Rolling bearing fault diagnosis method, computer device and computer program product

Abstract

The application is suitable for the technical field of rolling bearing fault diagnosis, and provides a rolling bearing fault diagnosis method, computer equipment and computer program product. The method comprises the following steps: obtaining a vibration signal of a rolling bearing to be diagnosed and constructing a corresponding time-frequency feature; training a double-path parallel hybrid model based on an impact feature driven gate control mechanism based on a time-frequency feature sample set. The model comprises units such as time-frequency local information extraction, impact feature extraction, long short-term memory network modeling, Transformer branch modeling and impact driven gate fusion. The time-frequency feature is input into the model to obtain a diagnosis result comprising a fault category label, a prediction probability, a timestamp and a running state identifier. The application can effectively solve the problems of insufficient feature extraction of rolling bearing vibration signals under complex working conditions and difficult collaborative modeling of time series and global features, and improve the precision and stability of fault diagnosis.

Description

Technical Field

[0001] This application belongs to the field of rolling bearing fault diagnosis technology, and in particular relates to a rolling bearing fault diagnosis method, computer equipment and computer program product. Background Technology

[0002] Rolling bearings, as key basic components in rotating machinery, are widely used in nuclear power plant main coolant pumps, circulating fans, steam turbine units, and various industrial machinery. Their operating status directly affects the safety, reliability, and service life of the equipment. In the nuclear industry and other engineering fields with high reliability requirements, rolling bearings typically operate under high speed, high load, and complex environmental conditions for extended periods. Existing technologies have inherent limitations in the feature modeling stage. Traditional methods rely on manually designed time-domain indicators (such as peak factor and kurtosis) or frequency-domain indicators (such as spectral energy distribution). These feature extraction processes are highly dependent on expert experience, their expressive power is limited by preset rules, and they are difficult to adapt to complex scenarios such as varying operating conditions and loads, resulting in weak feature generalization ability and susceptibility to noise interference. Summary of the Invention

[0003] In view of this, embodiments of this application provide a rolling bearing fault diagnosis method, computer equipment, and computer program product to solve the technical problems of insufficient extraction of rolling bearing vibration signal features under complex working conditions, difficulty in co-modeling of temporal and global features, and insufficient stability of multi-category fault diagnosis.

[0004] The first aspect of this application provides a method for diagnosing rolling bearing faults, including: The vibration signal of the rolling bearing to be diagnosed is acquired and the corresponding time-frequency characteristics are constructed based on the vibration signal; Training a dual-path parallel hybrid model based on the time-frequency feature sample set of rolling bearings and the impact feature-driven gate control mechanism. The dual-path parallel hybrid model includes a time-frequency local information extraction unit, a feature sequence combination unit, an impact feature extraction unit, a long short-term memory network modeling unit, a Transformer branch modeling unit, an impact-driven gate control fusion unit, and a classification mapping unit. The time-frequency features are input into the dual-path parallel hybrid model to obtain the diagnostic results, which include fault category labels, predicted probabilities, timestamps, and corresponding operating status identifiers.

[0005] In one implementation of the first aspect, the training of a dual-path parallel hybrid model based on an impact feature-driven gate control mechanism using a time-frequency feature sample set based on rolling bearings includes: The vibration signal of the rolling bearing is collected to obtain the original vibration time series data, and a corresponding time-frequency feature sample set is constructed based on the original vibration time series data. The time-frequency feature sample set includes a training set, and the training set includes time-frequency feature samples. A dual-path parallel hybrid model based on the impact feature-driven gate control mechanism is constructed, which includes a time-frequency local information extraction unit, a feature sequence combination unit, an impact feature extraction unit, a long short-term memory network modeling unit, a Transformer branch modeling unit, an impact-driven gating fusion unit, and a classification mapping unit. The time-frequency feature samples in the training set are input into the dual-path parallel hybrid model to obtain the predicted classification label; Based on the predicted classification labels, the dual-path parallel hybrid model is jointly optimized with gating constraints. Return to the step of inputting the time-frequency feature samples in the training set into the dual-path parallel hybrid model to obtain the predicted classification label, until the dual-path parallel hybrid model converges after joint optimization.

[0006] In one implementation of the first aspect, the step of inputting time-frequency feature samples from the training set into the dual-path parallel hybrid model to obtain predicted classification labels includes: The time-frequency feature samples are input into the convolutional local enhancement unit in the time-frequency local information extraction unit to enhance the local frequency band energy changes and transient impact edges to obtain the locally enhanced feature matrix. The feature sequence data is obtained by performing dimensional rearrangement and time series reconstruction on the locally enhanced feature matrix through the feature sequence combination unit. The original vibration time series data is input into the impact feature extraction unit to extract impact features and obtain an impact feature vector. The feature sequence data is input into the long short-term memory network of the long short-term memory network modeling unit to obtain a temporal feature representation; The feature sequence data is input into the Transformer branch modeling unit to obtain a global feature representation; The impact feature vector is input into the gating generation module to obtain the gating factor, and the time-series feature representation and the global feature representation are fused based on the gating factor to obtain the fused feature. The gating factor reflects the impact intensity level of the time-frequency feature sample. The fused features are input into the fully connected layer and nonlinear activation function of the classification mapping unit to obtain a classification vector, and the running state corresponding to the dimension with the highest probability in the classification vector is determined as the predicted classification label.

[0007] In one implementation of the first aspect, the original vibration time series data is input into an impact feature extraction unit to extract impact features and obtain an impact feature vector, including: Calculate the kurtosis, peak factor, high-frequency energy ratio, and envelope spectral energy of the original vibration time series data; An impact feature vector is constructed based on the kurtosis, peak factor, high-frequency energy ratio, and envelope spectrum energy. The impact feature vector is normalized to obtain the final impact feature vector.

[0008] In one implementation of the first aspect, the gating constraint joint optimization of the dual-path parallel hybrid model based on the predicted classification label includes: The gated joint optimization loss is calculated based on the predicted classification labels and the following formula. : ; in, t To predict the time window to which the segmented samples corresponding to the classification label belong, This is the classification loss, used to measure the difference between the predicted category and the true label; in, For gated response consistency loss, The constraint weighting coefficients for the gating response consistency loss are: , The impact strength normalization mapping function, The gate factor; in, To smooth out the stability loss of the gate, The constraint weight coefficients are those for the gated smoothing stability loss. , The gating factor for the current time window. The gating factor for the previous time window; The gradient is calculated based on the joint optimization loss under the gating constraints, and the parameters of the long short-term memory network modeling unit, the Transformer branch modeling unit, and the gating generation module in the dual-path parallel hybrid model are optimized.

[0009] In one implementation of the first aspect, the gating constraint joint optimization of the dual-path parallel hybrid model based on the predicted classification label further includes: In each training iteration, the distribution of gating factors of the time-frequency feature samples of the current training batch is statistically analyzed and statistical analysis results are generated. If the statistical analysis results indicate that the probability that the weight of any feature in the temporal feature representation and the global feature representation is lower than the preset minimum weight is greater than the preset maximum probability, or the correlation coefficient between the gating factor and the impact intensity corresponding to the impact feature vector is less than the preset correlation threshold, then at least one of the following operations is performed: adjusting the constraint weight coefficient, optimizing the learning rate of the gating generation module, and enhancing the gradient update ratio of the feature whose weight is lower than the preset minimum weight is greater than the preset maximum probability.

[0010] In one implementation of the first aspect, the step of acquiring the vibration signal of the rolling bearing to obtain raw vibration time series data and constructing a corresponding time-frequency feature sample set based on the raw vibration time series data includes: The vibration signal of the rolling bearing is collected to obtain the original vibration time series data, which includes the sampling timestamp, vibration amplitude and corresponding operating status identifier; By unifying the sequence length of the original vibration time series data, time series data with consistent length are obtained. Based on a preset window length and a preset sliding step size, the time series data is segmented and sampled using a sliding window, and then normalized to obtain segmented samples. Time-frequency analysis is performed on the segmented samples to obtain time-frequency feature samples, and the time-frequency feature samples have the corresponding running status identifiers of the original vibration time series data; A time-frequency feature sample set is constructed based on the aforementioned time-frequency feature samples. The time-frequency feature sample set includes a training set, a validation set, and a test set.

[0011] In one implementation of the first aspect, before unifying the sequence length of the original vibration time series data to obtain time series data of consistent length, the method further includes: A random noise signal is generated based on the target signal-to-noise ratio and superimposed on the original vibration time series.

[0012] A second aspect of this application provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect.

[0013] A third aspect of this application provides a computer program product including a computer program that, when run, causes the method described in the first aspect to be executed.

[0014] The first aspect of the embodiments of this application has the following beneficial effect: by constructing a dual-path gating modeling structure driven by impact features and a joint optimization mechanism, a complete technical closed loop is realized from signal preprocessing, feature enhancement, dual-path parallel modeling to adaptive fusion diagnosis. This scheme not only makes full use of the multi-scale structural information of vibration signals, but also ensures the response consistency and smooth stability of the gating mechanism by introducing physical prior constraints. It effectively overcomes the problem of incomplete representation caused by the fragmented feature modeling in traditional methods, and has the advantages of being able to collaboratively process the multi-scale characteristics of vibration signals and significantly improving the accuracy and robustness of fault diagnosis.

[0015] It is understood that the beneficial effects of the second and third aspects mentioned above can be found in the relevant descriptions in the first aspect above, and will not be repeated here. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram illustrating the implementation process of the rolling bearing fault diagnosis method provided in the embodiments of this application; Figure 2 This is a schematic diagram illustrating the implementation process of the rolling bearing fault diagnosis method provided in the embodiments of this application; Figure 3 This is a schematic diagram illustrating the implementation process of the rolling bearing fault diagnosis method provided in the embodiments of this application; Figure 4 This is a schematic diagram illustrating the implementation process of the rolling bearing fault diagnosis method provided in the embodiments of this application; Figure 5 This is a schematic diagram of the dual-path parallel hybrid model architecture provided in the embodiments of this application; Figure 6 This is a schematic diagram of the computer device provided in an embodiment of this application; Figure 7 This is a schematic diagram of a computer program product provided in an embodiment of this application. Detailed Implementation

[0018] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0019] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0020] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0021] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0022] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0024] Rolling bearings, as key basic components in rotating machinery, are widely used in main coolant pumps, circulating fans, steam turbine units, and various industrial machinery in nuclear power plants. In the nuclear industry and other engineering fields with high reliability requirements, rolling bearings typically operate under high speed, high load, and complex environmental conditions for extended periods. Existing rolling bearing fault diagnosis technologies mainly rely on manually designed features combined with shallow machine learning models, or use a single deep learning model for diagnosis under complex operating conditions and strong noise environments. However, bearing vibration signals often exhibit significant non-stationarity and nonlinear characteristics. Manual feature extraction relies on experience and has limited generalization ability; while single long short-term memory networks have low computational efficiency when processing long sequences with high sampling rates and limited ability to model global features, and single Transformer models lack sensitivity to local temporal evolution features. Therefore, under complex operating conditions and noise interference, how to fully exploit the temporal features of rolling bearing vibration signals while effectively modeling global features and improving the accuracy and stability of fault diagnosis is a key problem that existing technologies urgently need to solve.

[0025] Based on the above issues, such as Figure 1 As shown in the figure, a rolling bearing fault diagnosis method provided in this application embodiment includes the following steps: Step S1: Obtain the vibration signal of the rolling bearing to be diagnosed and construct the corresponding time-frequency characteristics based on the vibration signal.

[0026] In applications, acquiring the vibration signal of the rolling bearing to be diagnosed can refer to collecting the mechanical vibration generated by the rolling bearing during actual operation using vibration sensors installed in the bearing housing or adjacent structures. This vibration is then converted into an electrical signal, amplified and filtered by a signal conditioning module, and finally converted into raw vibration time-series data containing sampling timestamps, vibration amplitude, and corresponding operating status identifiers by an analog-to-digital converter. Constructing corresponding time-frequency features based on the vibration signal can refer to using time-frequency analysis methods to convert the one-dimensional non-stationary vibration signal into a two-dimensional feature representation containing both time and frequency information. Specifically, a continuous wavelet transform (CWT) is used to decompose the raw vibration time-series data into multiple scales. By selecting a mother wavelet function with good time-frequency localization characteristics, projection analysis is performed on the vibration signal at different scales and time positions, calculating wavelet coefficients at different scales to form a two-dimensional time-frequency feature matrix. This step aims to eliminate the influence of signal non-stationarity, providing information-rich and structured input data for subsequent deep learning models, thereby significantly improving the separability between different fault types under complex and noisy operating conditions.

[0027] Step S2: Train a dual-path parallel hybrid model based on the impact feature-driven gate control mechanism using the time-frequency feature sample set of rolling bearings. The dual-path parallel hybrid model includes a time-frequency local information extraction unit, a feature sequence combination unit, an impact feature extraction unit, a long short-term memory network modeling unit, a Transformer branch modeling unit, an impact-driven gate control fusion unit, and a classification mapping unit.

[0028] In applications, training based on the time-frequency feature sample set of rolling bearings can refer to using pre-divided time-frequency feature data containing training, validation, and test sets to iteratively optimize the parameters of a constructed neural network model until the model converges. The dual-path parallel hybrid model based on the impact feature-driven gate control mechanism is a composite architecture that integrates physical statistical features and deep learning.

[0029] In application, the time-frequency local information extraction unit is used to enhance the local frequency band energy changes and transient impact edges in the time-frequency features to obtain the locally enhanced feature matrix; The feature sequence combination unit is used to perform dimensional rearrangement and time series reconstruction on the locally enhanced feature matrix to obtain feature sequence data; The impact feature extraction unit is used to extract impact features from the original vibration time series data to obtain an impact feature vector. The long short-term memory network modeling unit is used to generate the temporal feature representation of the feature sequence data; The Transformer branch modeling unit is used to generate a global feature representation of the feature sequence data; The impact-driven gating fusion unit is used to generate a gating factor based on the impact feature vector and fuse the temporal feature representation and the global feature representation based on the gating factor to obtain a fused feature. The gating factor reflects the impact intensity level of the time-frequency feature sample. The classification mapping unit is used to generate a classification vector based on the fusion features and determine the running state corresponding to the dimension with the highest probability in the classification vector as the predicted classification label.

[0030] Step S3: Input the time-frequency features into the dual-path parallel hybrid model to obtain the diagnostic results. The diagnostic results include fault category labels, predicted probabilities, timestamps, and corresponding operating status identifiers.

[0031] In application, fault category labels and predicted classification labels have the same meaning. The difference is that predicted classification labels are generated during the training phase, while fault category labels are generated during the actual diagnosis phase.

[0032] In application, inputting time-frequency features into the dual-path parallel hybrid model can refer to feeding the two-dimensional time-frequency feature matrix obtained after processing the rolling bearing vibration signal to be diagnosed into a trained and parameter-fixed dual-path parallel hybrid model to perform forward propagation calculations. The diagnostic result can refer to the structured judgment information output by the model after a series of processes, including local information extraction, sequence reconstruction, dual-path feature modeling, impact-driven gating fusion, and classification mapping. The fault category label is the bearing operating state determined by the model according to the maximum probability principle, such as discrete categories like normal, inner race fault, outer race fault, or rolling element fault; the predicted probability is the confidence value of the corresponding fault category in the model's output classification vector, used to quantify the reliability of the diagnostic result; the timestamp is the sampling time information corresponding to the input vibration signal segment, used to achieve precise location and traceability of the fault occurrence time; and the corresponding operating status identifier is a description of the equipment's operating condition associated with the fault category label, such as high-speed heavy load or low-speed no-load.

[0033] This application constructs a complete diagnostic chain, from signal time-frequency conversion to dual-path parallel modeling and then to impact-driven adaptive fusion, through the synergistic effect of the aforementioned technical features. By utilizing the dual-path parallel structure of LSTM and Transformer to capture the temporal evolution and global dependencies of the signal, it overcomes the limitations of a single model. More importantly, it introduces a gated fusion mechanism driven by impact statistical features, enabling the model to dynamically adjust the contribution weights of temporal and global features based on the physical impact intensity of the vibration signal, achieving deep coupling between physical prior knowledge and data-driven learning. This ensures that the model prioritizes enhancing temporal modeling capabilities under strong impact fault scenarios and strengthens global modeling capabilities under weak impact or trend-based fault scenarios, achieving high-precision and robust rolling bearing fault diagnosis under complex working conditions.

[0034] In one embodiment, such as Figure 2 As shown, step S2 includes: Step S21: Collect the vibration signal of the rolling bearing to obtain the original vibration time series data, and construct the corresponding time-frequency feature sample set based on the original vibration time series data. The time-frequency feature sample set includes a training set, which includes time-frequency feature samples.

[0035] In applications, acquiring vibration signals from rolling bearings can refer to converting mechanical vibrations into electrical signals using vibration sensors installed in the bearing housing or adjacent structures. These signals are then amplified and filtered by a signal conditioning module, and finally converted into digital signals by an analog-to-digital converter. The data is then stored chronologically as one-dimensional time-series data containing sampling timestamps, vibration amplitudes, and corresponding operating status identifiers – this is the original vibration time-series data. Step S22: Construct a dual-path parallel hybrid model based on the impact feature-driven gate control mechanism, including a time-frequency local information extraction unit, a feature sequence combination unit, an impact feature extraction unit, a long short-term memory network modeling unit, a Transformer branch modeling unit, an impact-driven gating fusion unit, and a classification mapping unit.

[0036] In applications, the time-frequency local information extraction unit is typically composed of multi-layer convolutional neural networks, used to scan the input time-frequency feature matrix to extract local frequency band energy changes and transient impact edge information; the feature sequence combination unit is used to rearrange the dimensions and reconstruct the time series of the locally enhanced feature matrix, generating feature sequence data that maintains the time order; the impact feature extraction unit is independent of the dual-path system, used to directly calculate statistics such as kurtosis, peak factor, high-frequency energy ratio, and envelope spectrum energy from the original vibration time series data, constructing an impact feature vector reflecting the impact intensity level; the long short-term memory network modeling unit is composed of multi-layer LSTM cells stacked together, used to capture the temporal dynamic evolution law in the feature sequence; the Transformer branch modeling unit includes position encoding, multi-head self-attention mechanism, and feedforward network, used to model the global dependencies in the feature sequence; the impact-driven gating fusion unit receives the impact feature vector and generates a gating factor, and performs weighted fusion of the temporal feature representation output by the LSTM branch and the global feature representation output by the Transformer branch based on this factor; the classification mapping unit consists of fully connected layers and nonlinear activation functions, used to map the fused features to the probability distribution of fault categories. For example, the impulse feature extraction unit pre-defines a calculation formula, normalizes the kurtosis value and peak factor of the original signal, and concatenates them into a vector C. This vector C is then input into the mapping function f(C) in the gating generation module to generate the initial gating factor g. By constructing this dual-path structure that includes a physical statistical feature-driven mechanism, the foundation for the collaborative work of temporal features and global features can be established at the model architecture level, and the impulse statistics can be used as prior knowledge to guide the learning direction of the gating mechanism.

[0037] Step S23: Input the time-frequency feature samples in the training set into the dual-path parallel hybrid model to obtain the predicted classification label.

[0038] In application, inputting time-frequency feature samples into the model can refer to performing a forward propagation process. Specifically, the time-frequency feature samples in the training set are first processed by the time-frequency local information extraction unit and the feature sequence combination unit to be transformed into feature sequence data in a standard format. This feature sequence data is simultaneously fed into the Long Short-Term Memory Network modeling unit and the Transformer branch modeling unit to obtain the temporal feature representation F1 and the global feature representation F2, respectively. At the same time, the original vibration time series data corresponding to the time-frequency feature sample is fed into the impact feature extraction unit to obtain the impact feature vector, and then the current gating factor g is calculated by the gating generation module. The impact-driven gating fusion unit uses the gating factor g to adaptively weight and fuse F1 and F2 to obtain the fused feature F. Finally, the classification mapping unit outputs the predicted probabilities of various fault states and determines the category with the highest probability as the predicted classification label. For example, in one iteration, the input sample, after processing, yields a gating factor g=0.7, indicating a high impact intensity. The fusion unit automatically assigns a higher weight (0.7) to the temporal feature F1 and a lower weight (0.3) to the global feature F2, resulting in a predicted classification label of inner circle fault. This step realizes a complete inference chain from raw data to fault identification results and provides a direct predictive basis for subsequent error calculation and parameter optimization.

[0039] Step S24: Perform gated constraint joint optimization on the dual-path parallel hybrid model based on the predicted classification labels.

[0040] In applications, gated joint optimization can refer to constructing a joint loss function and calculating gradients based on this joint loss to update model parameters. Specifically, during the optimization process, the weight and bias parameters in the Long Short-Term Memory network modeling unit, the Transformer branch modeling unit, and the gated generation module are synchronously updated using the backpropagation algorithm based on the value of the joint loss function.

[0041] Step S25: Return to step S23, which involves inputting time-frequency feature samples from the training set into the dual-path parallel hybrid model to obtain the predicted classification label, until the dual-path parallel hybrid model converges after joint optimization.

[0042] In application, convergence criteria typically include the classification accuracy on the validation set no longer significantly improving, the joint loss function value stabilizing, or fluctuations within a preset threshold across multiple training epochs. During iteration, the model continuously adjusts its internal parameters to minimize the joint loss, gradually enhancing its dual-path feature extraction capability and making the gating factor more sensitive and smooth in response to impact intensity. For example, setting a maximum training epoch of 100 epochs, when the validation set accuracy remains above 99.5% from epoch 85 to 90 and the loss function decreases by less than 1e-4, the model is considered converged, training is stopped, and the model parameters at this point are saved as the final fault diagnosis model. This closed-loop iterative mechanism ensures that the model can fully exploit the deep features in the vibration signal while satisfying physical constraints, ultimately achieving a rolling bearing fault diagnosis capability that combines high accuracy and high robustness.

[0043] This application constructs a complete training closed loop through the synergistic effect of the aforementioned technical features, encompassing data acquisition, model building, forward inference, constraint optimization, and iterative convergence.

[0044] In one embodiment, such as Figure 3 As shown, step S23 further includes: Step S231: Input the time-frequency feature samples into the convolutional local enhancement unit in the time-frequency local information extraction unit, and perform enhancement processing on the local frequency band energy changes and transient impact edges to obtain the locally enhanced feature matrix.

[0045] In applications, the convolutional local enhancement unit comprises multiple convolutional layers, batch normalization layers, and nonlinear activation functions. The size of the convolutional kernel can be set according to the resolution of the time-frequency plot, such as a 3×3 or 5×5 window, to capture energy abrupt changes within a local frequency band. Local frequency band energy changes can refer to amplitude fluctuations of the vibration signal over time within a specific frequency range, while transient impact edges can refer to the bright stripes or speckled boundaries that appear on the time-frequency plot when a fault occurs. Through convolution operations, this unit can suppress background noise and highlight these key features, thus obtaining a locally enhanced feature matrix. For example, when a spalling fault occurs in the inner ring of a rolling bearing, it appears as periodic vertical bright stripes on the time-frequency plot. The convolutional local enhancement unit enhances the contrast of these stripes through edge detection operators, making it easier for subsequent models to identify the fault mode. This local enhancement processing effectively improves the signal-to-noise ratio of weak impact signals, providing a more discriminative feature basis for subsequent sequence modeling.

[0046] Step S232: The feature sequence data is obtained by rearranging the dimensions and reconstructing the time series of the locally enhanced feature matrix through the feature sequence combination unit.

[0047] In applications, the feature sequence combination unit converts the two-dimensional locally enhanced feature matrix into a one-dimensional time series format to adapt to the input requirements of Long Short-Term Memory (LSTM) networks and Transformer models. The specific operations performed by this unit include dimensional rearrangement and time series reconstruction: first, the height dimension (frequency axis) of the feature matrix is ​​flattened or aggregated, retaining the width dimension (time axis) as the sequence length; then, the feature vectors are rearranged according to chronological order to ensure that temporal causality is not compromised. Feature sequence data can refer to a set of feature vectors containing temporal evolution information, and its dimensions are typically [sequence length, feature dimension]. For example, if the size of the locally enhanced feature matrix is ​​64×128 (frequency points × time steps), the feature sequence combination unit can reconstruct it into a sequence of length 128, with 64 feature values ​​per time step. Through this dimensional rearrangement and reconstruction, the multi-scale frequency information from time-frequency analysis is preserved, while also meeting the input data format requirements of deep learning sequence models, achieving a seamless bridge from image features to time-series features.

[0048] Step S233: Input the original vibration time series data into the impact feature extraction unit to extract impact features and obtain the impact feature vector.

[0049] In applications, the impact feature extraction unit can refer to a statistical feature calculation module independent of the deep neural network backbone. Its function is to directly extract physical statistics reflecting the mechanical impact intensity from the original vibration time series data, providing an interpretable physical basis for the gating fusion mechanism. By independently extracting impact features, the distortion of physical statistics due to gradient optimization during deep network training is avoided, ensuring the objectivity and accuracy of the gating mechanism.

[0050] Step S234: Input the feature sequence data into the long short-term memory network of the long short-term memory network modeling unit to obtain the temporal feature representation.

[0051] In applications, the Long Short-Term Memory (LSTM) network modeling unit can be a temporal feature extraction branch built on an LSTM architecture. Its role is to capture the dynamic dependencies and local temporal patterns of vibration signals over time. The LSTM network includes three key gating mechanisms: a forget gate, an input gate, and an output gate, as well as a cell state transfer path. In this structure, the forget gate determines how much historical information from the previous time step is discarded, the input gate controls the proportion of new information written at the current time step, and the output gate generates a hidden state output based on the current cell state. By sequentially inputting the obtained feature sequence data into the network, the LSTM can update its internal state at each time step, ultimately outputting a temporal feature representation containing long-term dependencies. For example, when processing the gradual vibration changes during the development of bearing faults, the LSTM can retain early weak anomalous signals through memory units and correlate them with the current strong impact signal, thus forming a complete temporal feature representation. This temporal feature representation is mainly used to characterize the local dynamic evolution of the fault signal.

[0052] Step S235: Input the feature sequence data into the Transformer branch modeling unit to obtain the global feature representation.

[0053] In applications, the Transformer branch modeling unit can refer to a global feature extraction branch built on a self-attention mechanism. Its role is to process the entire feature sequence in parallel, capturing long-range dependencies and global structural features between any two time points in the sequence. The Transformer encoder layer includes a multi-head self-attention mechanism, a feedforward neural network, residual connections, and layer normalization modules. In this structure, the input feature sequence is first positionally encoded to preserve temporal order information. Subsequently, through the calculation of query vectors, key vectors, and value vectors, the self-attention mechanism is used to assign correlation weights between different time steps. The multi-head attention structure allows the model to simultaneously focus on different parts of the sequence in different feature subspaces, thereby enhancing its ability to perceive global patterns. For example, when there are two impact events that are far apart in a vibration signal, the Transformer branch can directly establish the correlation between them through the self-attention mechanism, without being limited by the sequence length, and thus output a global feature representation rich in global contextual information. This global feature representation complements the temporal feature representation, focusing on the macroscopic structure and microscopic evolution of the signal, respectively.

[0054] Step S236: Input the impact feature vector into the gating generation module to obtain the gating factor, and fuse the temporal feature representation and the global feature representation based on the gating factor to obtain the fused feature. The gating factor reflects the impact intensity level of the time-frequency feature sample.

[0055] In application, the gating generation module maps the obtained impact feature vector into a scalar or vector between 0 and 1, i.e., a gating factor. Instead of directly generating the gating factor using a conventional fully connected layer or a standard sigmoid function, the gating generation module is based on the rolling bearing failure impact mechanism. It constructs a normalized impact feature vector using kurtosis, peak factor, high-frequency energy ratio, and envelope spectrum energy. Furthermore, it introduces the impact coupling strength exponent, the impact change in adjacent time windows, and a dual-threshold impact response function to generate the gating factor. This allows the gating factor to reflect the impact intensity, impact abrupt change degree, and periodic fault characteristics of the vibration signal.

[0056] Specifically, let the normalized impact eigenvector of the t-th time window be... for:

[0057] in, Let be the normalized kurtosis of the t-th time window. Let be the normalized peak factor for the t-th time window. Let be the normalized high-frequency energy ratio for the t-th time window. Let be the normalized envelope spectrum energy of the t-th time window.

[0058] Constructing the shock coupling strength index for the t-th time window :

[0059] in, to These are learnable or preset parameters. , , These are used to characterize the coupling relationship between the spike impact, peak impact, high-frequency energy, and envelope spectrum modulation in the t-th time window, respectively.

[0060] Construct the gated response variable for the t-th time window :

[0061] in, to For gating mapping parameters, Used to characterize the amount of impact change between adjacent time windows (the t-th time window and the (t-1)-th time window). The gating factor used to suppress non-physical mutations of the gating factor in adjacent time windows is based on the gating factor of the (t-1)th time window.

[0062] Gating factor for the t-th time window generated based on double-threshold impulse response function : when hour:

[0063] when hour:

[0064] when hour:

[0065] in, For the Sigmoid function, For strong impact threshold, It is a weak impact threshold, and , For strong impact enhancement coefficient, It is a weak impact suppression coefficient. is the gating time smoothing coefficient, and .

[0066] The gating factor is a key parameter reflecting the impact intensity level of the current time-frequency feature sample. Its value directly determines the contribution weight of the temporal feature representation and the global feature representation in the final fused feature. Specifically, when the impact feature vector indicates that the sample has a high-intensity transient impact (such as severe local damage), the gating generation module outputs a larger gating factor, causing the fusion process to focus more on preserving the temporal feature representation output by the LSTM branch. Conversely, when the impact intensity is low or the signal mainly exhibits a trend change, the gating factor is smaller, and the fusion process focuses more on the global feature representation output by the Transformer branch. The fused features are obtained through the formula... The calculation yielded, where As the gating factor, For temporal feature representation, For global feature representation, This represents element-wise multiplication. For example, if the calculated gating factor is 0.8, then 80% of the information in the fused features comes from temporal features and 20% comes from global features, thus achieving adaptive weighting for different fault mechanisms. Through this impact-driven gating fusion mechanism, the model can dynamically adjust the feature emphasis according to the physical characteristics of the signal, significantly improving its diagnostic adaptability under complex operating conditions.

[0067] Step S237: The fully connected layer and non-linear activation function of the fused feature input classification mapping unit are processed to obtain the classification vector, and the running state corresponding to the dimension with the highest probability in the classification vector is determined as the predicted classification label.

[0068] In applications, the classification mapping unit can refer to the decision module at the end of the model, which maps high-dimensional fused features to a specific fault category space. This unit specifically includes at least one fully connected layer and a non-linear activation function (such as the Softmax function). The fully connected layer learns the non-linear mapping relationship between the fused features and fault categories, compressing the fused features to a dimension equal to the number of fault categories. The non-linear activation function converts the output of the fully connected layer into a probability distribution, forming a classification vector. Each dimension in the classification vector corresponds to a preset operating state (such as normal, inner race fault, outer race fault, rolling element fault, etc.), and its value represents the probability that the sample belongs to that state. The process of determining the predicted classification label is to execute the maximum probability principle: traverse all dimensions of the classification vector, find the dimension with the largest value, and use the operating state identifier corresponding to that dimension as the final predicted classification label output. For example, if the classification vector is [0.02, 0.95, 0.01, 0.02], corresponding to four states respectively, then the inner race fault state corresponding to the highest probability of 0.95 is determined as the predicted classification label. Through this series of processes, the model completes end-to-end reasoning from the original vibration signal to a clear fault diagnosis result.

[0069] This application constructs a complete technical chain—signal preprocessing, dual-path parallel modeling, impact-driven fusion, and classification decision-making—through the synergistic effect of the aforementioned technical features. Specifically, the time-frequency local information extraction unit and the feature sequence combination unit work together to transform non-stationary vibration signals into structured sequence data suitable for deep learning, preserving both the energy distribution information in the frequency domain and the evolutionary logic in the time domain. Based on this, the Long Short-Term Memory network modeling unit and the Transformer branch modeling unit work in parallel to capture the local temporal dynamics and global dependencies of the signal, overcoming the limitations of a single model in feature extraction. More importantly, the impact feature extraction unit, independent of the backbone network, calculates impact statistics reflecting the physical essence in real time and converts them into a gating factor through a gating generation module. This gating factor, as prior physical knowledge, dynamically adjusts the fusion ratio of dual-path features, enabling the model to automatically favor temporal features under strong impact conditions and automatically favor global features under weak impact or stable conditions, achieving a deep fusion of data-driven and physical mechanisms. Finally, the classification mapping unit outputs accurate diagnostic results based on the fused high-discriminative features. This collaborative mechanism not only improves the accuracy of fault identification, but also enhances the robustness and interpretability of the model under different noise levels and operating conditions.

[0070] In one embodiment, step S233, which involves inputting the original vibration time series data into an impact feature extraction unit to extract impact features and obtain an impact feature vector, further includes: Step S2331: Calculate the kurtosis, peak factor, high-frequency energy ratio, and envelope spectrum energy of the original vibration time series data.

[0071] In applications, the original vibration time series data can refer to bearing vibration signal segments after length unification and sliding window segmentation. These segments contain time-domain amplitude and frequency-domain distribution information reflecting the bearing's operating state. Kurtosis is a statistical measure characterizing the sharpness of the signal waveform and the significance of the impact pulse. It quantifies the probability of extreme values ​​occurring in the signal distribution; a higher value indicates a stronger transient impact component in the signal. Crest factor is the ratio of the signal peak value to the effective value (root mean square value), reflecting the prominence of the instantaneous peak relative to the average energy level, effectively identifying sudden, strong impact events. High-frequency energy ratio refers to the proportion of signal energy within a preset high-frequency band to the total energy, used to characterize the proportion of high-frequency resonance components caused by mechanical impact or friction, distinguishing between low-frequency background noise and high-frequency fault impacts. Envelope spectrum energy is the sum of energy extracted from the fault characteristic frequency and its harmonics after envelope demodulation of the signal. It characterizes the intensity of periodic impact modulation signals, particularly suitable for identifying periodic impacts caused by local defects in the inner ring, outer ring, or rolling elements.

[0072] Specifically, the processor first reads the raw vibration time series data within the current time window t. The kurtosis value is obtained by calculating the ratio of its fourth central moment to its standard deviation. Simultaneously, the peak factor is obtained by calculating the ratio of the maximum absolute value of the sequence to the root mean square value. Then on Perform a Fast Fourier Transform to count frequencies above the set cutoff frequency. The ratio of the energy of the spectral components to the total spectral energy yields the high-frequency energy ratio. Subsequently, the Hilbert transform was used to obtain... The envelope signal is obtained, and its spectrum is analyzed. The spectral energy within a specific fault band is accumulated to obtain the envelope spectral energy. For example, when a bearing experiences early outer ring spalling failure, periodic decaying pulses will be superimposed on the vibration signal. The calculated kurtosis value may increase from the normal 3.0 to over 8.5, the peak factor from 2.5 to 6.0, and the proportion of high-frequency energy will significantly increase. Meanwhile, the envelope spectrum energy will show obvious energy accumulation at the characteristic frequency of the outer ring failure. Through this multi-dimensional statistical calculation, the impact intensity can be comprehensively quantified from three orthogonal dimensions: amplitude distribution, frequency domain energy distribution, and modulation domain energy, avoiding misjudgments caused by noise interference from a single indicator.

[0073] Step S2332: Construct an impact feature vector based on kurtosis, peak factor, high-frequency energy ratio, and envelope spectrum energy.

[0074] In application, the impact feature vector is a multidimensional array formed by arranging the four independently calculated physical statistics in a preset order, serving as the physical driving input for the subsequent gating fusion mechanism. The construction process of this vector aims to integrate the dispersed time-domain, frequency-domain, and modulation-domain features into a unified mathematical expression, so that the neural network can simultaneously perceive the impact characteristics of different dimensions.

[0075] Specifically, the system will use the kurtosis calculated above. Peak factor High-frequency energy ratio and envelope spectrum energy By combining them sequentially, an initial impact feature vector is formed. Kurtosis and peak factor primarily respond to the transient changes in the time domain of the signal, high-frequency energy ratios respond to the broadband excitation in the frequency domain, while envelope spectrum energy specifically responds to the periodic modulation structure of the signal. This combination allows the constructed impact feature vector to not only contain information about the intensity of the impact but also implicitly contain information about the nature of the impact (such as whether it is random noise or a periodic fault), thus providing a clear physical basis for the subsequent generation of gating factors.

[0076] Step S2333: Normalize the impact feature vector to obtain the final impact feature vector.

[0077] Specifically, the initial impact feature vector is processed using the minimum-maximum normalization or Z-score normalization method. Each component in the vector is processed. For example, for the kurtosis component, the maximum and minimum kurtosis values ​​in the training set samples are used to map it to the [0, 1] interval; for the envelope spectrum energy, it is similarly scaled according to its statistical distribution range. After normalization, the final impact feature vector is obtained. C , Here, P represents the normalized kurtosis, and P represents the normalized peak factor. To normalize the high-frequency energy ratio, To normalize the envelope spectrum energy, all components are now of the same order of magnitude. This step ensures that the input impact feature vector maintains a stable numerical distribution regardless of changes in the amplitude of the original vibration signal. This allows the subsequent gating generation module to focus on learning the logical relationships between features rather than differences in numerical scale. The resulting final impact feature vector... It retains all the key statistical information about the fault impact in the original signal and has good numerical stability, laying a solid foundation for the subsequent accurate generation of gating factors that reflect the impact intensity level.

[0078] This application achieves high-quality extraction and standardized representation of impact features through the synergistic effect of the aforementioned technical features. Specifically, by simultaneously calculating kurtosis, peak factor, high-frequency energy ratio, and envelope spectrum energy, a four-dimensional impact fingerprint is constructed, encompassing time-domain waveform morphology, frequency-domain energy distribution, and modulation-domain periodic characteristics. This solves the problem of traditional single indicators being susceptible to noise interference and having incomplete representation. Furthermore, by constructing these four orthogonal physical quantities as vectors and normalizing them, dimensional barriers are eliminated, allowing impact features of different properties to be expressed in a balanced manner within the same space. This processing method works closely with the subsequent gated fusion unit: the normalized multi-dimensional impact feature vector serves as the physical prior knowledge input to the gated generation module, directly determining the fusion weights of the temporal branch (LSTM) and the global branch (Transformer). This ensures that under strong impact conditions, the model can adaptively emphasize the extraction of temporal dynamic features, while under weak impact or stable conditions, it emphasizes the modeling of global dependencies, thereby significantly improving the robustness and generalization ability of the rolling bearing fault diagnosis model in complex noise environments.

[0079] In one implementation, step S24, which involves performing gated constraint joint optimization on the dual-path parallel hybrid model based on the predicted classification label, includes: Step S241: Calculate the joint optimization loss with gate constraints based on the predicted classification label and the following formula. : ; in, t To predict the time window to which the segmented samples corresponding to the classification label belong, This is the classification loss, used to measure the difference between the predicted category and the true label; in, For gated response consistency loss, The constraint weighting coefficients for the gating response consistency loss are: , The impact strength normalization mapping function, The gate factor; in, To smooth out the stability loss of the gate, The constraint weight coefficients are those for the gated smoothing stability loss. , The gating factor for the current time window. The gating factor for the previous time window; The gradient is calculated based on the joint optimization loss under the gating constraints, and the parameters of the long short-term memory network modeling unit, the Transformer branch modeling unit, and the gating generation module in the dual-path parallel hybrid model are optimized.

[0080] In applications, gated constraint joint optimization loss This can refer to the overall objective function used to guide the parameter updates of a dual-path parallel hybrid model, which is composed of a weighted average of three parts: classification loss, gated response consistency loss, and gated smoothness stability loss. Classification loss It is obtained by comparing the predicted classification labels output by the model with the actual operating state labels of the samples. The cross-entropy loss function is typically used, and its purpose is to measure the model's basic accuracy in fault category identification tasks, ensuring that the model possesses basic discriminative ability. Gated response consistency loss. The loss term is obtained by calculating the squared Euclidean distance between the gating factor g and the impact intensity normalization mapping function f(C). The impact intensity normalization mapping function f(C) is a priori physical quantity calculated based on the extracted impact feature vector C. For example, statistical quantities such as kurtosis and maxima in the impact feature vector can be mapped to the 0-1 interval through linear normalization or a lightweight neural network. The purpose of this loss term is to force the gating factor g to approach the ideal value determined by the physical statistical characteristics during training, preventing the gating mechanism from degenerating into a purely data-driven black-box weight and deviating from the physical essence. (Gating smoothing stability loss) It is achieved by calculating the gating factor of the current time window t. Gating factor compared to the previous time window t-1 The squared difference between the two values ​​is used to constrain the variation of the gating factor between adjacent time segments, preventing drastic jumps in the gating value due to noise interference, thereby ensuring the temporal continuity of the diagnostic results. For example, when setting the constraint weight coefficient... It is 0.5. When the value is 0.3, if a sample's impact feature is significant enough to cause f(C) to approach 1, while the network-generated g is only 0.2, then... This will generate a large gradient, forcing g to move closer to 1; at the same time, if the previous time... If it is 0.8, then It will inhibit The tendency to abruptly jump to 1.0 is mitigated, allowing for a smooth transition. Through the construction of this ternary joint loss, synergistic optimization of classification accuracy, physical consistency, and temporal stability is achieved.

[0081] The above steps aim to address the issue of response distortion or temporal jitter in gating factors under single-classification loss by constructing a multi-dimensional loss function system. The main execution entity is the optimizer module during training, which calculates the comprehensive loss value in real time based on the predicted classification label of the current batch and historical gating states. By introducing a physical prior f(C) as an anchor point and combining it with temporal smoothing constraints, this step guides the model to automatically enhance the weights of the Long Short-Term Memory (LSTM) network modeling units under strong impact conditions, while increasing the contribution of the Transformer branch modeling units under weak impact or stable conditions. This ensures a smooth and natural switching process, effectively avoiding frequent oscillations in diagnostic results and providing an optimization objective with rich physical constraint information for subsequent parameter gradient calculations.

[0082] Step S242 involves calculating the gradient based on the gated joint optimization loss and optimizing the parameters of the Long Short-Term Memory network modeling unit, Transformer branch modeling unit, and gated generation module in the dual-path parallel hybrid model. In practice, parameter optimization can refer to using the backpropagation algorithm to optimize the gated joint optimization loss calculated above. The process of iteratively updating the weight parameters of each network unit. Specifically, the system first updates the total loss... The partial derivatives of the parameters at each layer of the model are calculated to obtain gradient information. This gradient information not only includes the direction for improving classification accuracy but also incorporates constraints that make the gating factor approximate the physical impact intensity and maintain temporal smoothness. Subsequently, stochastic gradient descent or its variants (such as the Adam optimizer) are used to update the weights of the forget gate, input gate, and output gate in the Long Short-Term Memory network modeling unit, update the self-attention matrix and feedforward network weights in the Transformer branch modeling unit, and update the connection weights in the gating generation module that map the impact features to the gating factor. For example, when When the value is large, the gradient generated by backpropagation will significantly adjust the parameters of the gating generation module, making its output g value more accurately reflect the strength of the input impact feature vector C; when When the gradient is large, the gradient update tends to smooth the gating generation strategy of adjacent time steps. Through this joint optimization mechanism, the Long Short-Term Memory network modeling unit and the Transformer branch modeling unit no longer learn features in isolation, but form an adaptive collaborative working mode for the impact characteristics of the current vibration signal under the dynamic adjustment of the gating generation module.

[0083] This application achieves model training driven by multiple objectives through the construction of a joint optimization mechanism comprising classification loss, gating response consistency loss, and gating smoothness stability loss. Based on this, the classification loss ensures the model's basic ability to identify fault categories, while the gating response consistency loss, as a physical constraint, forces the gating factor to maintain a strong correlation with the impact statistical characteristics of the vibration signal, preventing the gating mechanism from deviating from its physical nature during training. This allows the model to reliably activate the long short-term memory network branch, which excels at capturing transient features, when facing strong impact signals. Simultaneously, the gating smoothness stability loss further constrains the evolution trajectory of the gating factor over time, avoiding drastic fluctuations in the gating value due to signal noise, thereby ensuring the consistency of diagnostic results in continuous monitoring scenarios. By leveraging the weighted combination of these three types of loss functions, the long short-term memory network modeling unit, the Transformer branch modeling unit, and the gated generation module in the dual-path parallel hybrid model are trained together into an organic whole. This retains the powerful feature extraction capabilities of deep learning while incorporating physical rules defined by expert knowledge. Ultimately, when dealing with non-stationary, noisy rolling bearing vibration signals, the model can adaptively balance the contribution ratio of local temporal features and global dependent features, significantly improving the accuracy, interpretability, and engineering applicability stability of fault diagnosis.

[0084] In one embodiment, step S24 further includes: Step S243: In each training iteration, perform statistical analysis on the gating factor distribution of the time-frequency feature samples of the current training batch and generate statistical analysis results.

[0085] In practice, statistical analysis can refer to the gating factor generated for all samples in the current batch. Statistical calculations and distribution characteristic assessments were performed to quantify the activity levels of each branch and the consistency of the physical response of the gating mechanism in the dual-path parallel hybrid model. Gating factor. It is a scalar or vector parameter obtained by inputting the impact feature vector into the gating generation module. Its value directly determines the contribution weight of the temporal feature representation (from the Long Short-Term Memory network modeling unit) and the global feature representation (from the Transformer branch modeling unit) in the fused feature. The statistical analysis results specifically include two core indicators: one is the branch weight suppression probability, which is the gating factor in the current batch. This leads to a certain path feature weight ( or The first factor is the percentage of samples with a weight lower than the preset minimum weight; the second is the physical correlation coefficient, i.e., the calculation of the gating factor. Sequence and corresponding shock eigenvector (or its normalized mapping value) The Pearson correlation coefficient or Spearson rank correlation coefficient between the samples. For example, setting the preset minimum weight to 0.1, the preset maximum probability to 80%, and the preset correlation threshold to 0.7; if in the current iteration of 128 samples, 110 samples have a Pearson correlation coefficient or Spearson rank correlation coefficient. If the value is less than 0.1, the branch weight suppression probability is approximately 85.9%; if calculated... A correlation coefficient of 0.45 with the impact intensity index indicates that the gating factor has failed to effectively follow changes in impact intensity. This step aims to monitor the health status of the model in real time during training, preventing a branch from failing for a long period due to gradient vanishing or local optima, thereby ensuring that the dual-path structure always works collaboratively.

[0086] Step S244: If the statistical analysis results indicate that the probability of any feature weight being lower than the preset minimum weight is greater than the preset maximum probability, or the correlation coefficient between the gate factor and the impact intensity corresponding to the impact feature vector is less than the preset correlation threshold, then at least one of the following operations is performed: adjusting the constraint weight coefficient, optimizing the learning rate of the gate generation module, and enhancing the gradient update ratio of the feature whose weight is lower than the preset minimum weight is greater than the preset maximum probability.

[0087] In application, the adjustment operation is an adaptive correction mechanism triggered by the aforementioned statistical analysis results, used to dynamically adjust the training strategy to correct model degradation problems. When the branch weight suppression probability exceeds the preset maximum probability, it means that one of the two paths (such as the LSTM branch or the Transformer branch) is over-suppressed, and the model tends to degenerate into a single-path model. At this time, it is necessary to perform an operation to enhance the gradient update ratio of the weak branch. Specifically, this can be achieved by multiplying the backpropagation gradient of that branch by an amplification factor greater than 1, forcing the model to focus on the ignored feature dimensions; or by adjusting the constraint weight coefficient. (Weight of gating response consistency loss), increase The value is used to strengthen the physical dependence of the gating factor on the impact characteristics, forcing the gating mechanism to reallocate weights. When the correlation coefficient between the gating factor and the impact intensity is less than the preset correlation threshold, it indicates that the gating generation module has failed to correctly learn the mapping relationship that emphasizes temporal sequence when the impact is strong and global sequence when the impact is weak. At this time, it is necessary to optimize the learning rate of the gating generation module, such as temporarily reducing the learning rate of the module to avoid parameter divergence, or using an optimizer with momentum to reset its state to escape local minima. For example, if the probability of detecting that the LSTM branch weight is lower than 0.1 reaches 90% (greater than the set 80%), the system automatically adjusts the gradient update ratio coefficient of the branch from 1.0 to 1.5, and... The learning rate is increased from 0.5 to 0.8; if the correlation coefficient is detected to be only 0.3 (less than the set 0.7), the learning rate of the gating generation module is reduced from 0.001 to 0.0005 and a regularization term is added. This step, by introducing a dynamic intervention mechanism, solves the problem of poor adaptability of fixed hyperparameters during training on complex non-stationary data, effectively avoiding the model from falling into a rigid state dominated by a single feature, and ensuring the effectiveness and robustness of the impact-driven gating mechanism throughout the entire training lifecycle.

[0088] This application achieves intelligent self-healing and dynamic optimization of the model training process through the synergistic effect of the aforementioned technical features. By statistically analyzing the distribution characteristics of the gating factors in real time during each iteration, the originally black-box neural network training process is transformed into an observable and controllable closed-loop system. When statistical analysis reveals an imbalance in the dual-path feature weights or a disconnect between the gating response and the physical mechanism (impact intensity), the system can immediately trigger targeted adjustment strategies: on the one hand, by enhancing the gradient update ratio of weak branches, the suppressed feature extraction path is forcibly activated to prevent the model from degenerating into a simple LSTM or Transformer single model, ensuring the parallel mining capability of temporal dynamic features and global dependency features; on the other hand, by dynamically adjusting the constraint weight coefficients and learning rate, the mapping deviation of the gating generation module is corrected, ensuring that it strictly follows the physical laws driven by impact. This dynamic diagnosis and treatment mechanism not only overcomes the defect that static loss function weights may fail in long-term training, but also significantly improves the model's transfer adaptability when facing new working conditions and new fault types of data, ultimately enabling the trained dual-path parallel hybrid model to maintain stable structural response characteristics and high-precision diagnostic performance under different impact intensity levels.

[0089] In one embodiment, such as Figure 4 As shown, step S21 further includes the following steps: Step S211: Collect the vibration signal of the rolling bearing to obtain the original vibration time series data. The original vibration time series data includes the sampling timestamp, vibration amplitude and corresponding operating status identifier.

[0090] In applications, raw vibration time series data refers to a digital signal sequence continuously collected during bearing operation by vibration sensors installed in the rolling bearing housing or adjacent structural locations, and then converted from analog to digital. This data originates from equipment lifecycle monitoring data or experimental bench simulation data, and its purpose is to provide the most basic physical quantity input for subsequent fault diagnosis models. Specifically, the acquisition process must ensure that the sampling frequency meets the Nyquist sampling theorem and the requirements of bearing fault characteristic frequency analysis to ensure that the signal fully reflects the local defects and structural changes of the bearing. The sampling timestamp records the precise acquisition time of each data point, supporting the time traceability of subsequent diagnostic results; the vibration amplitude reflects the intensity information of the bearing vibration and is the direct basis for extracting fault features; the operating status identifier is a manual or automatic labeling of the current signal segment's operating condition (such as normal, inner ring fault, outer ring fault, rolling element fault, etc.), serving as a true label for supervised learning. For example, in a nuclear main pump bearing monitoring scenario, vibration signals are acquired at a sampling frequency of 20kHz, and each data record includes a microsecond-level timestamp, an acceleration amplitude in m / s², and an inner ring wear status label. By clearly defining the three core elements of the data, the traceability of data lineage is ensured, laying the data foundation for the subsequent generation of structured diagnostic results containing timestamps and status identifiers. This step aims to obtain high-quality, clearly labeled raw data and eliminate model training bias caused by missing data or inconsistent labels.

[0091] Step S212: Unify the sequence length of the original vibration time series data to obtain time series data with consistent length.

[0092] In this context, time series data of uniform length refers to processing raw signal segments with different runtimes into standardized sequences with the same number of data points. Since the duration of signals collected under different operating conditions in actual engineering may vary, directly inputting them into the model would lead to dimensionality mismatch. Therefore, this step aims to eliminate the compatibility barrier of sample length differences on the model's input layer. The method for setting a uniform sequence length can be determined based on the model's preset input dimension. Specific operations include truncation or zero-padding: when the original signal length is greater than the preset length, the central or initial segment containing effective impact characteristics is truncated; when the original signal length is less than the preset length, zero values ​​are added to the end of the sequence until the preset length is reached. For example, if the model input layer requires a sequence length of 4096 points, but a normal state signal only has 3800 points, then 296 zeros are added to the end; if a fault signal is as long as 5000 points, then the first 4096 points are truncated, or the middle 4096 points are truncated based on the energy concentration area. This standardization process ensures that all samples have a uniform baseline scale before entering the sliding window segmentation, avoiding calculation errors caused by varying sequence lengths and significantly improving the efficiency of batch training.

[0093] Step S213: Based on the preset window length and preset sliding step size, the time series data is sampled in segments by sliding window and normalized to obtain segmented samples.

[0094] In applications, segmented samples refer to short signal segments extracted from time series of uniform length and normalized in amplitude. This step aims to expand the number of samples through overlap sampling without altering the original signal characteristics, enhancing the data's coverage of local impact features and suppressing interference caused by differences in amplitude dimensions between different samples. The preset window length defines the time span of a single sample, typically set to cover at least one complete bearing rotation cycle or a typical impact response decay time. The preset sliding step size controls the movement distance between adjacent windows; a smaller step size results in a higher overlap rate, thus increasing sample density. Normalization specifically employs Z-Score normalization or max-min normalization to map the signal amplitude to a uniform interval (e.g., 0 to 1, or a mean of 0 and a variance of 1). For example, setting the window length to 1024 points and the sliding step size to 256 points (i.e., 75% overlap), a 4096-point sequence can be truncated to obtain 13 segmented samples. Subsequently, the mean is subtracted from each sample and divided by the standard deviation to eliminate differences in absolute amplitude values ​​caused by load fluctuations. This approach not only maximizes the use of limited fault data, especially capturing the details of transient impact events, but also effectively avoids overfitting the model to the absolute amplitude of the signal, enhancing the model's generalization ability under different load conditions.

[0095] Step S214: Perform time-frequency analysis on the segmented samples to obtain time-frequency feature samples. The time-frequency feature samples have the corresponding running status identifiers of the original vibration time series data.

[0096] In this process, time-frequency feature samples refer to the transformation of one-dimensional time-domain segmented samples into a two-dimensional matrix representation that simultaneously contains time and frequency information. Since rolling bearing fault signals exhibit significant non-stationarity and transient impact characteristics, relying solely on time-domain waveforms is insufficient to fully characterize the fault evolution. Therefore, this step aims to increase the signal's dimensionality and accurately extract the transient impact and modulation features of the fault using a joint time-frequency distribution. Specifically, the time-frequency analysis employs Continuous Wavelet Transform (CWT). By selecting a mother wavelet function with good time-frequency localization characteristics (such as the Morlet or Gabor wavelet), projection analysis is performed on the segmented samples at different scales and time positions to calculate wavelet coefficients and form a two-dimensional time-frequency diagram. The operating status label remains unchanged during the transformation process, directly inherited from the original segmented samples, ensuring a one-to-one correspondence between features and labels. For example, performing Morlet wavelet transform on the normalized segmented samples, with a scale range of 1 to 128, generates a time-frequency energy distribution map of size 128×1024. In the map, the horizontal axis represents time, the vertical axis represents frequency, and the color intensity represents energy intensity, clearly demonstrating the energy accumulation phenomenon of fault impact in specific frequency bands. Through CWT processing, background noise and fault impact components can be effectively separated, solving the problem that traditional Fourier transform cannot locate transient events. This provides intuitive and information-rich input data for subsequent convolutional neural networks to extract local frequency band energy changes.

[0097] Step S215: Construct a time-frequency feature sample set based on the time-frequency feature samples. The time-frequency feature sample set includes a training set, a validation set, and a test set.

[0098] The time-frequency feature sample set refers to a standardized dataset formed by dividing all generated time-frequency feature samples according to a preset ratio. This step provides an independent and consistently distributed data source for model training, hyperparameter tuning, and final performance evaluation, preventing overfitting and objectively evaluating its generalization ability. The construction process first randomly shuffles all time-frequency feature samples with state labels to eliminate bias caused by the data collection order. Then, the samples are divided into training, validation, and test sets according to a preset ratio (e.g., 7:2:1 or 8:1:1). The training set drives iterative updates of model parameters; the validation set monitors model performance and adjusts hyperparameters (such as learning rate and constraint weight coefficients) during training; the test set is only used to evaluate the final diagnostic accuracy after model training is complete. For example, after randomly shuffling 10,000 time-frequency feature samples, 7,000 are selected as the training set and input into a dual-path parallel hybrid model for gated constraint joint optimization; 2,000 are selected as the validation set to monitor loss convergence; and the remaining 1,000 are used as the test set to calculate the final classification accuracy and confusion matrix. This rigorous data partitioning mechanism ensures the robustness of the model under different operating conditions, providing reliable data support for actual deployment on engineering sites.

[0099] This application constructs an end-to-end reproducible pipeline from raw sensor data to model-ready features through the synergistic effect of the technical features described above. Specifically, the timestamps and state identifiers in the raw vibration time series data provide a metadata foundation for the traceability of subsequent diagnostic results; the unified sequence length operation eliminates the dimensionality barrier of time series modeling, enabling standardized execution of sliding window segmentation; the combination of sliding window segmentation and normalization not only maximizes the sample coverage of impact events with limited data, but also suppresses amplitude interference caused by changes in operating conditions; continuous wavelet transform time-frequency analysis transforms one-dimensional non-stationary signals into two-dimensional structured representations, naturally adapting to the convolution operation of subsequent time-frequency local information extraction units, accurately characterizing the transient impact and modulation features of bearing failures; finally, the standardized dataset partitioning ensures the effectiveness of supervised learning, enabling the dual-path parallel hybrid model based on the impact feature-driven gate control mechanism to learn fully on the training set, be stably optimized on the validation set, and exhibit excellent generalization performance on the test set.

[0100] In one embodiment, before step S212, which unifies the sequence length of the original vibration time series data, the method further includes: Step S216: Generate a random noise signal based on the target signal-to-noise ratio and superimpose it onto the original vibration time series.

[0101] In applications, the target signal-to-noise ratio (SNR) can refer to a pre-set ratio of signal energy to noise energy, used to simulate the interference intensity level of rolling bearings under different complex operating conditions. This target SNR can be set according to the noise environment of the actual application scenario. For example, in the high-noise environment of a nuclear industry main coolant pump, a lower target SNR (such as 0dB or 5dB) can be set, while in a general industrial fan environment, a higher target SNR (such as 20dB) can be set. Random noise signals typically use Gaussian white noise, whose spectrum covers a wide frequency band, and can simulate unpredictable electromagnetic interference and mechanical coupling vibrations in real environments. Specifically, the signal power of the original vibration time series is first calculated, and then... Calculate the required noise power, where The target signal-to-noise ratio (SNR) is set in dB. Next, a random noise sequence with zero mean and variance matched to the noise power is generated. Finally, the generated random noise signal is superimposed point-by-point onto the original vibration time series to obtain a noisy vibration time series. This method of actively injecting controllable noise early in the data preprocessing stage enables the construction of training samples covering multiple levels of interference. This forces subsequent feature extraction units and the dual-path parallel hybrid model to learn to extract robust fault features in a strong noise environment, thereby significantly improving the model's generalization ability and anti-interference performance under complex field conditions.

[0102] This application generates random noise signals based on the target signal-to-noise ratio and superimposes them onto the original vibration time series. This, combined with subsequent steps such as sequence length unification, sliding window segmentation, and time-frequency feature construction, forms a complete data augmentation and robust pre-training mechanism. This mechanism not only solves the problem of poor model generalization caused by the inconsistent distribution of clean laboratory data and high-noise field data, but also ensures the full coupling of noise features with subsequent time-frequency analysis, impact feature extraction, and gating fusion processes. Based on this, the model can fully adapt to background disturbances of varying intensities during the training phase, enabling the impact-driven gating mechanism to accurately respond to real fault impacts even when faced with real noise disturbances. This effectively avoids overfitting the model to low-noise data and significantly improves the stability and reliability of the rolling bearing fault diagnosis system under strong noise and non-stationary operating conditions.

[0103] In one embodiment, the method further includes Step S4: Calculate the overall classification accuracy and recall of the dual-path parallel hybrid model based on the validation set data in the time-frequency feature sample set, and calculate the F1 score and generate a confusion matrix to evaluate the classification performance of the dual-path parallel hybrid model.

[0104] In practice, this step aims to quantify the model's generalization ability and fine-grained recognition accuracy on unseen data through multi-dimensional statistical indicators.

[0105] Specifically, the validation set data reserved in the time-frequency feature sample set is input into the already trained dual-path parallel hybrid model, keeping the model parameters frozen, and forward propagation is performed to obtain the predicted classification labels. Subsequently, the matching between the predicted labels and the true labels is statistically analyzed, and the aforementioned indicators are calculated respectively. For example, if the validation set contains 1000 samples, of which 200 are inner ring fault samples, and the model correctly identifies 195, then the recall rate for inner ring faults is 97.5%; if the confusion matrix shows that 10 normal state samples are misclassified as outer ring faults, it suggests that the model needs to further distinguish the feature differences between normal vibration and slight outer ring impact.

[0106] Step S5: To verify the stability of the model in different training rounds, record the following data: the trend of loss during training, the trend of accuracy on the validation set, and the fluctuation of the gating factor distribution. If the model's performance fluctuates little in multiple training rounds, it indicates that the model has good stability.

[0107] Specifically, after each training iteration, the current loss value and validation set accuracy are automatically recorded, and the gating factor g value of all samples in the current batch is extracted for statistical analysis. For example, a sliding window is set to statistically analyze the variance of the gating factor. If the variance of the gating factor exceeds a preset threshold (e.g., 0.15) under normal operating conditions, or if there is a drastic jump between adjacent time windows, it indicates that the gating mechanism's response to the impact intensity is unstable, which may lead to abnormal oscillations in the dual-path feature fusion weights. By continuously tracking these dynamic indicators, non-steady-state behaviors during model training can be effectively identified. If the fluctuation range of various indicators in multiple training iterations is controlled within a preset range (e.g., accuracy fluctuation is less than 1%, and the gating factor distribution remains consistent), the model is judged to have good engineering applicability and stability, and can adapt to the high reliability requirements of equipment condition monitoring in complex environments such as the nuclear industry.

[0108] Table 1 shows the statistical results of the performance of different models under multiple experiments.

[0109]

[0110] 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.

[0111] In one embodiment, such as Figure 5As shown, this application also provides a dual-path parallel hybrid model architecture 500, including a time-frequency local information extraction unit 501, a feature sequence combination unit 502, an impact feature extraction unit 503, a long short-term memory network modeling unit 504, a Transformer branch modeling unit 505, an impact-driven gating fusion unit 506, and a classification mapping unit 507.

[0112] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 6 As shown, the computer device 6 of this embodiment includes: at least one processor 60 ( Figure 6 (Only one is shown in the diagram) a processor, a memory 61, and a computer program 62 stored in the memory 61 and executable on the at least one processor 60, wherein the processor 60 executes the computer program 62 to implement the steps in any of the above embodiments of the rolling bearing fault diagnosis method.

[0113] The computer device 6 can be a desktop computer, laptop, handheld computer, or cloud server, etc. This computer device may include, but is not limited to, a processor 60 and a memory 61. Those skilled in the art will understand that... Figure 6 The computer device 6 is merely an example and does not constitute a limitation on the computer device 6. It may include more or fewer components than shown, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0114] The processor 60 can be a Central Processing Unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0115] In some embodiments, the memory 61 may be an internal storage unit of the computer device 6, such as a hard disk or memory of the computer device 6. In other embodiments, the memory 61 may be an external storage device of the computer device 6, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 6. Furthermore, the memory 61 may include both internal and external storage units of the computer device 6. The memory 61 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory 61 can also be used to temporarily store data that has been output or will be output.

[0116] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0117] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0118] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0119] like Figure 7 As shown, this application embodiment provides a computer program product 70, including a computer program 62. When the computer program 62 is run, the steps in the above embodiments of the rolling bearing fault diagnosis methods are executed.

[0120] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a device / terminal equipment, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0121] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0122] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0123] In the embodiments provided in this application, it should be understood that the disclosed terminal devices and methods can be implemented in other ways. For example, the terminal device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0124] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0125] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for diagnosing rolling bearing faults, characterized in that, include: The vibration signal of the rolling bearing to be diagnosed is acquired and the corresponding time-frequency characteristics are constructed based on the vibration signal; Training a dual-path parallel hybrid model based on the time-frequency feature sample set of rolling bearings and the impact feature-driven gate control mechanism. The dual-path parallel hybrid model includes a time-frequency local information extraction unit, a feature sequence combination unit, an impact feature extraction unit, a long short-term memory network modeling unit, a Transformer branch modeling unit, an impact-driven gate control fusion unit, and a classification mapping unit. The time-frequency features are input into the dual-path parallel hybrid model to obtain the diagnostic results, which include fault category labels, predicted probabilities, timestamps, and corresponding operating status identifiers. The system comprises the following components: a time-frequency local information extraction unit for enhancing local frequency band energy changes and transient impact edges in time-frequency features to obtain a locally enhanced feature matrix; a feature sequence combination unit for performing dimensional rearrangement and time series reconstruction on the locally enhanced feature matrix to obtain feature sequence data; an impact feature extraction unit for extracting impact features from the original vibration time series data to obtain an impact feature vector; a long short-term memory network modeling unit for generating a temporal feature representation of the feature sequence data; a Transformer branch modeling unit for generating a global feature representation of the feature sequence data; an impact-driven gating fusion unit for generating a gating factor based on the impact feature vector and fusing the temporal feature representation and the global feature representation based on the gating factor to obtain a fused feature, wherein the gating factor reflects the impact intensity level of the time-frequency feature sample; and a classification mapping unit for generating a classification vector based on the fused feature and determining the running state corresponding to the dimension with the highest probability in the classification vector as the predicted classification label. The training of the time-frequency feature sample set based on rolling bearings for a dual-path parallel hybrid model based on an impact feature-driven gate control mechanism includes: The vibration signal of the rolling bearing is collected to obtain the original vibration time series data, and a corresponding time-frequency feature sample set is constructed based on the original vibration time series data. The time-frequency feature sample set includes a training set, and the training set includes time-frequency feature samples. A dual-path parallel hybrid model based on the impact feature-driven gate control mechanism is constructed, which includes a time-frequency local information extraction unit, a feature sequence combination unit, an impact feature extraction unit, a long short-term memory network modeling unit, a Transformer branch modeling unit, an impact-driven gating fusion unit, and a classification mapping unit. The time-frequency feature samples in the training set are input into the dual-path parallel hybrid model to obtain the predicted classification label; Based on the predicted classification labels, the dual-path parallel hybrid model is jointly optimized with gating constraints. Return to the step of inputting the time-frequency feature samples in the training set into the dual-path parallel hybrid model to obtain the predicted classification label, until the dual-path parallel hybrid model converges after joint optimization; The step of inputting the time-frequency feature samples from the training set into the dual-path parallel hybrid model to obtain the predicted classification label includes: The time-frequency feature samples are input into the convolutional local enhancement unit in the time-frequency local information extraction unit to enhance the local frequency band energy changes and transient impact edges to obtain the locally enhanced feature matrix. The feature sequence data is obtained by performing dimensional rearrangement and time series reconstruction on the locally enhanced feature matrix through the feature sequence combination unit. The original vibration time series data is input into the impact feature extraction unit to extract impact features and obtain an impact feature vector. The feature sequence data is input into the long short-term memory network of the long short-term memory network modeling unit to obtain a temporal feature representation; The feature sequence data is input into the Transformer branch modeling unit to obtain a global feature representation; The impact feature vector is input into the gating generation module to obtain the gating factor, and the time-series feature representation and the global feature representation are fused based on the gating factor to obtain the fused feature. The gating factor reflects the impact intensity level of the time-frequency feature sample. The fused features are input into the fully connected layer and nonlinear activation function of the classification mapping unit to obtain a classification vector, and the running state corresponding to the dimension with the highest probability in the classification vector is determined as the predicted classification label. The original vibration time series data is input into the impact feature extraction unit to extract impact features and obtain an impact feature vector, including: Calculate the kurtosis, peak factor, high-frequency energy ratio, and envelope spectral energy of the original vibration time series data; An impact feature vector is constructed based on the kurtosis, peak factor, high-frequency energy ratio, and envelope spectrum energy. The impact feature vector is normalized to obtain the final impact feature vector; The gating constraint joint optimization of the dual-path parallel hybrid model based on the predicted classification label includes: The gated joint optimization loss is calculated based on the predicted classification labels and the following formula. : ； in, t To predict the time window to which the segmented samples corresponding to the classification label belong, This is the classification loss, used to measure the difference between the predicted category and the true label; in, For gated response consistency loss, The constraint weighting coefficients for the gating response consistency loss are: , The impact strength normalization mapping function, The gate factor; in, To smooth out the stability loss of the gate, The constraint weight coefficients are those for the gated smoothing stability loss. , The gating factor for the current time window. The gating factor for the previous time window; The gradient is calculated based on the joint optimization loss under the gating constraints, and the parameters of the long short-term memory network modeling unit, the Transformer branch modeling unit, and the gating generation module in the dual-path parallel hybrid model are optimized.

2. The rolling bearing fault diagnosis method as described in claim 1, characterized in that, The gating constraint joint optimization of the dual-path parallel hybrid model based on the predicted classification label further includes: In each training iteration, the distribution of gating factors of the time-frequency feature samples of the current training batch is statistically analyzed and statistical analysis results are generated. If the statistical analysis results indicate that the probability that the weight of any feature in the temporal feature representation and the global feature representation is lower than the preset minimum weight is greater than the preset maximum probability, or the correlation coefficient between the gating factor and the impact intensity corresponding to the impact feature vector is less than the preset correlation threshold, then at least one of the following operations is performed: adjusting the constraint weight coefficient, optimizing the learning rate of the gating generation module, and enhancing the gradient update ratio of the feature whose weight is lower than the preset minimum weight is greater than the preset maximum probability.

3. The rolling bearing fault diagnosis method as described in claim 1 or 2, characterized in that, The process involves acquiring vibration signals from rolling bearings to obtain raw vibration time-series data and constructing a corresponding time-frequency feature sample set based on the raw vibration time-series data, including: The vibration signal of the rolling bearing is collected to obtain the original vibration time series data, which includes the sampling timestamp, vibration amplitude and corresponding operating status identifier; By unifying the sequence length of the original vibration time series data, time series data with consistent length are obtained. Based on a preset window length and a preset sliding step size, the time series data is segmented and sampled using a sliding window, and then normalized to obtain segmented samples. Time-frequency analysis is performed on the segmented samples to obtain time-frequency feature samples, and the time-frequency feature samples have the corresponding running status identifiers of the original vibration time series data; A time-frequency feature sample set is constructed based on the aforementioned time-frequency feature samples. The time-frequency feature sample set includes a training set, a validation set, and a test set.

4. The rolling bearing fault diagnosis method as described in claim 3, characterized in that, To obtain time series data of consistent length by unifying the sequence length of the original vibration time series data, the process previously included: A random noise signal is generated based on the target signal-to-noise ratio and superimposed on the original vibration time series.

5. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 4.

6. A computer program product, characterized in that, Includes a computer program, which, when run, causes the method described in any one of claims 1 to 4 to be performed.

Images