A rotating machinery fault diagnosis method and system based on TISRNet
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN UNIV OF SCI & TECH
- Filing Date
- 2026-02-06
- Publication Date
- 2026-06-19
Smart Images
Figure CN122241050A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of rotating machinery fault diagnosis technology, specifically relating to a rotating machinery fault diagnosis method and system based on TISRNet. Background Technology
[0002] Rotating machinery, as a key component of modern industrial equipment and energy systems, plays an irreplaceable role in equipment drive systems. Its stable and efficient operation is crucial for ensuring production safety and improving economic benefits. As a critical supporting structure of rotating machinery, rolling bearings not only undertake important functions such as reducing friction, supporting the rotor, and maintaining motion accuracy, but are also among the most prone to failure. Statistics show that approximately 30% of rotating machinery failures originate from bearing failure, while in wind turbines, about 80% of gearbox failures are related to bearing problems. Taking large wind turbines as an example, as an important unit for clean energy power generation, they play an increasingly important role in the global energy structure transformation. However, wind turbines typically operate under extremely complex outdoor conditions. Affected by the operating environment and load fluctuations, rolling bearings are highly susceptible to failure due to fatigue, wear, or fracture, thus threatening the safety and reliability of the entire system.
[0003] Existing methods for fault diagnosis of rotating machinery mainly fall into two categories: signal processing and deep learning. Traditional signal processing methods extract time-frequency features and suppress noise through wavelet transform, empirical mode decomposition, and envelope spectrum analysis, but their performance depends on manual parameter settings and they lack adaptability to complex operating conditions. In recent years, end-to-end diagnostic methods based on convolutional neural networks have been extensively studied.
[0004] Existing methods for training improved DRSN (Deep Residual Shrinkage Network) on wavelet time-frequency images can improve the accuracy of fault identification in noisy environments. Although this approach achieves certain results in feature extraction and noise resistance, its signal enhancement and classification processes remain independent and lack physical prior constraints, thus the physical interpretability and robustness of the model still need improvement.
[0005] However, the above scheme has two major drawbacks: (1) The signal enhancement process mainly relies on data-driven or empirical filtering, lacks a constraint mechanism based on physical models such as energy operators, and is difficult to effectively enhance the transient energy changes of impact faults, resulting in the feature extraction stage being still relatively sensitive to noise; (2) Signal enhancement and fault diagnosis are mostly run as independent modules, and the enhancement process fails to be optimized in coordination with the classification target, which limits the generalization ability of the model under different noise levels and operating conditions. Summary of the Invention
[0006] The technical problem to be solved by this invention is to provide a rotating machinery fault diagnosis method and system based on TISRNet, which can simultaneously optimize noise suppression, feature extraction and classification performance.
[0007] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: a method for diagnosing rotating machinery faults based on TISRNet, comprising the following steps: S1: Collect vibration signals of rotating machinery under different operating conditions; S2: Construct a TISRNet model including a data preprocessing module, a trainable signal enhancement module, and a fault diagnosis backbone network; input vibration signals and perform fault diagnosis; specifically including: S21: The vibration signal is divided into datasets and subjected to sliding sampling through the data preprocessing module; S22: Adaptive enhancement of the sampled signal guided by physical information is performed through a trainable signal enhancement module, and an enhanced signal is output. S23: Deep feature extraction and fault identification of enhanced signals are performed through the fault diagnosis backbone network; S3: The TISRNet model is trained using a joint optimization strategy of primary and secondary tasks; S4: Test the trained TISRNet model, determine the fault category based on the maximum probability principle, and output the probability distribution of each type of fault.
[0008] According to the above scheme, the specific steps in step S21 are as follows: The collected vibration signals were divided into training and testing sets; The sliding window length is determined based on the bearing fault characteristic frequency, and the divided data is sampled using the sliding window.
[0009] According to the above scheme, the specific steps in step S22 are as follows: S221: Extract the instantaneous energy features of the sampled signal through the TEO energy operator, and stabilize the feature distribution through batch normalization (BN). S222: The instantaneous energy features are extracted using multiple sets of parallel one-dimensional convolutions with different kernel sizes to extract the impact and modulation features at different time scales; the feature maps output by each convolution branch are spliced together in the channel dimension to form a multi-scale feature tensor; S223: Introducing a channel attention mechanism to recalibrate the multi-scale feature tensor, obtaining statistical information of each channel through global average pooling (GAP), and calculating the weight coefficient of each channel through the channel attention mechanism (ECA) to highlight the scale features that contribute more to fault detection. S224: Input the multi-scale features after channel recalibration into a one-dimensional fusion convolutional layer, and perform adaptive linear combination of features of different scales in the time dimension through convolution operation, thereby realizing the dynamic fusion of multi-scale information; S225: Apply a learnable threshold function to the fused features to generate enhanced weights; S226: Multiply the enhanced weights point by point with the preprocessed signal and sum the residuals to output the enhanced signal.
[0010] Furthermore, in step S222, the parallel one-dimensional convolutions with different kernel sizes include short convolution kernels and long convolution kernels; the short convolution kernels are used to capture instantaneous impacts, and the long convolution kernels are used to extract periodic envelope features. use This indicates the TEO output signal. This represents the size of the i-th convolutional kernel. M Indicates the number of convolution kernels. This represents a one-dimensional convolution, where each convolutional branch outputs a feature. f i for: (1); This represents a cascaded operation along the channel dimension, a multi-scale feature tensor. F norm for: (2).
[0011] Furthermore, in step S225, using This represents the learnable parameters shared across all channels and time steps, contributing to the fused features. z Apply a soft thresholding function with a learnable threshold to generate boosting weights. for: (3).
[0012] According to the above scheme, in step S226, using This indicates the preprocessed signal. To enhance weight, A learnable residual scaling factor to enhance the signal. for: (4).
[0013] According to the above scheme, in step S3, the cross-entropy loss is used as the main task loss function to backpropagate the entire model; the spectral entropy loss is used as the auxiliary task loss function to backpropagate only to the TFE module.
[0014] A rotating machinery fault diagnosis system includes a data acquisition module, an end-to-end TISRNet model, a training module, and a testing module; The data acquisition module is used to collect vibration signals of rotating machinery under different operating conditions; The end-to-end TISRNet model comprises a data preprocessing module, a trainable signal enhancement module, and a fault diagnosis backbone network connected sequentially. The model uses a physical information loss function to jointly constrain signal enhancement and classification tasks. The data preprocessing module samples vibration signals and divides them into training and testing sets. The trainable signal enhancement module includes a TEO energy operator, a multi-scale convolution module, a channel recalibration module, a one-dimensional fusion convolutional layer, a soft thresholding function module, and a residual-based enhancement signal generation module, used for adaptive enhancement of the sampled signals guided by physical information, outputting the enhanced signal. The fault diagnosis backbone network includes a one-dimensional convolutional layer, multiple RPSBU units, and a fully connected classification layer, used for deep feature extraction and fault category identification of the enhanced signal. The RPSBU units use shrinking blocks to adaptively soft-threshold the features of the enhanced signal. The soft thresholding function includes a two-layer fully connected network shared across all channels and dynamically adjusted according to the energy distribution of the input signal. The physical information loss function includes a main task loss function and an auxiliary task loss function. The main task loss function uses the cross-entropy loss function to supervise the classification task and optimize the fault identification performance. The auxiliary task loss function introduces an auxiliary loss based on spectral entropy to constrain the energy concentration of the enhanced signal in the frequency domain, so that the signal enhancement process maintains physical consistency with the diagnostic target.
[0015] Furthermore, the TEO energy operator is used to extract the instantaneous energy characteristics of the signal; The multi-scale convolution module includes multiple sets of parallel one-dimensional convolutions with different kernel sizes, used to extract impact and modulation features at different time scales; The channel recalibration module is used to introduce multi-scale feature tensors into the channel attention mechanism for channel recalibration. It obtains statistical information of each channel through global average pooling (GAP) and calculates the sparse weights of each channel through the channel attention mechanism (ECA) to highlight scale features that contribute more to fault detection. One-dimensional fusion convolutional layers are used to enable multi-scale features after channel weighting to automatically learn the weighting coefficients of different scale features at each time step, thereby achieving dynamic fusion of multi-scale information to obtain fused features. The soft thresholding function module is used to apply a learnable threshold to the fused features using a soft thresholding function, generating enhanced weights; The residual-based enhancement signal generation module is used to generate the final enhanced signal output by modulating the preprocessed signal of the enhancement weight domain point by point and combining residual links.
[0016] A computer memory storing a computer program executable by a computer processor, the computer program executing a TISRNet-based method for diagnosing rotating machinery faults.
[0017] The beneficial effects of this invention are as follows: 1. The present invention provides a rotating machinery fault diagnosis method and system based on TISRNet. By constructing an energy operator-guided signal enhancement module and jointly training it with the fault diagnosis network in an end-to-end manner, the consistency between the enhanced features and the classification target is maintained, so that the signal enhancement results can adaptively meet the requirements of the diagnostic task, and the synchronous optimization of noise suppression, feature extraction and classification performance is achieved.
[0018] 2. This invention introduces an energy operator sensitive to instantaneous energy at the network front end, making full use of physical characteristics, effectively highlighting the characteristics of impact-type faults and suppressing background noise, and realizing adaptive signal enhancement under physical prior constraints, thereby significantly improving the model's feature enhancement capability in low signal-to-noise ratio environments.
[0019] 3. This invention effectively avoids the feature bias introduced by traditional staged processing, thereby achieving effective enhancement of impact faults in complex noise environments, significantly improving the robustness, stability and physical interpretability of the model, and thus improving the accuracy of rotating machinery fault identification.
[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 This is a flowchart of an embodiment of the present invention.
[0023] Figure 2 This is a structural diagram of the TFE module according to an embodiment of the present invention.
[0024] Figure 3 This is a structural diagram of the RPSBU according to an embodiment of the present invention.
[0025] Figure 4 This is a flowchart of an embodiment of the present invention.
[0026] Figure 5 This is a test set accuracy graph of each model in the embodiments of the present invention on the WUST dataset.
[0027] Figures 6a to 6e This is a comparison chart of signal enhancement effects under a noise level of SNR=-4dB according to an embodiment of the present invention.
[0028] Figures 7a to 7e This is a comparison chart of signal enhancement effects under a noise level of SNR=-6dB according to an embodiment of the present invention. Detailed Implementation
[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0030] Example 1 See Figure 1 The specific steps of a TISRNet-based method for diagnosing rotating machinery faults are as follows: S1: Collect vibration signals of rotating machinery under different working conditions; S2: Construct an end-to-end TISRNet (Teager energy operator-informed shrinkage residual network) model, which consists of a data preprocessing module, a TFE (TEO-inspired physics-aware feature-oriented signal enhancement) module, and a PRSN (Residual shrinkage network with pre-shrinkage strategy) network. The model sets a physical information loss function to jointly constrain signal enhancement and classification tasks.
[0031] The data preprocessing module is used to divide the vibration signal collected in step S1 into datasets and perform sliding window sampling; the original data is divided into training set and test set; the divided data are sampled by sliding window, and the length of the sliding window is determined according to the bearing fault characteristic frequency to ensure that the sample contains complete fault cycle information.
[0032] like Figure 2 As shown, the TFE module takes the original vibration signal as input and generates an enhanced signal through the following steps: First, the instantaneous energy features of the signal are extracted by the TEO (Teager energy operator) to highlight the impulse component of the signal, and the amplification effect of high-frequency noise is suppressed by Batch Normalization (BN) to stabilize the feature distribution.
[0033] Secondly, the TEO output signal is input into a multi-scale convolution module, which consists of multiple sets of parallel one-dimensional convolutions with different kernel sizes, used to extract impact and modulation features at different time scales. Short convolution kernels are used to capture instantaneous impacts, while long convolution kernels are used to extract periodic envelope features. The calculation formula is as follows: (1) in, This indicates the TEO output signal. This represents the size of the i-th convolutional kernel. Represents one-dimensional convolution. M This indicates the number of convolution kernels.
[0034] The feature maps output from each convolutional branch are concatenated along the channel dimension to form a multi-scale feature tensor, as shown in the following equation: (2) in, This indicates a cascading operation at the channel dimension.
[0035] Furthermore, a channel attention mechanism is introduced into the multi-scale feature tensor for channel recalibration. Statistical information of each channel is obtained through global average pooling (GAP), and the weight coefficients of each channel are calculated through the channel attention mechanism (ECA) to highlight the scale features that contribute more to fault detection.
[0036] Then, the multi-scale features that have undergone channel recalibration are input into a one-dimensional fusion convolutional layer. Through convolution operations, the features of different scales are adaptively linearly combined in the time dimension, thereby realizing the dynamic fusion of multi-scale information.
[0037] Next, the fusion features z Apply a soft thresholding function with a learnable threshold to generate the final boosting weights. : (3) in, These are learnable parameters that are shared across all channels and time steps. This operation adaptively suppresses low-energy components while preserving dominant impulse characteristics, thereby achieving signal sparsity and robust enhancement.
[0038] Finally, the weight will be increased. With the original signal The signal is enhanced by applying the signal point by point and using residual superposition. (4) in, The original signal, To enhance weight, This is a learnable residual scaling factor. The enhanced signal is input into a subsequent diagnostic network for deep feature extraction and classification.
[0039] The PRSN network is used for deep feature extraction and fault identification of signals enhanced by the TFE module. It employs a residual network structure with a pre-shrinkage mechanism to improve the stability and noise resistance of feature extraction. The network consists of one-dimensional convolutional layers, multiple RPSBU (Residual Pre-shrinkage Building Unit) units, and a fully connected classification layer.
[0040] The RPSBU unit is the core improved module in this embodiment, and its structure is as follows: Figure 3 As shown, unlike the "convolution followed by contraction" structure in traditional DRSN, this embodiment places the soft-threshold feature contraction operation before the convolution calculation. Specifically, the input features are first adaptively soft-thresholded using contraction blocks. This soft-threshold function is learned by a two-layer fully connected network and shared across all channels, and can be dynamically adjusted according to the energy distribution of the input signal.
[0041] The physical information loss function mainly consists of two parts: the main task loss function and the auxiliary task loss function. The main task loss function uses the cross-entropy loss function to supervise the classification task and optimize the fault identification performance. The auxiliary task loss function introduces an auxiliary loss based on spectral entropy to constrain the energy concentration of the enhanced signal in the frequency domain, so that the signal enhancement process maintains physical consistency with the diagnostic target.
[0042] Through the joint optimization of the primary and auxiliary tasks, the signal enhancement module, during training, not only updates its parameters based on classification error but is also guided by spectral entropy constraints, thereby achieving a synergistic balance between "diagnostic effectiveness" and "physical consistency." This synergistic optimization mechanism ensures that the enhanced signal output by the signal enhancement module strengthens the feature components that contribute to fault category discrimination while maintaining the integrity of the signal's physical characteristics, avoiding distorted enhancement results that sacrifice physical authenticity for the sake of classification accuracy. Thus, adaptive coupling and synergistic improvement of the signal enhancement and fault diagnosis processes under a unified optimization objective are achieved.
[0043] S3: Input the training set into the constructed TISRNet model for training.
[0044] During training, a primary and secondary task structure optimization strategy is adopted. Specifically, cross-entropy loss is used as the primary task loss function, with its scope covering the entire model during backpropagation, while spectral entropy loss is used as the secondary task loss function, with backpropagation only applied to the TFE module. In practice, gradient flow control at the structural level is achieved through steps such as freezing the main network parameters, backpropagating the gradients of the enhancement modules separately, and restoring the network parameter states. This effectively ensures the independence of the enhancement module optimization from classification performance.
[0045] S4: Input the test set samples into the trained TISRNet model, and the model outputs the fault probability distribution for each category. Finally, the fault category is determined according to the maximum probability principle, and the enhanced signal is visualized and analyzed to verify the physical effectiveness of the signal enhancement module and the noise resistance performance of the model.
[0046] This embodiment constructs an energy operator-guided signal enhancement module and trains it in an end-to-end manner with a fault diagnosis network. This maintains consistency between the enhanced features and the classification target, enabling the signal enhancement results to adaptively meet the requirements of the diagnostic task. This achieves simultaneous optimization of noise suppression, feature extraction, and classification performance.
[0047] Example 2 The steps in this embodiment are the same as in Embodiment 1, the difference being that each step is applied to a specific instance, combined with... Figure 4 The technical solution of this embodiment will be described in detail. The fault classification result can be obtained by inputting the collected vibration signal into the TISRNet model. The specific steps are as follows: (I) Collecting experimental data 1. This embodiment uses a self-constructed rolling bearing fault diagnosis dataset. The experimental platform is the BTS100 rolling bearing test bench, which consists of an inverter, a motor, a motor shaft, and a bearing housing.
[0048] 2. The experiment used SKF6006 deep groove ball bearings. Wire EDM was used to machine defects onto the bearing surface, and three types of failures were set: outer ring failure, inner ring failure, and rolling element failure. Each type of failure was further divided into several subcategories based on the degree of damage, as detailed in Table 1.
[0049] 3. Further incorporate scenarios where multiple rolling elements are damaged simultaneously in rolling element bearings to increase the complexity of fault types and reflect the true vibration characteristics under non-stationary operating conditions.
[0050] Table 1 Fault Classification of Self-built Dataset
[0051] (II) Data Preprocessing 1. To simulate noise interference under complex operating conditions, Gaussian white noise of different intensities is superimposed on the original signal. The signal-to-noise ratio (SNR) is calculated as follows: (5) in and These represent the effective power of the signal and noise, respectively.
[0052] In this embodiment, SNR values of -6dB, -4dB, -2dB, 0dB, 2dB, 4dB, and 6dB are selected, representing the noise levels.
[0053] 2. After noise superposition is completed, the noisy signal is sampled using a sliding window method, where the window length is determined by the fault characteristic frequency as 4663. The first 80% of the data is used to generate the training set, and the last 20% of the data is used to generate the test set.
[0054] (III) Model Training and Optimization Strategies In this embodiment, a primary-secondary task collaborative optimization mechanism is adopted during the model training phase: 1. The primary task uses cross-entropy loss to supervise fault classification performance, while the secondary task uses spectral entropy loss to constrain the frequency domain concentration of the signal enhancement module's output. The two losses are optimized independently through gradient isolation, ensuring that the secondary task enhances the signal without interfering with the primary task's classification accuracy.
[0055] 2. The AdamW optimizer is used for the main task and the Adam optimizer is used for the auxiliary task. The learning rate of both is dynamically updated through a cosine annealing learning rate adjustment strategy to accelerate convergence and improve stability.
[0056] (IV) Experimental Design and Results Analysis To verify the effectiveness of the method of this invention, DRSN, WDCNN, ClassBD, and a diagnostic model based on MSCNN-LSTM structure were selected as comparison objects, and accuracy, F1 score, and FPR score were used as evaluation indicators. The specific calculation formulas for F1 and FPR are as follows: (6) (7) In this context, TP, TN, FP, and FN represent the number of true positives, true negatives, false positives, and false negatives, respectively.
[0057] Figure 5The accuracy of each model under different noise conditions is compared. Table 2 shows the F1 score of each model under different noise conditions, and Table 3 shows the FPR score of each model under different noise conditions. The experimental results show that the method of the present invention can maintain high diagnostic accuracy under multiple noise levels; under strong noise conditions with a signal-to-noise ratio (SNR) of -6dB, the classification accuracy still exceeds 90%.
[0058] Table 2. F1 scores of each model on the WUST dataset
[0059] Table 3. FPR scores of each model on the WUST dataset
[0060] Figures 6a to 6e The original signal and signal enhancement results for different fault categories are shown under noise conditions with SNR=-4dB. Figures 7a to 7e The original signal and signal enhancement results for different fault categories are shown under noise conditions of SNR=-6dB; these results demonstrate that the signal enhancement module constructed in this embodiment can effectively recover the impact characteristics masked by noise.
[0061] This embodiment constructs a signal enhancement module capable of strengthening impact characteristics, and achieves joint optimization of signal enhancement and fault identification through a collaborative optimization network integrating enhancement and diagnosis. This embodiment can improve the accuracy and reliability of fault diagnosis, while enhancing signal interpretability, and is suitable for monitoring rotating machinery in complex noise environments. This embodiment can be further extended to more types of mechanical equipment, more complex operating conditions, and real-time online diagnostic scenarios.
[0062] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0063] Example 3 This embodiment is used to implement the principle of the above method embodiment to construct a rotating machinery fault diagnosis system based on TISRNet, including a data acquisition module, an end-to-end TISRNet model, a training module and a testing module; The data acquisition module is used to collect vibration signals of rotating machinery under different operating conditions; The end-to-end TISRNet model comprises a data preprocessing module, a TFE module, and a PRSN network connected sequentially. The model uses a physical information loss function to jointly constrain signal enhancement and classification tasks. The data preprocessing module divides the original signal into training and testing sets, then generates samples through sliding window sampling. The TFE module includes a TEO energy operator, a multi-scale convolution module, a channel-weighted fusion module, a one-dimensional fusion convolutional layer, a soft thresholding function module, and multiplication and stacking modules to generate enhanced signals from the preprocessed signals. The PRSN network includes one-dimensional convolutional layers, multiple RPSBU units, and a fully connected classification layer, employing a residual structure with a pre-shrinking mechanism for deep feature extraction and fault identification of the enhanced signals. The RPSBU units use shrinking blocks to adaptively soft-threshold the features of the enhanced signals. The soft thresholding function comprises a two-layer fully connected network shared across all channels and dynamically adjusted according to the energy distribution of the input signal. The physical information loss function includes a main task loss function and an auxiliary task loss function. The main task loss function uses the cross-entropy loss function to supervise the classification task and optimize the fault identification performance. The auxiliary task loss function introduces an auxiliary loss based on spectral entropy to constrain the energy concentration of the enhanced signal in the frequency domain, so that the signal enhancement process maintains physical consistency with the diagnostic target.
[0064] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.
[0065] This embodiment also includes a processor, a communication interface, a memory, and a communication bus; wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the program is executed by the processor, the processor performs the steps of a TISRNet-based rotating machinery fault diagnosis method.
[0066] This embodiment also provides a computer-readable storage medium storing executable instructions that, when executed by a processor, enable the processor to implement a TISRNet-based method for diagnosing rotating machinery faults.
[0067] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.
[0068] Furthermore, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0069] This application is described with reference to the flowchart of the method and computer program product according to Embodiment 1 and the block diagram of the device (system) according to Embodiment 3. It should be understood that each step or block in the flowchart or block diagram, as well as combinations of steps or blocks in the flowchart or block diagram, can be implemented by computer program instructions.
[0070] These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which are executable by the processor of the computer or other programmable data processing device, produce instructions for implementing the process. Figure 1 One or more processes or boxes Figure 1 A TISRNet-based rotating machinery fault diagnosis system that specifies functions in one or more boxes.
[0071] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes or boxes Figure 1 The function specified in one or more boxes.
[0072] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes or boxes Figure 1 The steps of a TISRNet-based rotating machinery fault diagnosis method are specified in one or more boxes.
[0073] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.
Claims
1. A fault diagnosis method for rotating machinery based on TISRNet, characterized in that: Includes the following steps: S1: Collect vibration signals of rotating machinery under different operating conditions; S2: Construct a TISRNet model including a data preprocessing module, a trainable signal enhancement module, and a fault diagnosis backbone network; input vibration signals and perform fault diagnosis; specifically including: S21: The vibration signal is divided into datasets and subjected to sliding sampling through the data preprocessing module; S22: Adaptive enhancement of the sampled signal guided by physical information is performed through a trainable signal enhancement module, and an enhanced signal is output. S23: Deep feature extraction and fault identification of enhanced signals are performed through the fault diagnosis backbone network; S3: The TISRNet model is trained using a joint optimization strategy of primary and secondary tasks; S4: Test the trained TISRNet model, determine the fault category based on the maximum probability principle, and output the probability distribution of each type of fault.
2. The rotating machinery fault diagnosis method based on TISRNet according to claim 1, characterized in that: The specific steps in step S21 are as follows: The collected vibration signals were divided into training and testing sets; The sliding window length is determined based on the bearing fault characteristic frequency, and the divided data is sampled using the sliding window.
3. The method for diagnosing rotating machinery faults based on TISRNet according to claim 1, characterized in that: The specific steps in step S22 are as follows: S221: Extract the instantaneous energy features of the sampled signal through the TEO energy operator, and stabilize the feature distribution through batch normalization (BN). S222: The instantaneous energy features are extracted using multiple sets of parallel one-dimensional convolutions with different kernel sizes to extract the impact and modulation features at different time scales; the feature maps output by each convolution branch are spliced together in the channel dimension to form a multi-scale feature tensor; S223: Introducing a channel attention mechanism to recalibrate the multi-scale feature tensor, obtaining statistical information of each channel through global average pooling (GAP), and calculating the weight coefficient of each channel through the channel attention mechanism (ECA) to highlight the scale features that contribute more to fault detection. S224: Input the multi-scale features after channel recalibration into a one-dimensional fusion convolutional layer, and perform adaptive linear combination of features of different scales in the time dimension through convolution operation, thereby realizing the dynamic fusion of multi-scale information; S225: Apply a learnable threshold function to the fused features to generate enhanced weights; S226: Multiply the enhanced weights point by point with the preprocessed signal and sum the residuals to output the enhanced signal.
4. The rotating machinery fault diagnosis method based on TISRNet according to claim 3, characterized in that: In step S222, the multiple sets of parallel one-dimensional convolutions with different kernel sizes include short convolution kernels and long convolution kernels; Short convolutional kernels are used to capture instantaneous impacts, while long convolutional kernels are used to extract periodic envelope features; use This indicates the TEO output signal. This represents the size of the i-th convolutional kernel. M Indicates the number of convolution kernels. This represents a one-dimensional convolution, where each convolutional branch outputs a feature. f i for: (1)； This represents a cascaded operation along the channel dimension, a multi-scale feature tensor. F norm for: (2)。 5. The rotating machinery fault diagnosis method based on TISRNet according to claim 3, characterized in that: In step S225, using This represents the learnable parameters shared across all channels and time steps, contributing to the fused features. z Apply a soft thresholding function with a learnable threshold to generate boosting weights. for: (3)。 6. The method for diagnosing rotating machinery faults based on TISRNet according to claim 1, characterized in that: In step S226, using This indicates the preprocessed signal. To enhance weight, A learnable residual scaling factor to enhance the signal. for: (4)。 7. The rotating machinery fault diagnosis method based on TISRNet according to claim 1, characterized in that: In step S3, the cross-entropy loss is used as the main task loss function and applied to the entire model for backpropagation; the spectral entropy loss is used as the auxiliary task loss function and is only applied to the TFE module for backpropagation.
8. A rotating machinery fault diagnosis system for use in the rotating machinery fault diagnosis method based on TISRNet as described in any one of claims 1 to 7, characterized in that: It includes a data acquisition module, an end-to-end TISRNet model, a training module, and a testing module; The data acquisition module is used to collect vibration signals of rotating machinery under different operating conditions; The end-to-end TISRNet model comprises a data preprocessing module, a trainable signal enhancement module, and a fault diagnosis backbone network connected in sequence. The model uses a physical information loss function to jointly constrain signal enhancement and classification tasks. The data preprocessing module samples vibration signals and divides them into training and testing sets. The trainable signal enhancement module includes a TEO energy operator, a multi-scale convolution module, a channel recalibration module, a one-dimensional fusion convolutional layer, a soft thresholding function module, and a residual-based enhanced signal generation module, used for adaptive enhancement of the sampled signals guided by physical information, outputting the enhanced signal. The fault diagnosis backbone network includes a one-dimensional convolutional layer, multiple RPSBU units, and a fully connected classification layer, which are used for deep feature extraction and fault category identification of the enhanced signal; the RPSBU units are used to perform adaptive soft thresholding on the features of the enhanced signal using shrink blocks. The soft thresholding function consists of a two-layer fully connected network that is shared across all channels and dynamically adjusted according to the energy distribution of the input signal; The physical information loss function includes a main task loss function and an auxiliary task loss function. The main task loss function uses the cross-entropy loss function to supervise the classification task and optimize the fault identification performance. The auxiliary task loss function introduces an auxiliary loss based on spectral entropy to constrain the energy concentration of the enhanced signal in the frequency domain, so that the signal enhancement process maintains physical consistency with the diagnostic target.
9. A rotating machinery fault diagnosis system according to claim 8, characterized in that: The TEO energy operator is used to extract the instantaneous energy characteristics of a signal; The multi-scale convolution module includes multiple sets of parallel one-dimensional convolutions with different kernel sizes, used to extract impact and modulation features at different time scales; The channel recalibration module is used to introduce multi-scale feature tensors into the channel attention mechanism for channel recalibration. It obtains statistical information of each channel through global average pooling (GAP) and calculates the sparse weights of each channel through the channel attention mechanism (ECA) to highlight scale features that contribute more to fault detection. One-dimensional fusion convolutional layers are used to enable multi-scale features after channel weighting to automatically learn the weighting coefficients of different scale features at each time step, thereby achieving dynamic fusion of multi-scale information to obtain fused features. The soft thresholding function module is used to apply a learnable threshold to the fused features using a soft thresholding function, generating enhanced weights; The residual-based enhancement signal generation module is used to generate the final enhanced signal output by modulating the preprocessed signal of the enhancement weight domain point by point and combining residual links.
10. A computer memory, characterized in that: It contains a computer program that can be executed by a computer processor, which performs a TISRNet-based rotating machinery fault diagnosis method as described in any one of claims 1 to 7.