent fault feature extraction.

[0039] Example 4: Based on Example 1, the optimization method for convolutional neural networks based on adaptive mode decomposition includes step S3 as follows: The ratio of the sum of squares of the amplitude of the k-th modal component to the sum of squares of the amplitude of the vibration signal sample is determined as the relative energy of the k-th modal component; The ratio of the fourth central moment of the k-th modal component to its fourth standard deviation is determined as the kurtosis of the k-th modal component. The correlation coefficient between the k-th modal component and the vibration signal sample is determined based on the covariance and standard deviation of the two, where k is the modal component number.

[0040] In this embodiment, the formula for calculating the relative energy of the k-th modal component is: ; in, Let be the relative energy of the k-th modal component; For the k-th modal component at the sampling point The amplitude at that point; For the standardized vibration signal sample at the sampling point The amplitude at the given point; the numerator is the energy of the modal component, the denominator is the total energy of the original signal sample, and the relative energy reflects the proportion of information carried by the mode.

[0041] In this embodiment, the kurtosis calculation formula for the k-th modal component is: ; in, Let be the kurtosis of the k-th modal component; For the k-th modal component at the sampling point The amplitude at that point; The mean of the k-th modal component; The sample length is given by the numerator, the fourth central moment by the denominator, and the kurtosis is a dimensionless index. The kurtosis of a normal signal is approximately 3 (which follows a Gaussian distribution). Signals containing impact components have a kurtosis significantly greater than 3. Periodic impact pulses generated by rotating machinery bearing failures can cause modal kurtosis to reach 5 to 15 or even higher.

[0042] In this embodiment, the formula for calculating the correlation coefficient between the k-th modal component and the vibration signal sample is as follows: ; in, The correlation coefficient between the k-th modal component and the original signal; The mean of the vibration signal samples is given by the formula. The Pearson correlation coefficient is calculated, with a value ranging from -1 to 1. The closer the absolute value of the correlation coefficient is to 1, the stronger the linear correlation between the mode and the original signal. The closer it is to 0, the weaker the correlation.

[0043] The working principle and beneficial effects of the above technical solution: Relative energy evaluates the effectiveness of modes from the perspective of energy distribution. Modes with a high energy proportion usually carry more useful information, while modes with a low energy proportion are often noise modes. Kurtosis is highly sensitive to impact pulses. Periodic impacts from rotating machinery faults (such as bearing pitting and gear cracks) significantly increase mode kurtosis. Therefore, kurtosis is an important indicator for distinguishing fault-characteristic modes from stationary background noise. Correlation coefficient measures the consistency between a mode and the original signal. Modes with a high correlation coefficient are more representative of the original signal characteristics, while modes with a low correlation coefficient may be spurious modes generated by excessive VMD decomposition. These three indicators, based on the different physical meanings of energy, impact, and correlation, are used in a joint evaluation to ensure that the selected VMD parameters can produce high-quality, fault-sensitive, and highly representative modal components.

[0044] Example 5: Based on Example 4, the convolutional neural network optimization method based on adaptive mode decomposition further includes step S3 as follows: Calculate the arithmetic mean of relative energy, arithmetic mean of kurtosis, and arithmetic mean of correlation coefficient for each modal component corresponding to the same candidate parameter combination; Read the three pre-stored evaluation weights and calculate the comprehensive evaluation value according to the following formula: ; in, The number of candidate modes, As a candidate penalty factor, , and In order, they are the arithmetic mean of relative energy, the arithmetic mean of kurtosis, and the arithmetic mean of correlation coefficient. , and For evaluation weights; The candidate parameter combination with the largest comprehensive evaluation value is determined as the target parameter combination, and the number of candidate modes and the candidate penalty factor in it are determined as the target number of modes and the target penalty factor, respectively.

[0045] In this embodiment, the arithmetic mean is calculated as follows: for the same candidate parameter combination corresponding Each modal component is used to calculate its relative energy. , cliff and the absolute value of the correlation coefficient Then calculate the arithmetic mean. , , The correlation coefficient is taken as an absolute value because negative and positive correlations are equivalent in characterizing the strength of the correlation between the mode and the original signal.

[0046] In this embodiment, to avoid the impact of numerical scale differences between indicators of different dimensions on the comprehensive evaluation, the three arithmetic means are normalized before calculating the comprehensive evaluation value: the corresponding values ​​of each candidate parameter are combined. , , Subtract the minimum value of the indicator from all candidate parameter combinations, and then divide by the difference between the maximum and minimum values ​​of the indicator, so that the normalized indicator values ​​all range from 0 to 1. The normalization process makes the three indicators weighted on the same scale, avoiding the dominance of indicators with large numerical ranges in the evaluation results.

[0047] In this embodiment, the evaluation weight , and The sum of the three is 1; the actual configuration used in this algorithm is... , , The kurtosis weight is set to the highest because kurtosis is most sensitive to the impact characteristics of faults and is a core indicator for fault diagnosis. The relative energy and correlation coefficient weights are equal to balance the information content and representativeness of the modes. This set of weight values ​​was determined by grid search and cross-validation of 200 sets of bearing vibration signals under different working conditions. The fault diagnosis accuracy corresponding to the VMD parameter selection results using this weight combination is the highest.

[0048] The working principle and beneficial effects of the above technical solution: This invention evaluates the overall quality of the mode set corresponding to each candidate parameter combination by calculating the arithmetic mean of three indicators, avoiding the influence of individual abnormal modes on the evaluation results. It constructs a comprehensive evaluation value through weighted summation, integrating the evaluations of energy distribution, impact characteristics, and signal correlation into a single indicator, thus transforming a multi-objective optimization problem into a single-objective optimization problem and facilitating the determination of the optimal parameter combination. Normalization eliminates differences in the dimensions and numerical scales of different indicators, ensuring that the evaluation weights truly reflect the importance of each indicator. Finally, by selecting the candidate parameter combination with the largest comprehensive evaluation value as the target parameter combination, it achieves adaptive optimization of VMD parameters without the need for manual experience or repeated trial and error.

[0049] Example 6: Based on Example 1, the optimization method for convolutional neural networks based on adaptive mode decomposition includes step S4 as follows: Read the pre-stored energy threshold and correlation coefficient threshold; The relative energy of each modal component corresponding to the target parameter combination is compared with the energy threshold, and the correlation coefficient of each modal component is compared with the correlation coefficient threshold. Modal components with relative energy not less than the energy threshold and correlation coefficient not less than the correlation coefficient threshold are defined as effective modal components.

[0050] In this embodiment, modal screening employs a dual threshold judgment condition: for the k-th modal component, if its relative energy... Not less than the energy threshold And the absolute value of its correlation coefficient Not less than the correlation coefficient threshold If the modal component is positive, it is considered a valid modal component. The correlation coefficient is also taken as an absolute value because the negatively correlated mode (opposite phase) also contains information of the original signal. The "AND" logic of the double threshold makes the valid mode satisfy both the energy and correlation conditions at the same time, filtering out both low-energy noise modes and false modes that are weakly correlated with the original signal.

[0051] In this embodiment, the energy threshold The following method was used to determine the energy threshold: 100 sets of normal and faulty bearing vibration signals under different working conditions were statistically analyzed. The optimal VMD parameters were used to decompose each set of signals and analyze the energy distribution of each mode. It was found that the mode with an energy ratio higher than 5% contained more than 95% of the effective information of the signal, while the mode with an energy ratio lower than 5% was mainly high-frequency noise and redundant components. Therefore, the energy threshold was set to 0.05.

[0052] In this embodiment, the correlation coefficient threshold The correlation coefficient was determined as follows: For the same 100 sets of signal samples, the correlation coefficient distribution between each mode and the original signal was analyzed. It was found that the modes with an absolute correlation coefficient higher than 0.3 had a high similarity to the original signal waveform and could effectively characterize the features of the original signal. The modes with an absolute correlation coefficient lower than 0.3 were often pseudo modes generated by VMD over-decomposition or interference components unrelated to the features of the original signal. Therefore, the correlation coefficient threshold was set to 0.3.

[0053] The working principle and beneficial effects of the above technical solution: This invention employs a dual-threshold screening mechanism to perform secondary screening of modes based on the target parameter combination. The energy threshold ensures that the retained modes carry sufficient signal energy from an information content perspective, while the correlation coefficient threshold ensures that the retained modes are highly correlated with the original signal from a correlation perspective. This dual constraint removes noisy and redundant modes while preserving fault characteristic modes to the greatest extent possible. Compared to directly using all modes or screening based on only a single indicator, the dual-threshold mechanism can more accurately distinguish between effective and ineffective modes, significantly improving the signal-to-noise ratio and fault characteristic saliency of the reconstructed signal.

[0054] Example 7: Based on Example 6, the convolutional neural network optimization method based on adaptive mode decomposition further includes step S4 as follows: Modal filtering flags are set according to whether each modal component is a valid modal component. The modal filtering flag for valid modal components is 1, and the modal filtering flag for other modal components is 0. Each modal component is multiplied by its corresponding modal screening marker, and the multiplication results are superimposed according to the sampling point positions to obtain the reconstructed vibration signal.

[0055] In this embodiment, the formula for calculating the reconstructed vibration signal is: ; in, To reconstruct the vibration signal at the sampling point The amplitude at that point; For the mode selection flag of the k-th modal component, when When it is an effective modal component ,otherwise ; For the k-th modal component at the sampling point The amplitude at that point; The target number of modes is determined by multiplying the markers with the modal components point by point. Valid modes marked with a modal screening mark of 1 are fully retained, while invalid modes marked with a modal screening mark of 0 are filtered out (reset to zero after multiplication). All modes are superimposed according to the sampling point positions to obtain the reconstructed vibration signal, which has the same length as the original sample (2048 sampling points).

[0056] In this embodiment, taking a VMD decomposition result containing 6 modal components as an example, if the 1st, 3rd, and 5th modes are valid modal components, and the 2nd, 4th, and 6th modes are invalid modal components, then the modality filtering label vector is: The reconstructed vibration signal is Invalid mode , , It is filtered out and does not participate in the reconstruction.

[0057] In this embodiment, compared with the original vibration signal sample, the reconstructed vibration signal, combined with the dataset test conditions of Example 11 and the quantitative evaluation of the subsequent ablation model experiment, effectively improves the signal-to-noise ratio by 5 to 10 dB after precise screening and isolation of the effective modal dual thresholds, and improves the spectral amplitude of the fault feature frequency by 20% to 50%, providing a higher quality input signal for subsequent convolutional neural network feature extraction.

[0058] The working principle and beneficial effects of the above technical solution: This invention achieves automatic retention of effective modes and automatic filtering of invalid modes through 0-1 encoding of modal screening tags and point-by-point multiplication operations. The operation is simple and computationally efficient. By superimposing effective modes, a reconstructed vibration signal is obtained. This signal retains the fault-related feature components in the original signal and filters out noise and redundant components. This reduces the computational load of subsequent processing and improves the saliency of fault features, providing high-quality input for convolutional neural networks.

[0059] Example 8: Based on Example 1, the optimization method for convolutional neural networks based on adaptive mode decomposition includes step S5 as follows: The reconstructed vibration signal is input in parallel into the first-scale one-dimensional convolution branch, the second-scale one-dimensional convolution branch, and the third-scale one-dimensional convolution branch, with the kernel length of the three one-dimensional convolution branches increasing sequentially. The reconstructed vibration signal is convolved and nonlinearly activated by three one-dimensional convolutional branches to obtain the corresponding scale branch features. The features of the three scale branches are aligned along the time dimension and then spliced ​​along the channel dimension to obtain multi-scale fused features.

[0060] In this embodiment, the kernel length of the first-scale one-dimensional convolution branch is 16, the stride is 1, the padding method is same padding (i.e., the output length is the same as the input length), the number of convolution kernels is 32, and the activation function is ReLU; the receptive field of the first-scale branch is 16 sampling points, corresponding to a time length of 1.25ms (sampling frequency 12800Hz). This scale is suitable for capturing short-term features such as local impact pulses.

[0061] In this embodiment, the kernel length of the second-scale one-dimensional convolution branch is 64, the stride is 1, the padding method is same padding, the number of kernels is 32, and the activation function is ReLU; the receptive field of the second-scale branch is 64 sampling points, corresponding to a time length of 5ms. This scale is suitable for capturing the combined features of local impacts and short-period oscillations.

[0062] In this embodiment, the kernel length of the third-scale one-dimensional convolution branch is 128, the stride is 1, the padding method is same padding, the number of kernels is 32, and the activation function is ReLU; the receptive field of the third-scale branch is 128 sampling points, corresponding to a time length of 10ms. This scale is suitable for capturing periodic modulation features and long-period change trends.

[0063] In this embodiment, the calculation process of the one-dimensional convolutional layer is as follows: for the reconstructed vibration signal The j-th convolutional kernel at the sampling point The output at this point is obtained by weighted summation of the signal values ​​within the convolution kernel window, adding a bias, and then calculating using an activation function; the ReLU activation function expression is: It can introduce nonlinearity while maintaining a positive response.

[0064] In this embodiment, three convolutional branches process the same reconstructed vibration signal in parallel. The input dimension is (batch size, 1, 2048) and the output dimension is (batch size, 32, 2048). Since the same padding is used, the time dimension length of the output of the three branches is the same, so no additional alignment operation is required. The outputs of the three branches are concatenated along the channel dimension to obtain a multi-scale fusion feature with a dimension of (batch size, 96, 2048), where 96 = 32 + 32 + 32 is the total number of channels.

[0065] In this embodiment, the rotating machinery fault features have multi-scale characteristics: the impact pulse width of the bearing outer ring fault is about 0.5 to 2 ms (corresponding to 6 to 26 sampling points), the impact pulse width of the inner ring fault is slightly longer, and the impact pulse width of the rolling element fault is shorter; the periodic modulation period of the fault feature frequency is usually 5 to 20 ms; through three branches with different convolution kernel lengths, fault features at different time scales can be captured respectively, realizing the comprehensive extraction of multi-scale features.

[0066] The working principle and beneficial effects of the above technical solution: This invention reconstructs vibration signals through parallel processing of three one-dimensional convolutional branches with progressively increasing kernel lengths. The small-scale convolutional kernel (length 16) has a small receptive field, is sensitive to local impact pulses, and can capture transient impacts generated by faults; the medium-scale convolutional kernel (length 64) has a medium receptive field, and can capture combination patterns and short-period features of impact pulses; the large-scale convolutional kernel (length 128) has a large receptive field, and can capture macroscopic features such as periodic modulation and long-period trends. The features of the three scales are spliced ​​and fused into multi-scale fusion features through channel dimension, realizing a comprehensive expression of fault features from micro to macro and from local to global, avoiding the one-sidedness of single-scale feature extraction, and improving the completeness and robustness of fault features.

[0067] Example 9: Based on Example 8, the convolutional neural network optimization method based on adaptive mode decomposition includes steps S5 and S6 as follows: Global max pooling is performed along the time dimension on each feature channel of the multi-scale fused feature to obtain the corresponding channel descriptor; The channel attention weights for each feature channel are generated based on the channel descriptors and the channel attention parameters obtained during training. The attention weights of each channel are multiplied by their corresponding feature channels to obtain the channel-enhanced features; The channel enhancement features are input into the Softmax classifier to obtain the state category probability of each preset rotating machinery state category, and the rotating machinery state category corresponding to the maximum state category probability is determined as the diagnostic result. The network parameters of the one-dimensional convolutional branch, the channel attention parameters, and the classification parameters of the Softmax classifier are obtained by training with vibration signal training samples with state category labels and the cross-entropy loss function.

[0068] In this embodiment, global max pooling extracts the maximum value of each feature channel along the time dimension of 2048 sampling points. The channel descriptor of the c-th feature channel is the maximum feature value of the channel at all time positions. After global max pooling, 96 feature channels are obtained to obtain a 96-dimensional channel descriptor vector. Max pooling is used instead of average pooling because the characteristics of rotating machinery faults are sparse impact pulses. Max pooling can capture the peak response of the channel and is more suitable for characterizing the impact characteristic intensity.

[0069] In this embodiment, the channel attention weights are calculated as follows: a 96-dimensional channel descriptor vector is input into a two-layer fully connected network. The first layer compresses the 96-dimensional vector to 16-dimensional vectors and activates it using ReLU. The second layer expands the 16-dimensional vectors back to 96-dimensional vectors and normalizes the weight values ​​to the 0-1 range using a Sigmoid activation function. This structure is a compression-activation structure with a compression ratio of 6 (96 / 16=6), which reduces the number of parameters to avoid overfitting while still learning the nonlinear interaction relationships between channels. The Sigmoid activation function expression is: .

[0070] In this embodiment, the channel enhancement feature is calculated as follows: the attention weight of the c-th channel is multiplied by the feature value of the channel at all time positions. When the attention weight is close to 1, the corresponding channel is enhanced, and when the weight is close to 0, the corresponding channel is suppressed. By weighting each channel, the fault-related feature channels are adaptively enhanced, while redundant or noisy channels are suppressed.

[0071] In this embodiment, the channel enhancement features are processed by global average pooling to obtain a 96-dimensional feature vector, which is then input into a Softmax classifier. The classifier is a two-layer fully connected network: the first layer has dimensions (64, 96) and uses the ReLU activation function; the second layer has dimensions (4, 64) and uses the Softmax activation function, outputting the probability distribution of four state categories. The Softmax function expression is as follows: ,in The output score for the c-th category, The total number of categories; the diagnosis result is the category with the highest probability.

[0072] In this embodiment, the network training uses the Adam optimizer with an initial learning rate of 0.001, which decays to 0.5 times the original rate every 50 epochs. The batch size is 64, and the number of training epochs is 200. The training samples consist of 2000 samples in total, with 500 samples for each of the four state categories. The samples in each category are expanded to 500 through data augmentation (adding Gaussian white noise of different intensities, random cropping, and random scaling). The cross-entropy loss function is used to measure the difference between the predicted probability distribution and the true label distribution. By minimizing the cross-entropy loss, the network parameters are trained to make the diagnostic results approximate the true labels.

[0073] The working principle and beneficial effects of the above technical solution: This invention extracts the peak response of each feature channel as a channel descriptor using global max pooling. This descriptor reflects the channel's sensitivity to impact features, providing a basis for generating channel attention weights. Through a compressed-excitation structure channel attention mechanism, it learns the importance weights of each channel, adaptively enhancing fault-related feature channels and suppressing irrelevant channels, significantly improving the network's sensitivity to key features and its anti-interference ability. Through end-to-end training, the parameters of the convolutional branch, attention module, and classifier are co-optimized, and the network automatically learns the mapping relationship from reconstructed vibration signals to fault categories, eliminating the need for manual feature extraction rule design. Through the cross-entropy loss function and Adam optimizer, the network training converges quickly and has good generalization performance, improving the accuracy of fault diagnosis.

[0074] Example 10: This example provides a convolutional neural network optimization system based on adaptive mode decomposition, such as... Figure 2 As shown, it includes: Vibration sensors are installed on rotating machinery to collect raw vibration signals during the operation of the rotating machinery. The memory is used to store the computer program, the set of candidate modalities, the set of candidate penalty factors, evaluation weights, energy thresholds, correlation coefficient thresholds, and convolutional neural network parameters; and The processor, connected to the vibration sensor and memory, is used to execute a computer program to perform any of the rotating machinery fault diagnosis methods described above on the raw vibration signal and output the rotating machinery state category.

[0075] In this embodiment, the vibration sensor is a PCB 352C33 type ICP accelerometer with a sensitivity of 100mV / g and a frequency response range of 0.5Hz to 10kHz. It is mounted on the bearing housing at the horizontal or vertical position closest to the outer ring of the bearing via a magnetic base or bolts. The sensor output signal is transmitted to the data acquisition card via a shielded cable.

[0076] In this embodiment, the system also includes a data acquisition card, model NI USB-6366, with a sampling frequency set to 12.8kHz or 25.6kHz, an analog-to-digital conversion accuracy of 16 bits, and an input voltage range of ±10V. The data acquisition card converts the analog signals output by the sensor into digital signals and transmits them to the processor via the USB interface.

[0077] In this embodiment, the storage consists of DDR4 memory with a capacity of not less than 8GB and a solid-state drive with a capacity of not less than 256GB. The memory is used to store intermediate data and network parameters during runtime, while the solid-state drive is used to persistently store computer programs, configuration parameters, and historical diagnostic records. The candidate modality set, candidate penalty factor set, evaluation weights, energy thresholds, and correlation coefficient thresholds are stored in the configuration file config.yaml, and the convolutional neural network parameters (including convolutional layer weights, channel attention parameters, and classifier weights) are stored in the model file model.pth.

[0078] In this embodiment, the processor is a multi-core CPU with a main frequency of not less than 2.5 GHz or a GPU with CUDA computing capabilities; the processor reads and executes the computer program in the memory, the computer program includes a VMD decomposition module, a parameter optimization module, a modal screening module, a convolutional neural network module, and a diagnostic output module; the processor obtains the original vibration signal from the vibration sensor, and sequentially performs signal preprocessing, VMD parameter optimization, modal screening and reconstruction, multi-scale feature extraction, attention weighting, and fault classification, and finally outputs the rotating machinery state category.

[0079] In this embodiment, the system also includes a display and an alarm device; the display shows the vibration signal waveform, spectrum, VMD decomposition results, reconstructed signal waveform, and diagnostic results in real time; when the diagnostic results indicate a fault state, the alarm device issues an audible and visual alarm signal to remind the operator to handle the situation promptly.

[0080] In this embodiment, the system's diagnostic delay (the time from acquiring vibration signals to outputting diagnostic results) is less than 1 second, meeting the requirements for online real-time diagnosis; the system can monitor multiple rotating machines simultaneously, and achieve synchronous diagnosis of multiple devices through multi-channel data acquisition and multi-threaded parallel processing.

[0081] The working principle and beneficial effects of the above technical solution: This invention integrates a vibration sensor, data acquisition card, processor, memory, and display alarm device into a complete fault diagnosis system, realizing an automated closed loop from signal acquisition, processing and analysis to result output. The vibration sensor collects the operating status information of rotating machinery in real time, and the processor executes the fault diagnosis method of this invention to intelligently analyze the vibration signal, output accurate fault categories, and issue timely alarms, providing a reliable basis for equipment maintenance. The system has the advantages of strong real-time performance, high accuracy, and flexible deployment, and is suitable for online monitoring and fault diagnosis of rotating machinery in industrial scenarios such as power, petrochemical, and metallurgy.

[0082] Example 11: This example experimentally verifies the convolutional neural network optimization method based on adaptive mode decomposition.

[0083] In this embodiment, rolling bearing vibration data published by the Bearing Data Center of Case Western Reserve University are used as experimental data. The data used includes normal state data and drive-end bearing fault data. The sampling frequency of the drive-end data is 12kHz, and the fault locations include the inner ring, rolling elements, and outer ring of the bearing. Fault states are divided according to different fault diameters, thus forming 10 preset rotating machinery state categories together with the normal state.

[0084] In this embodiment, each vibration signal sample contains 1024 sampling points. This sample length is an experimental parameter set for the 12kHz publicly available bearing data used in this embodiment, and is a different feasible parameter configuration than the 2048 sampling points set for the 12.8kHz sampling signal in Embodiment 1. The preset sample length can be set according to the sampling frequency, rotating machinery speed, fault characteristic period, and computing resources, and is not limited to 1024 or 2048 sampling points.

[0085] In this embodiment, to reduce data correlation caused by the same original signal being sliced ​​into adjacent windows and then entering the training and test sets separately, vibration signal samples from the same source data file are divided into the same data subset, and the partitioning is performed at the file level according to the ratio of 70% for the training set, 15% for the validation set, and 15% for the test set. The training set is used to train the parameters of the one-dimensional convolutional branch, the channel attention module, and the classifier; the validation set is used to select the number of training epochs and network hyperparameters; and the test set is used to evaluate the fault diagnosis performance after training. A fixed random seed of 2026 is used in the experiment to ensure that the data partitioning and model initialization process can be repeated. The experimental data and parameter settings are shown in Table 1.

[0086] Table 1 Experimental data and parameter settings In this embodiment, the processing flow of the method of the present invention is shown in Table 2. First, the original vibration signal is standardized with zero mean and unit variance, and sliced ​​according to a preset sample length. Then, variational mode decomposition is performed using candidate parameter combinations composed of candidate mode numbers and candidate penalty factors, respectively, to calculate the relative energy, kurtosis, and correlation coefficient of the corresponding modal components, and the target parameter combination is determined based on the comprehensive evaluation value. Next, effective modal components are screened and the vibration signal is reconstructed based on the energy threshold and correlation coefficient threshold. Finally, the reconstructed vibration signal is input into a multi-scale one-dimensional convolution branch, and the rotating machinery state category is obtained through channel attention weighting and Softmax classification.

[0087] Table 2 Processing flow of the method of the present invention In this embodiment, to verify the combined effect of adaptive variational mode decomposition, effective mode selection, multi-scale convolution, and channel attention, a basic one-dimensional convolutional neural network, a combination model of variational mode decomposition and convolutional neural network, a combination model of convolutional neural network and bidirectional long short-term memory network, a combination model of variational mode decomposition, convolutional neural network, and bidirectional long short-term memory network, a complex hybrid model including convolutional block attention modules, and the method of this invention are compared under the same data partitioning conditions. The structure of each model and the purpose of comparison are shown in Table 3.

[0088] Table 3 Comparison Model Settings In this embodiment, accuracy, macro-average precision, macro-average recall, and macro-average F1 score are used to evaluate the diagnostic performance of each model, and the number of parameters that need to be determined through training in each model is counted. The macro-average precision, macro-average recall, and macro-average F1 score are determined by calculating the precision, recall, and F1 score for each state category separately, and then taking the arithmetic mean of the corresponding metrics for all state categories, thereby reducing the impact of differences in the number of samples for each state category on the evaluation results. The diagnostic results of each model on the same test set are as follows: Figure 3 As shown.

[0089] Depend on Figure 3 It can be seen that when the original vibration signal is processed directly using CNN, the accuracy is 0.9628 and the macro-average F1 value is 0.9624; after using VMD-CNN, the accuracy is improved to 0.9804 and the macro-average F1 value is improved to 0.9803. This shows that decomposing the vibration signal into modal components with relatively concentrated frequency structure is beneficial to reducing the influence of non-stationary noise and irrelevant frequency components on convolution feature extraction.

[0090] The method of this invention achieves an accuracy of 0.9954, a macro-average precision of 0.9957, a macro-average recall of 0.9950, and a macro-average F1 score of 0.9953. Compared to CNN, the method of this invention improves accuracy by 0.0326 and macro-average F1 score by 0.0329; compared to VMD-CNN, the method of this invention improves both accuracy and macro-average F1 score by 0.0150; and compared to VMD-CNN-BiLSTM, the method of this invention improves both accuracy and macro-average F1 score by 0.0053. These results demonstrate that joint evaluation of relative energy, kurtosis, and correlation coefficient, effective modality selection, multi-scale convolution, and channel attention can synergistically improve the discriminability between different bearing state features.

[0091] The accuracy of VMD-CNN-BiLSTM-CBAM is 0.9970, which is 0.0016 higher than the method of this invention, but the number of parameters is 1.41M. The number of parameters of the method of this invention is 0.93M, which is 0.48M less than that of the complex hybrid model, a reduction of approximately 34.04%. This shows that, under the experimental conditions of this embodiment, the method of this invention can achieve diagnostic performance close to that of a complex hybrid model with fewer trainable parameters, without introducing a bidirectional long short-term memory network and a complete convolutional block attention module.

[0092] To further determine the contribution of each technical module to the diagnostic results, ablation experiments were conducted under the aforementioned dataset, data partitioning, and training conditions.

[0093] In this embodiment, experiments A1 through A5 are set up. Experiment A1 uses fixed variational mode decomposition parameters and a CNN to verify the effect of adaptive parameter selection; Experiment A2 uses adaptive variational mode decomposition but does not perform effective mode selection, and inputs all mode components into a CNN after superposition to verify the effect of effective mode selection; Experiment A3 uses adaptive variational mode decomposition and effective mode selection, but uses single-scale one-dimensional convolution to verify the effect of multi-scale convolution; Experiment A4 uses adaptive variational mode decomposition, effective mode selection, and multi-scale one-dimensional convolution, but does not use channel attention to verify the effect of channel attention; Experiment A5 uses the complete method of this invention.

[0094] Ablation test results as follows Figure 4 As shown. By Figure 4 It can be seen that when the VMD parameters are fixed in Experiment A1, the accuracy is 0.9836, which is 0.0118 lower than that of the complete method. This indicates that the selection of the number of candidate modes and the candidate penalty factor will affect the quality of mode decomposition. Constructing a comprehensive evaluation value using relative energy, kurtosis and correlation coefficient can improve the parameter selection results.

[0095] Experiment A2 retains adaptive VMD but does not perform effective mode screening, with an accuracy of 0.9872, which is 0.0082 lower than the complete method. This indicates that low-energy or low-correlation modes may contain noise or redundant components. Screening effective modes by using energy thresholds and correlation coefficient thresholds is beneficial to increasing the proportion of fault-related components in the reconstructed vibration signal.

[0096] Experiment A3, using single-scale one-dimensional convolution, achieved an accuracy of 0.9905, 0.0049 lower than the complete method, indicating that convolutional kernels of different lengths can respond to features at different time scales, such as local shocks, short-period oscillations, and periodic modulation. Experiment A4, without channel attention, achieved an accuracy of 0.9921, 0.0033 lower than the complete method, demonstrating that generating channel attention weights based on channel descriptors can enhance fault-related feature channels and suppress redundant feature channels.

[0097] The above ablation experiments were conducted under the same data partitioning and training conditions, with comparisons obtained by changing the corresponding functional modules. Experimental results show that the diagnostic performance of the method in this invention does not solely stem from increasing the size of the convolutional network, but rather from the synergistic effect between adaptive VMD parameter selection, effective modality screening, multi-scale feature extraction, and channel attention weighting.

[0098] The experimental data in Example 11 are used to illustrate the feasibility and technical effects of the present invention under specific publicly available bearing data and specific training conditions, and do not limit the present invention to the bearing data from Case Western Reserve University, the 12kHz sampling frequency, the 1024 sampling points, the 10 state categories, or the specific parameters listed in Tables 1 to 3. In practical applications, the preset sample length, candidate parameter set, number of state categories, and convolutional neural network parameters can be adjusted according to the structure, rotational speed, sampling frequency, fault type, and processor computing power of the rotating machinery to be diagnosed.

[0099] At different sampling frequencies, the sample length can be adjusted to ensure that a single sample covers at least one preset fault feature period. Under different diagnostic tasks, state categories can be set according to normal state, fault location, fault severity, or fault combination, and the number of output nodes of the Softmax classifier can be adjusted accordingly. The technical solution of this invention can be implemented as long as the target VMD parameters are jointly determined using relative energy, kurtosis, and correlation coefficients, effective modes are screened using energy and correlation, and feature extraction and classification are performed using multi-scale convolution and channel attention.

[0100] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A convolutional neural network optimization method based on adaptive mode decomposition, characterized in that, include: Step S1: Collect the raw vibration signal of the rotating machinery using a vibration sensor and preprocess it to obtain a vibration signal sample; Step S2: Decompose the vibration signal sample using each candidate parameter combination consisting of candidate mode number and candidate penalty factor to obtain the corresponding modal components; Step S3: Calculate the relative energy, kurtosis and correlation coefficient of each modal component, and weight the mean values ​​of each index corresponding to the same candidate parameter combination to obtain the comprehensive evaluation value. The candidate parameter combination with the largest comprehensive evaluation value is determined as the target parameter combination. Step S4: From the modal components corresponding to the target parameter combination, determine the modal components with relative energy not less than the energy threshold and correlation coefficient not less than the correlation coefficient threshold as effective modal components, and superimpose them to obtain the reconstructed vibration signal; Step S5: Extract scale branch features of the reconstructed vibration signal through multiple one-dimensional convolutional branches with different kernel lengths, align and splice the scale branch features into multi-scale fusion features, perform global max pooling on each feature channel along the time dimension, and generate channel attention weights based on the pooling results. Step S6: Use channel attention weights to weight the multi-scale fusion features and determine the state category of rotating machinery based on the obtained channel enhancement features.

2. The convolutional neural network optimization method based on adaptive mode decomposition as described in claim 1, characterized in that, Step S1 includes: Calculate the mean and standard deviation of the original vibration signal, and then standardize it with zero mean and unit variance based on the mean and standard deviation to obtain the standardized vibration signal; The standardized vibration signal is sliced ​​according to the preset sample length to obtain multiple vibration signal samples.

3. The convolutional neural network optimization method based on adaptive mode decomposition as described in claim 1, characterized in that, Step S2 includes: Read the pre-stored set of candidate modal numbers and candidate penalty factors; The candidate modes in the candidate mode number set are combined with the candidate penalty factors in the candidate penalty factor set to obtain the candidate parameter combinations; The candidate parameter combinations are traversed, and variational mode decomposition is performed on the same vibration signal sample using each candidate parameter combination to obtain the modal components corresponding to each candidate parameter combination.

4. The convolutional neural network optimization method based on adaptive mode decomposition as described in claim 1, characterized in that, Step S3 includes: The first The ratio of the sum of squares of the amplitudes of the first modal component to the sum of squares of the amplitudes of the vibration signal samples is determined as the ratio of the sum of squares of the amplitudes of the second modal component to the sum of squares of the amplitudes of the third modal component. The relative energy of each modal component; The first The ratio of the fourth-order central moment of each modal component to its fourth standard deviation is determined as the... Kurtosis of each modal component; According to the The covariance and standard deviation between each modal component and the vibration signal sample are used to determine the first modal component. The correlation coefficients between each modal component and the vibration signal sample, where... This represents the modal component number.

5. The convolutional neural network optimization method based on adaptive mode decomposition as described in claim 4, characterized in that, Step S3 also includes: Calculate the arithmetic mean of relative energy, arithmetic mean of kurtosis, and arithmetic mean of correlation coefficient for each modal component corresponding to the same candidate parameter combination; Read the three pre-stored evaluation weights and calculate the comprehensive evaluation value according to the following formula: ; in, The number of candidate modes, As a candidate penalty factor, , and In order, they are the arithmetic mean of relative energy, the arithmetic mean of kurtosis, and the arithmetic mean of correlation coefficient. , and For evaluation weights; The candidate parameter combination with the largest comprehensive evaluation value is determined as the target parameter combination, and the number of candidate modes and the candidate penalty factor in it are determined as the target number of modes and the target penalty factor, respectively.

6. The convolutional neural network optimization method based on adaptive mode decomposition as described in claim 1, characterized in that, Step S4 includes: Read the pre-stored energy threshold and correlation coefficient threshold; The relative energy of each modal component corresponding to the target parameter combination is compared with the energy threshold, and the correlation coefficient of each modal component is compared with the correlation coefficient threshold. Modal components with relative energy not less than the energy threshold and correlation coefficient not less than the correlation coefficient threshold are defined as effective modal components.

7. The convolutional neural network optimization method based on adaptive mode decomposition as described in claim 6, characterized in that, Step S4 also includes: Modal filtering flags are set according to whether each modal component is a valid modal component. The modal filtering flag for valid modal components is 1, and the modal filtering flag for other modal components is 0. Each modal component is multiplied by its corresponding modal screening marker, and the multiplication results are superimposed according to the sampling point positions to obtain the reconstructed vibration signal.

8. The convolutional neural network optimization method based on adaptive mode decomposition as described in claim 1, characterized in that, Step S5 includes: The reconstructed vibration signal is input in parallel into the first-scale one-dimensional convolution branch, the second-scale one-dimensional convolution branch, and the third-scale one-dimensional convolution branch, with the kernel length of the three one-dimensional convolution branches increasing sequentially. The reconstructed vibration signal is convolved and nonlinearly activated by three one-dimensional convolutional branches to obtain the corresponding scale branch features. The features of the three scale branches are aligned along the time dimension and then spliced ​​along the channel dimension to obtain multi-scale fused features.

9. The convolutional neural network optimization method based on adaptive mode decomposition as described in claim 8, characterized in that, Steps S5 and S6 include: Global max pooling is performed along the time dimension on each feature channel of the multi-scale fused feature to obtain the corresponding channel descriptor; The channel attention weights for each feature channel are generated based on the channel descriptors and the channel attention parameters obtained during training. The attention weights of each channel are multiplied by their corresponding feature channels to obtain the channel-enhanced features; The channel enhancement features are input into the Softmax classifier to obtain the state category probability of each preset rotating machinery state category, and the rotating machinery state category corresponding to the maximum state category probability is determined as the diagnostic result. The network parameters of the one-dimensional convolutional branch, the channel attention parameters, and the classification parameters of the Softmax classifier are obtained by training with vibration signal training samples with state category labels and the cross-entropy loss function.

10. A convolutional neural network optimization system based on adaptive mode decomposition, characterized in that, include: Vibration sensors are installed on rotating machinery to collect raw vibration signals during the operation of the rotating machinery. The memory is used to store the computer program, the set of candidate modal numbers, the set of candidate penalty factors, the evaluation weights, the energy threshold, the correlation coefficient threshold, and the parameters of the convolutional neural network. as well as A processor, connected to a vibration sensor and a memory, is used to execute a computer program to perform the rotating machinery fault diagnosis method according to any one of claims 1 to 9 on the raw vibration signal and output the rotating machinery state category.

Citation Information

Patent Citations

  • One-dimensional convolutional neural network rolling bearing fault diagnosis method fusing channel attention

    CN115034267A

  • Bearing fault diagnosis method based on data layer feature fusion and convolutional neural network

    CN116186590A