A multilayer collaborative denoising method for intelligent diagnosis of swash plate type axial piston pump failure

Abstract

The application discloses a multilayer collaborative denoising method for intelligent fault diagnosis of swash plate type axial piston pump. In view of the defects of insufficient denoising of non-stationary signals, modal aliasing and dynamic mode decomposition (DMD) susceptible to noise interference in the prior art, an integrated feature extraction scheme of deep coupling of variational mode decomposition (VMD) and wavelet packet layered collaborative denoising and DMD is provided. The method adaptively decomposes the original signal into a plurality of intrinsic mode components (IMF) through VMD, and selects effective components based on a correlation coefficient; then, each effective IMF is finely adaptively soft threshold denoised by using a wavelet packet; finally, the reconstructed high signal-to-noise ratio pure signal is input into DMD for dynamic feature analysis. The application can effectively suppress strong background noise, accurately extract fault features and significantly improve fault recognition precision, and can be widely applied to intelligent fault diagnosis and condition monitoring of rotary machines such as axial piston pumps.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment fault diagnosis technology, specifically to a signal processing method, and more particularly to a multi-layer collaborative noise reduction method for intelligent fault diagnosis of swashplate axial piston pumps, which is applicable to the condition monitoring and fault diagnosis of rotating machinery under strong noise and non-steady operating conditions. Background Technology

[0002] As a core power component of hydraulic systems, swashplate axial piston pumps are widely used in engineering machinery, aerospace, and shipbuilding industries due to their advantages such as high pressure, large flow rate, and high efficiency. However, their operating conditions are complex, involving long-term exposure to high pressure, high speed, and heavy loads. Key components such as the slippers, cylinder block, and distributor plate are prone to wear, fatigue, and other failures. Therefore, real-time monitoring of their operating status and intelligent fault diagnosis are crucial for ensuring equipment safety and preventing significant economic losses.

[0003] In the engineering practice of fault diagnosis for swashplate axial piston pumps, the vibration sensing signals acquired in the field typically possess inherent characteristics of nonlinear modulation, non-stationary time-varying nature, and strong background noise interference. The signals are mixed with various interference components, including environmental electromagnetic noise, random mechanical impacts, hydraulic fluid pulsations, and harmonics from the transmission pair meshing. These interferences cause weak fault features to be submerged by strong noise, resulting in severe mode aliasing within the signal. Traditional signal processing methods struggle to effectively extract fault features, significantly reducing fault identification accuracy.

[0004] Currently, existing technologies for processing this type of signal mostly employ a single-mode processing approach, which has the following main drawbacks:

[0005] 1. Variational mode decomposition (VMD) or empirical mode decomposition (EMD) is used alone to separate the components of the original signal, but no effective hierarchical noise reduction is configured, resulting in a large amount of residual noise in the decomposed components.

[0006] 2. While using wavelet denoising, mean filtering, adaptive filtering, and other methods alone to suppress interference can remove some noise, it can easily cause the loss of effective signal details and lead to mode aliasing problems.

[0007] 3. Directly inputting the original signal without noise reduction or with insufficient noise reduction into Dynamic Mode Decomposition (DMD) for feature calculation results in noise directly participating in the matrix solution process, leading to severe distortion of the extracted modal features, high feature redundancy, and low discriminative power.

[0008] In summary, the three stages of decomposition, noise reduction, and feature extraction in the existing technology are independent of each other and lack a collaborative optimization mechanism. It is difficult to balance noise suppression, detail preservation, and feature accuracy under strong noise and complex working conditions, and thus cannot meet the high reliability requirements of intelligent fault diagnosis for high-end equipment. Summary of the Invention

[0009] To address the aforementioned problems in existing technologies, this invention provides a multi-layer collaborative noise reduction method for intelligent fault diagnosis of swashplate axial piston pumps. This invention constructs an integrated technical solution that combines "variational mode decomposition (VMD) and wavelet packet multi-scale noise reduction in a cascaded collaborative preprocessing manner, deeply coupled end-to-end with dynamic mode decomposition (DMD)."

[0010] A multi-layer collaborative noise reduction method for intelligent fault diagnosis of swashplate axial piston pumps includes the following steps:

[0011] S1: Collect the original vibration signals of the swashplate axial piston pump under various operating conditions, and preprocess the original vibration signals, including outlier removal and amplitude standardization.

[0012] S2: Adaptive decomposition of the preprocessed signal is performed using variational mode decomposition (VMD) to obtain several intrinsic mode components (IMFs), and effective IMFs are selected based on the correlation coefficients between each IMF and the original signal;

[0013] S3: Wavelet packet transform is used to decompose each effective IMF selected in step S2 into multiple scales, and adaptive soft thresholding is used to denoise the high-frequency coefficients after decomposition. Then, the denoised effective IMF is reconstructed.

[0014] S4: Linearly superimpose and reconstruct all the effective IMFs after noise reduction in step S3, and perform phase correction to obtain a clean reconstructed signal with high signal-to-noise ratio.

[0015] S5: Construct a Hankel matrix based on the pure reconstructed signal, perform singular value decomposition (SVD) to truncate the noise subspace, and solve for the eigenvalues ​​and mode vectors of dynamic mode decomposition (DMD) to extract multidimensional dynamic fault features;

[0016] S6: Normalize the multidimensional dynamic fault features to construct a standardized fault feature dataset, and input it into a one-dimensional convolutional neural network (1D-CNN) for fault pattern recognition and output diagnostic results.

[0017] Furthermore, in step S2 of the present invention, the correlation coefficient between each IMF component and the original signal is calculated. and with A threshold of >0.25 is used to remove noise-dominated spurious components and retain valid IMF components that carry fault information.

[0018] Furthermore, in step S3 of the present invention, the sym8 or db4 wavelet basis is used to perform 3-5 levels of full-band wavelet packet decomposition on the effective IMF components; the adaptive soft thresholding The calculation formula is: ,in The noise standard deviation of the wavelet coefficients. This represents the number of signal sampling points.

[0019] Furthermore, in step S5 of the present invention, the multidimensional dynamic fault characteristics include at least modal frequency, damping ratio, and modal energy.

[0020] Furthermore, in step S5 of the present invention, the truncation principle of the singular value decomposition (SVD) is: to retain singular values ​​whose energy percentage reaches a preset ratio (such as 95%) in order to eliminate noise subspaces.

[0021] Furthermore, the outlier removal in step S1 of the present invention employs... The criteria and the decision relation are as follows: ,in This is the original sampled data. The mean of the signal. The standard deviation is the signal standard deviation; the amplitude standardization uses zero-mean unit variance normalization, as shown in the formula: .

[0022] Furthermore, the present invention also includes introducing a block access heat statistics mechanism during the signal block processing to cache the effective signal segments that are multiplexed at high frequencies, so as to avoid repeated calculations.

[0023] Furthermore, the various operating conditions of the present invention include at least four typical operating conditions: normal state, slipper wear, cylinder block wear, and distributor plate wear.

[0024] Compared with the prior art, the present invention has the following significant advantages:

[0025] 1. Thorough noise reduction and modal purity: This invention employs a hierarchical collaborative noise reduction mechanism combining VMD and wavelet packets. VMD decomposes the signal into physically meaningful independent components at the source, avoiding modal aliasing; wavelet packets then perform full-band fine noise reduction on each component. The synergistic effect of these two mechanisms significantly improves the purity and signal-to-noise ratio of non-stationary signals, providing a high-quality data foundation for subsequent feature extraction.

[0026] 2. Precise features and high discriminative power: By using a clean signal with a high signal-to-noise ratio as input to the DMD, the interference of noise on modal calculation is fundamentally eliminated, making the extracted dynamic modal features (such as modal frequencies and damping ratios) more accurate and stable, with low feature redundancy and high discriminative power. Experiments show that the feature space contour coefficients processed by this method are significantly improved, and the clustering effect is excellent.

[0027] 3. Robust and adaptable: This invention is particularly suitable for high noise, nonlinear and non-stationary signal scenarios. It not only performs well in axial piston pump fault diagnosis (the fault classification accuracy can reach more than 99% on measured data with a signal-to-noise ratio of 10dB), but can also be extended to other fields such as rotating machinery condition monitoring, power signal analysis, and acoustic signal processing.

[0028] 4. Link collaboration and efficiency optimization: This invention constructs a complete integrated pipeline from acquisition to diagnosis and introduces a block adaptive caching mechanism to optimize computing efficiency, with good engineering real-time performance and deployment potential. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall process of the multi-layer collaborative noise reduction method for intelligent fault diagnosis of swashplate axial piston pumps provided by the present invention.

[0030] Figure 2 This is a comparison diagram of the signal waveforms processed using the method of this invention and other noise reduction methods.

[0031] Figure 3 This is a schematic diagram of the DMD modal feature results extracted using the method of the present invention.

[0032] Figure 4 This is a schematic diagram of the confusion matrix obtained after fault classification using the method of the present invention.

[0033] Figure 5 This is a comparison chart of the quality quantification of feature space clustering before and after processing using the method of this invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0035] In practical engineering scenarios for fault diagnosis of swashplate axial piston pumps, the original vibration sensing signals acquired on-site generally possess inherent characteristics such as nonlinear modulation, non-stationary time-varying nature, and strong background noise interference. The signals are also contaminated with various redundant interference components, including environmental electromagnetic noise, random mechanical impact interference, hydraulic fluid pulsation noise, and periodic meshing harmonic interference from the transmission pairs. These interferences directly cause engineering challenges such as weak fault features being submerged by strong noise, severe aliasing and overlap of internal signal modes, distortion and degradation of traditional algorithm feature extraction, and a significant reduction in fault identification accuracy.

[0036] Most existing traditional signal processing technologies adopt a discrete processing architecture with independent computation of single mode decomposition and separate execution of single filtering and noise reduction. The three stages of decomposition, noise reduction and feature extraction are independent and isolated from each other, and a full-link collaborative optimization coupling mechanism has not been established. It is impossible to simultaneously take into account the three core requirements of thorough noise suppression, complete details of weak faults and accurate dynamic mode analysis under strong noise and complex working conditions, which makes it difficult to meet the high-precision engineering requirements of intelligent fault diagnosis for high-end equipment.

[0037] To address the aforementioned technical bottlenecks, this invention proposes a cascaded collaborative preprocessing architecture combining Variational Mode Decomposition (VMD) and wavelet packet multi-scale denoising. It deeply couples the preprocessed clean output signal with Dynamic Mode Decomposition (DMD) end-to-end, constructing a complete, integrated, high-precision dynamic feature extraction pipeline for non-stationary time-varying signals. The overall technical approach of this invention strictly follows a progressive execution logic of signal decoupling → multi-scale hierarchical denoising → clean signal reconstruction → dynamic mode analysis. Simultaneously, drawing on the concept of transparent mapping management of embedded serial memory addresses, the control logic for block-ordered transmission, progressive caching optimization, and redundant component removal is migrated to the signal processing link. This achieves integrated automated execution of noise suppression, component selection, and feature extraction throughout the entire process, effectively reducing computational redundancy and improving the stability of the link operation.

[0038] The core technological innovation of this invention lies in the organic integration of the advantages of VMD adaptive modal decoupling and separation capabilities, wavelet packet full-band refined noise reduction and preservation capabilities, and DMD data-driven high-dimensional dynamic feature mining capabilities, constructing a highly robust cascaded signal processing link. This can transform the original non-stationary signal, characterized by strong noise and modal aliasing, into a standardized feature vector that is modally pure, feature-clear, and directly applicable to intelligent fault diagnosis of swashplate axial piston pumps.

[0039] This invention provides a multi-layer collaborative noise reduction method for intelligent fault diagnosis of swashplate axial piston pumps. It consists of five functional modules arranged in a serial data stream relationship: a signal preprocessing module, a VMD adaptive decomposition module, a wavelet packet hierarchical noise reduction module, a noise-reduced signal reconstruction module, and a DMD dynamic feature analysis module. The modules are connected in a serial data stream relationship, with the output of one module serving as the sole input to the next, ensuring consistency and reproducibility of signal processing.

[0040] 1) Signal preprocessing module: Completes outlier removal, centering, standardization, and dimension alignment;

[0041] 2) VMD Adaptive Decomposition Module: Completes multi-component decoupling of the signal, spurious component filtering, and effective component retention;

[0042] 3) Wavelet packet hierarchical denoising module: completes IMF component multi-scale decomposition, adaptive threshold denoising, and coefficient reconstruction;

[0043] 4) Noise reduction signal reconstruction module: Completes effective component weighted fusion, signal phase correction, and clean signal output;

[0044] 5) DMD Dynamic Feature Analysis Module: Completes spatiotemporal matrix construction, SVD decomposition, mode solving, feature selection and output.

[0045] Data interaction between modules strictly follows the process: original signal → preprocessing → VMD decomposition → wavelet packet denoising → signal reconstruction → DMD analysis → feature output, without introducing additional noise or human interference at any stage.

[0046] Please see Figure 1 This invention addresses the non-stationary vibration signals of axial piston pumps by employing a hierarchical, progressive serial processing chain, as shown in the attached diagram. The entire process consists of 9 consecutive steps, with each preceding output serving as the sole input for the following step. There are no parallel branches or manual intervention, resulting in high reproducibility and stability. The specific implementation steps are as follows:

[0047] Step 1: Acquisition of raw vibration signals under multiple operating conditions

[0048] For four typical operating conditions of swashplate axial piston pumps—normal condition, slipper wear, cylinder block wear, and distributor plate wear—raw time-series vibration signals are collected using vibration sensors, and the sampling frequency is set. Number of sampling points per sample A raw vibration signal dataset containing corresponding fault labels is constructed as the initial input for the entire process. : Sampling frequency; N: Number of sampling points for a single signal.

[0049] Step 2: Signal Preprocessing

[0050] Outlier removal and amplitude normalization were performed on the original vibration signal: [The following is a list of methods used in the original text, which is not translated as it's not part of the main text.] Criteria, based on decision relations Remove outlier abrupt changes caused by acquisition jitter and electromagnetic interference, and then... Zero-mean unit variance normalization eliminates dimensional differences and DC offset, resulting in a normalized output signal. Where x: original sampled data; μ: signal mean; σ: signal standard deviation; : Standardized signal.

[0051] Step 3: VMD Adaptive IMF Decomposition

[0052] Initialize the core VMD parameters such as the number of modal components K, penalty factor α, and convergence accuracy. Adaptively decompose the preprocessed non-stationary coupled signal into several independent, non-modal aliasing intrinsic mode components (IMFs). Calculate the correlation coefficient between each component and the original signal. ,by A threshold greater than 0.25 is used to remove noise-dominated spurious components, retaining only valid IMF components carrying fault information. Where K: total number of modes; α: VMD penalty coefficient; : Correlation coefficient of the k-th component.

[0053] Step 4: Wavelet Packet Multi-Scale Denoising

[0054] Selecting orthogonal wavelet bases and an appropriate number of decomposition levels, full-band wavelet packet decomposition is performed on all selected effective IMF components; according to... An adaptive soft threshold is calculated, and high-frequency noise figures are suppressed by contraction, while simultaneously preserving high- and low-frequency fault details, thus completing refined noise reduction processing for each IMF component. Noise reduction threshold; : Noise standard deviation of wavelet coefficients.

[0055] Step 5: Noise Reduction Signal Reconstruction

[0056] according to The effective IMF components after noise reduction are linearly superimposed and reconstructed according to the original time sequence. Simultaneously, phase distortion and time-domain glitches caused by decomposition and noise reduction are corrected, resulting in a pure vibration signal with high signal-to-noise ratio, prominent fault characteristics, and no mode mixing. : IMF components after single-channel noise reduction; The reconstructed pure signal.

[0057] Step 6: Hankel Matrix Construction and DMD Feature Extraction

[0058] A two-dimensional Hankel delay observation matrix is ​​constructed using a clean, denoised signal. Singular value decomposition is performed on the matrix, and the noise subspace is truncated according to the energy principle. A dynamic approximation operator is constructed based on principal components to solve for DMD eigenvalues ​​and mode vectors. Furthermore, multi-dimensional dynamic fault features such as modal frequency, damping ratio, and modal energy are calculated to complete feature extraction.

[0059] Step 7: Feature Standardization and Dataset Partitioning

[0060] use The extracted high-dimensional features of the DMD are normalized to eliminate amplitude differences between different feature dimensions; the training and test sets are randomly divided according to a fixed ratio to maintain a balanced distribution of fault categories, thus constructing a standardized fault feature dataset. Where F represents the original feature; Normalization characteristic; , : Eigenvalues.

[0061] Step 8: One-dimensional convolutional feature classification and recognition

[0062] The standardized feature dataset is input into the constructed one-dimensional convolutional neural network model, and then processed through convolution extraction, pooling dimensionality reduction, and fully connected mapping. The system categorizes and intelligently identifies fault modes for four types of axial piston pump operating conditions.

[0063] Step 9: Fault Diagnosis Output and Performance Evaluation

[0064] Relying on The classification accuracy is calculated, and the diagnostic results are quantitatively evaluated using metrics such as confusion matrix and precision. The advantages of this invention in noise reduction, feature extraction accuracy, and diagnostic performance improvement are verified by comparing it with traditional processing methods. Finally, a complete fault diagnosis conclusion and performance report are output. Accuracy; Correct sample; Total test samples.

[0065] To make the technical solution and effects of the present invention clearer, a specific embodiment is provided below in the scenario of slipper wear failure in a swashplate axial piston pump (signal-to-noise ratio). ):

[0066] 1. Data Acquisition: Select measured vibration data from a prototype with slipper wear fault, and set f , Measured signal-to-noise ratio of the original signal .

[0067] 2. Preprocessing: Calculate the signal mean. , ,in accordance with Two outlier points were removed, and all data were standardized.

[0068] 3. VMD Mode Decomposition: The number of modes K is set to 6, and the penalty factor α is set to 3000. The correlation coefficients of the 6 components are obtained. According to the threshold >0.25, filter and retain the first 4 valid IMF components.

[0069] 4. Wavelet packet layered noise reduction: Selected Wavelet basis, 4-level decomposition, and calculation of soft thresholding. And perform soft threshold noise reduction on each effective component.

[0070] 5. Signal Reconstruction: The four sets of denoised IMF components are superimposed to generate the reconstructed signal. After phase correction, the signal-to-noise ratio of the reconstructed signal was significantly improved to 22.3 dB. Please refer to [link / reference]. Figure 2The comparison of signal waveforms under different noise reduction methods shows that the combined noise reduction scheme of the present invention can effectively filter out background noise and completely preserve the fault impact characteristics.

[0071] 6. DMD Feature Extraction: A Hankel matrix is ​​constructed, and SVD is performed with 95% energy truncation to extract 10 dimensions of features, including modal frequencies, damping, and energy. Please refer to [link to relevant documentation]. Figure 3 The DMD modal feature extraction results effectively characterize the dynamic characteristics of different fault states.

[0072] 7. Dataset partitioning: Normalize the features to the 0~1 range and split them into 2100 training samples and 900 test samples in a 7:3 ratio.

[0073] 8. Fault classification: After training with a one-dimensional convolutional network, the overall classification accuracy of the test set reached 99.67%.

[0074] 9. Performance Evaluation: Please refer to [link / reference]. Figure 4 The confusion matrix of the fault classification results, combined with quantitative evaluation of indicators such as accuracy, verifies the fault identification accuracy and stability of this scheme under strong noise conditions.

[0075] 10. Feature Clustering Effect Analysis: This embodiment also uses the average silhouette coefficient to evaluate feature quality (coefficient range [-1, 1], the larger the value, the better the discrimination). Please refer to [link to relevant documentation]. Figure 5 The contour coefficient of the original signal features was only -0.1098, indicating severe feature overlap; however, after processing by this method, the contour coefficient was increased to 0.6201, significantly improving the clustering differentiation effect.

[0076] Regarding the implementation of key technologies, this invention also includes the following core algorithms:

[0077] 1. VMD decomposition without aliasing constraint technology: By adaptively adjusting the penalty factor and the number of modes, the bandwidth of each IMF component is minimized and the center frequency is discretized. This avoids the mode aliasing, endpoint effect and spurious component problems that are easy to occur in traditional EMD, and ensures that the physical meaning of the subsequent noise reduction object is clear.

[0078] 2. Wavelet Packet Adaptive Soft Threshold Noise Reduction Technology: Unlike the limitation of traditional wavelets that only decompose the low-frequency part, wavelet packets decompose both high-frequency and low-frequency components at the full scale and use an adaptive threshold based on statistical characteristics to achieve precise noise suppression while preserving weak fault characteristics.

[0079] 3. Lightweight DMD parsing technology based on SVD truncation: It adopts economical SVD to reduce computational complexity, removes noise subspace by singular value energy truncation, and retains only the principal components of the signal, so that DMD can run quickly on embedded platforms and improve the anti-interference of features.

[0080] 4. Block Adaptive Caching and Frequency Optimization Mechanism: During the iterative processing of signal blocks, a block access heat statistics mechanism is introduced to cache the effective signal segments that are reused at high frequencies. This effectively avoids the repeated execution of VMD decomposition and wavelet packet denoising operations, and significantly improves the efficiency of the entire signal processing process.

[0081] In addition, the present invention provides a variety of alternative implementation schemes to adapt to different application scenarios and hardware platforms:

[0082] 1. Alternative noise reduction solutions: In scenarios where wavelet packets cannot be used, variational thresholding or singular spectrum noise reduction can be used instead, while still retaining the overall architecture of VMD+DMD.

[0083] 2. Alternative decomposition scheme: On low computing power platforms, VMD can be replaced by LMD local mean decomposition, but the effective component screening logic based on the correlation coefficient must be retained.

[0084] 3. Alternative feature schemes: DMD can be replaced with singular value features, entropy features, or time-frequency domain statistical features, but the preprocessing architecture for joint denoising remains unchanged.

[0085] All alternatives retain the core innovation of this invention: the combined noise reduction preprocessing and clean signal-driven dynamic feature extraction.

[0086] The method of this invention has good compatibility in terms of hardware and software support environments. The hardware platform supports PCs, ARM embedded processors, and FPGA signal processing boards; the software environment supports MATLAB 2020b and above, Python 3.8+, PyTorch / TensorFlow, etc.; the signal types supported for processing include, but are not limited to, non-stationary time-series signals such as vibration signals, pressure signals, acoustic signals, and electrical signals.

[0087] In summary, the multi-layer collaborative noise reduction method provided by this invention effectively solves the engineering problem of accurately extracting fault characteristics of swashplate axial piston pumps under strong noise background, and has extremely high industrial application value and promotion prospects.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-layer collaborative noise reduction method for intelligent fault diagnosis of swashplate axial piston pumps, characterized in that, Includes the following steps: S1: Collect the original vibration signals of the swashplate axial piston pump under various operating conditions, and preprocess the original vibration signals, including outlier removal and amplitude standardization. S2: Adaptive decomposition of the preprocessed signal is performed using variational mode decomposition to obtain several intrinsic mode components, and effective intrinsic mode components are selected based on the correlation coefficients between each IMF and the original signal. S3: Wavelet packet transform is used to decompose each effective IMF selected in step S2 into multiple scales, and adaptive soft thresholding is used to denoise the high-frequency coefficients after decomposition. Then, the denoised effective intrinsic mode components are reconstructed. S4: Linearly superimpose and reconstruct all the effective intrinsic mode components after noise reduction in step S3, and perform phase correction to obtain a clean reconstructed signal with high signal-to-noise ratio. S5: Construct a Hankel matrix based on the pure reconstructed signal, perform singular value decomposition to truncate the noise subspace, and solve for the eigenvalues ​​and mode vectors of the dynamic mode decomposition to extract multidimensional dynamic fault features; S6: Normalize the multidimensional dynamic fault features to construct a standardized fault feature dataset, and input it into a one-dimensional convolutional neural network for fault pattern recognition, and output the diagnostic results.

2. The method according to claim 1, characterized in that, In step S2, the correlation coefficient between each IMF component and the original signal is calculated. and with A threshold of >0.25 is used to remove noise-dominated spurious components and retain the valid intrinsic mode components that carry fault information.

3. The method according to claim 1, characterized in that, In step S3, the effective intrinsic mode components are decomposed into 3-5 levels of full-band wavelet packets using the sym8 or db4 wavelet basis; the adaptive soft thresholding... The calculation formula is: ,in The noise standard deviation of the wavelet coefficients. This represents the number of signal sampling points.

4. The method according to claim 1, characterized in that, In step S5, the multidimensional dynamic fault characteristics include at least modal frequency, damping ratio, and modal energy.

5. The method according to claim 1, characterized in that, In step S5, the truncation principle of the singular value decomposition is to retain singular values ​​whose energy proportion reaches a preset ratio in order to eliminate the noise subspace.

6. The method according to claim 1, characterized in that, The outlier removal in step S1 adopts... The criteria and the decision relation are as follows: ,in This is the original sampled data. The mean of the signal. The standard deviation is the signal standard deviation; the amplitude standardization uses zero-mean unit variance normalization, as shown in the formula: .

7. The method according to claim 1, characterized in that, The method also includes introducing a block access heat statistics mechanism during the signal block processing to cache the effective signal segments that are frequently reused, so as to avoid repeated calculations.

8. The method according to claim 1, characterized in that, The various operating conditions include at least four typical conditions: normal condition, slipper wear, cylinder block wear, and distributor plate wear.

Images