Bearing fault diagnosis method and device based on time-frequency characteristics and physical information neural network

Abstract

The invention discloses a bearing fault diagnosis method and device based on time-frequency characteristics and a physical information neural network, and the method comprises the steps: collecting vibration signals of a bearing under different operation conditions, carrying out the time-frequency analysis, and generating a time-frequency spectrogram; based on the physical model and the operation parameters of the bearing, extracting physical characteristics related to the fault, and carrying out normalization and dimension reduction processing to obtain enhanced physical characteristics; establishing association mapping between enhanced physical features and fault modes, analyzing a corresponding relation between fault feature frequencies and vibration signals, and generating a structured physical feature vector; a deep neural network fusing the time-frequency features and the physical information is constructed, the marked fault data set is used for training, and a final diagnosis model is generated; according to the bearing fault diagnosis method, the time-frequency characteristics and the physical information are fused, so that high-precision, robust and interpretable bearing fault diagnosis is realized, and the bearing fault diagnosis method is suitable for complex industrial scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and in particular to a bearing fault diagnosis method and device based on time-frequency characteristics and physical information neural network. Background Art

[0002] Bearings are core components of mechanical equipment, widely used in motors, fans, pumps, gearboxes, and other equipment. Their operating condition directly impacts equipment reliability and lifespan. Bearing failures (such as inner ring cracks, outer ring wear, and rolling element damage) can cause increased vibration, decreased efficiency, and even equipment downtime, resulting in significant economic losses. Therefore, accurate and timely bearing fault diagnosis technology is crucial in the industrial sector.

[0003] Traditional diagnostic methods rely on time-domain or frequency-domain analysis of vibration signals, identifying fault modes by extracting statistical or spectral features. However, vibration signals in industrial environments are often non-stationary and subject to strong noise interference. Traditional methods have limited feature extraction capabilities, which can easily lead to misjudgments or missed detections. In recent years, deep learning methods (such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs)) have significantly improved diagnostic efficiency by automatically extracting features from vibration signals or time-frequency graphs. However, most of these methods are purely data-driven, relying on large amounts of labeled data and lacking understanding of the physical mechanisms of bearings. This results in insufficient generalization and robustness, and the physical meaning of faults is difficult to interpret, limiting their credibility in industrial applications.

[0004] Therefore, there is an urgent need for an intelligent diagnostic method that integrates time-frequency features and physical information to improve the accuracy, robustness and interpretability of bearing fault diagnosis and meet the needs of modern industry for high-reliability diagnosis. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems in the prior art.

[0006] The technical solution adopted by the present invention to solve the technical problem is to provide a bearing fault diagnosis method based on time-frequency characteristics and physical information neural network, including the following steps:

[0007] Collect and pre-process bearing vibration signals under different operating conditions, including normal conditions and various failure modes;

[0008] The pre-processed vibration signal is subjected to time-frequency analysis using continuous wavelet transform, and the time-frequency features are extracted to generate a time-frequency spectrum, which is then standardized.

[0009] Based on the bearing's physical model and operating parameters, fault-related physical features are extracted, including fault characteristic frequency, speed, load, temperature, and bearing geometric parameters. Enhanced physical features are obtained through z-score normalization and principal component analysis dimensionality reduction.

[0010] Establish an association mapping between enhanced physical features and fault modes, construct a semantic representation of physical features by analyzing the correspondence between fault feature frequencies and vibration signals, and generate a structured physical feature vector.

[0011] A deep neural network that integrates time-frequency features and physical information is constructed and trained using a labeled fault dataset to generate the final diagnostic model. The deep neural network adopts a dual-branch architecture. The first branch receives the standardized time-frequency spectrum and extracts deep time-frequency features based on ResNet18. The second branch receives the physical feature vector and uses Transformer to interact between features to obtain a second physical feature vector with the same dimension as the time-frequency features. The deep time-frequency features and the second physical feature vector are adaptively weighted using the ECA attention mechanism before entering the feature fusion layer, where fused features are obtained through weighted concatenation. The fully connected layer maps the fused features to fault categories and outputs the classification probability.

[0012] The final diagnostic model is used to achieve accurate classification and condition assessment of bearing faults.

[0013] Preferably, the multiple failure modes include: inner ring damage, outer ring damage, rolling element damage and retainer damage.

[0014] Preferably, the preprocessing includes denoising, signal segmentation and data enhancement.

[0015] Preferably, the continuous wavelet transform is used to perform time-frequency analysis on the preprocessed vibration signal, and the time-frequency features are extracted to generate a time-frequency spectrum. Specifically, the Morlet wavelet is used to obtain the time-frequency features, and then a high-resolution time-frequency spectrum is generated by continuously changing the scale and translation parameters.

[0016] Preferably, the bearing geometric parameters include rolling element diameter, pitch circle diameter, and contact angle.

[0017] Preferably, the bearing fault characteristic frequencies include inner ring fault frequency BPFI, outer ring fault frequency BPFO, rolling element fault frequency BSF and cage fault frequency FTF, which are respectively expressed as:

[0018]

[0019] Among them, f r is the bearing speed (unit: rpm), n is the number of rolling elements, d is the rolling element diameter, D is the bearing pitch circle diameter, and θ is the contact angle.

[0020] Preferably, the Transformer models complex interactions between features through a 2-layer encoder, each layer of encoder includes 8 self-attention and feedforward networks.

[0021] Preferably, the ECA attention mechanism is expressed as:

[0022]

[0023] Among them, z c is the description of the cth channel, x c (i, j) is the value of the c-th channel feature map at position (i, j), H and W are the height and width of the feature map, Conv1D is a one-dimensional convolution, the convolution kernel size k is adaptively selected, σ is the Sigmoid activation function, w c is the channel weight, is the weighted feature map. The ECA mechanism enhances the expression of fault-related features through low-complexity channel attention.

[0024] Preferably, the deep neural network integrating time-frequency features and physical information is constructed and trained using a labeled fault data set to generate a final diagnostic model, and the training includes the following steps:

[0025] The cross-entropy loss function is used to optimize the model, and the network weights are initialized through the pre-trained model in combination with the transfer learning strategy to enhance the generalization ability in small sample scenarios;

[0026] Verify the diagnostic performance of the model using the test set and evaluate the precision, recall, and F1 score;

[0027] The model's interpretability is revealed through feature weight heatmaps and t-SNE dimensionality reduction visualization. The contribution of time-frequency and physical features to fault classification is analyzed, and the relationship between fault features and physical mechanisms is clarified, providing guidance for model optimization.

[0028] Hyperparameter tuning is performed through grid search and Bayesian optimization, combined with L2 regularization and Dropout technology to reduce the risk of overfitting, optimize the model's convergence speed and generalization performance, and generate the final diagnostic model.

[0029] The present invention also provides a bearing fault diagnosis device based on time-frequency characteristics and physical information neural network, comprising:

[0030] A data acquisition module that collects and pre-processes vibration signals of the bearing under different operating conditions, including normal conditions and various fault modes;

[0031] The time-frequency feature acquisition module uses continuous wavelet transform to perform time-frequency analysis on the pre-processed vibration signal, extracts time-frequency features to generate a time-frequency spectrum, and performs standardization on the time-frequency spectrum;

[0032] The physical feature acquisition module extracts fault-related physical features based on the bearing's physical model and operating parameters, including fault characteristic frequency, speed, load, temperature, and bearing geometric parameters. Enhanced physical features are obtained through z-score normalization and principal component analysis dimensionality reduction.

[0033] The label acquisition module establishes an association mapping between enhanced physical features and fault modes. By analyzing the corresponding relationship between fault feature frequencies and vibration signals, it constructs a semantic representation of physical features and generates a structured physical feature vector.

[0034] The prediction model construction module constructs a deep neural network that integrates time-frequency features and physical information and trains it using a labeled fault dataset to generate the final diagnostic model. The deep neural network adopts a dual-branch architecture. The first branch receives the standardized time-frequency spectrum and extracts deep time-frequency features based on ResNet18. The second branch receives the physical feature vector and uses Transformer to interact between features to obtain a second physical feature vector with the same dimension as the time-frequency feature. The deep time-frequency features and the second physical feature vector are adaptively adjusted through the ECA attention mechanism, and then enter the feature fusion layer to obtain fused features through weighted splicing and integration. The fully connected layer maps the fused features to fault categories and outputs the classification probability.

[0035] The bearing fault diagnosis module uses the final diagnostic model to achieve accurate classification and status assessment of bearing faults.

[0036] The present invention has the following beneficial effects:

[0037] (1) Fusion of time-frequency and physical features to improve diagnostic accuracy: By extracting high-resolution time-frequency features through continuous wavelet transform (CWT), the dynamic changes of non-stationary vibration signals can be accurately captured. Combined with physical information such as bearing fault characteristic frequency, speed, load, temperature, and geometric parameters, a multi-dimensional feature space is formed. Through feature fusion, a deep integration of data-driven methods and physical mechanisms is achieved, which not only improves the comprehensiveness and accuracy of feature expression, but also enhances the model's ability to identify complex fault modes, providing a solid foundation for high-precision diagnosis.

[0038] (2) Residual network and attention mechanism enhance robustness: The ResNet18 residual network architecture is adopted to effectively alleviate the gradient vanishing problem of deep networks through residual connections. The ECA attention mechanism is combined to adaptively adjust channel weights, significantly enhancing the model's ability to extract key time-frequency features and physical features. This method can adapt to complex working conditions such as variable speed, variable load, and high noise, ensuring diagnostic stability and reliability under diverse working conditions.

[0039] (3) Visual analysis improves model interpretability: Through analytical methods such as feature weight heat maps and t-SNE dimensionality reduction visualization, we deeply explore the relationship between time-frequency features and fault feature frequencies, reveal the physical mechanism of the fault, significantly enhance the interpretability of the model, and provide a reliable diagnostic basis for industrial sites;

[0040] (4) Systematic framework to ensure generalization capability: A systematic diagnostic framework is constructed through a multi-step feature processing, model training, interpretability analysis and hyperparameter optimization process; this framework ensures the high accuracy, robustness and generalization capability of the model in various industrial scenarios, providing comprehensive support for the accurate classification and status assessment of bearing faults.

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A diagram showing the steps of a method according to an embodiment of the present invention;

[0043] Figure 2 Schematic diagram of a deep network model according to an embodiment of the present invention;

[0044] Figure 3 Schematic diagram of a device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] See also Figure 1 FIG. 1 is a diagram showing steps of a method according to an embodiment of the present invention, comprising the following steps:

[0046] S101, collecting and preprocessing vibration signals of the bearing under different operating conditions, wherein the operating conditions include a normal state and various fault modes;

[0047] S102, performing time-frequency analysis on the preprocessed vibration signal using continuous wavelet transform, extracting time-frequency features to generate a time-frequency spectrum, and performing standardization on the time-frequency spectrum;

[0048] S103, extracting fault-related physical features based on the bearing's physical model and operating parameters, including fault characteristic frequency, speed, load, temperature, and bearing geometric parameters. Enhanced physical features are obtained through z-score normalization and principal component analysis dimensionality reduction.

[0049] S104, establishing an association mapping between enhanced physical features and fault modes, constructing a semantic representation of the physical features by analyzing the corresponding relationship between the fault feature frequency and the vibration signal, and generating a structured physical feature vector;

[0050] S105, constructing a deep neural network that integrates time-frequency features and physical information and training it using the labeled fault dataset to generate the final diagnosis model;

[0051] S106, using the final diagnostic model to achieve accurate classification and status assessment of bearing faults.

[0052] Among them, the deep neural network adopts a dual-branch architecture. The first branch receives the standardized time-frequency spectrum and extracts the time-frequency deep features based on ResNet18; the second branch receives the physical feature vector and uses Transformer to interact between features to obtain a second physical feature vector consistent with the dimension of the time-frequency feature; the time-frequency deep features and the second physical feature vector are respectively adaptively adjusted through the ECA attention mechanism. After the channel weights are adjusted, they enter the feature fusion layer and obtain the fused features through weighted splicing and integration; the fully connected layer maps the fused features to fault categories and outputs the classification probability.

[0053] Specifically, the experimental data of the embodiments of the present invention are obtained from a bearing test bench at Case Western Reserve University, which is composed of a motor, a torque sensor, and a dynamometer. The present invention is illustrated using a drive-end SKF 6205 rolling bearing as an example.

[0054] Specifically, in S101, speed, load, temperature, and bearing geometry parameters are first extracted. Then, based on the speed and bearing geometry parameters, the fault characteristic frequency is calculated using a physical model. Finally, all physical characteristics (including the fault characteristic frequency, speed, load, temperature, and geometry parameters) are z-score normalized and PCA dimension reduction is performed to obtain enhanced physical characteristics. In this example, an accelerometer is used to collect vibration signals from an SKF 6205 bearing under various operating conditions. The SKF 6205 bearing is a deep groove ball bearing with an inner diameter of 25 mm, an outer diameter of 52 mm, and a width of 15 mm. Single-point faults are introduced using electrospark machining (EDM) technology on the bearing inner race, outer race, and rolling element, with diameters of 0.007, 0.014, and 0.021 inches, respectively. The collected data includes four states: normal, inner race fault, outer race fault, and rolling element fault. The sampling frequency is 12 kHz, and the motor speed is 1772 rpm. The vibration signal is preprocessed, including denoising, signal segmentation, and data enhancement, to improve data quality and diversity. Each sample is 1024 data points long. The dataset is partitioned using a sliding window with a window length of 1024 samples and a sliding step size of 512 points, ensuring a 50% overlap. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. It contains 10 fault categories, with 210 samples per category, for a total of 2100 samples. The dataset is divided into four groups, designated A, B, C, and D, based on loads of 0HP, 1HP, 2HP, and 3HP, respectively.

[0055] Specifically, in S102, a continuous wavelet transform is used to perform time-frequency analysis on the vibration signal. The Morlet mother wavelet is used with a scale range of 1 to 128 to generate a 128×128 pixel time-frequency spectrum. The center frequency of the Morlet wavelet is set to 5, and the sampling frequency is 12kHz. After the time-frequency spectrum is generated, z-score normalization is performed to optimize the distribution of neural network input data. During verification, fault signals such as inner race faults show periodic impact energy concentration at the characteristic frequency, proving the effectiveness of the continuous wavelet transform. The formula of the continuous wavelet transform (CWT) is defined as follows:

[0056]

[0057] Where x(t) is the input vibration signal, ψ(t) is the mother wavelet function (the present invention uses Morlet wavelet), represents its complex conjugate, a is the scale parameter, b is the translation parameter, X w (a, b) are the CWT coefficients, which represent the time-frequency characteristics of the signal at different scales and time positions. The CWT generates high-resolution time-frequency spectrograms by continuously varying the scale and translation parameters, making it suitable for analyzing nonstationary vibration signals.

[0058] Specifically, in S103, the bearing fault characteristic frequency is the core of the physical information, reflecting the periodic vibration characteristics of the bearing fault, including the inner race fault frequency (Ball Pass Frequency Inner, BPFI), the outer race fault frequency (Ball Pass Frequency Outer, BPFO), the rolling element fault frequency (Ball Spin Frequency, BSF), and the cage fault frequency (Fundamental Train Frequency, FTF). The main formula is as follows:

[0059]

[0060] Among them, f r is the bearing speed (in revolutions per minute), n is the number of rolling elements, d is the rolling element diameter, D is the bearing pitch diameter, and θ is the contact angle. These characteristic fault frequencies are reflected in the periodic impact of the vibration signal and, together with operating parameters such as speed, load, and temperature, constitute physical characteristics that enhance the model's understanding of the fault mechanism.

[0061] Physical characteristics were extracted based on the SKF 6205 bearing parameters: 9 rolling elements, 7.94 mm rolling element diameter, 39.04 mm pitch circle diameter, and a contact angle of 0 degrees. At a speed of 1772 rpm, the characteristic fault frequencies were calculated: the inner ring fault frequency was approximately 162.1 Hz, the outer ring fault frequency was approximately 104.9 Hz, the rolling element fault frequency was approximately 68.7 Hz, and the cage fault frequency was approximately 11.7 Hz. Operating parameters collected included speed and load, ranging from 0 hp to 3 hp. The five-dimensional features were normalized by z-score and reduced in dimension using principal component analysis to generate a five-dimensional physical feature vector. This vector includes the fault characteristic frequency, which is the typical fault frequency calculated based on the structural parameters and operating speed of the SKF 6205 bearing; the speed, which is the actual bearing operating speed (unit: RPM); the load, which is the external load borne by the bearing (ranging from 0 to 3HP); the temperature, which is the temperature data during the bearing operation; and the bearing geometric parameters, which refer to static structural information such as the number of rolling elements, rolling element diameter, and pitch circle diameter. These are combined and encoded into a one-dimensional value to express the differences in bearing structural characteristics.

[0062] Specifically, in S104, the correspondence between physical features and fault modes is analyzed to construct a semantic mapping model. The inner race fault frequency of 162.1 Hz correlates with high-energy regions in the time-frequency spectrum, indicating that the inner race damage manifests as periodic impacts. A five-dimensional physical feature vector is generated, and ten fault categories are mapped to label vectors using one-hot encoding. This supports subsequent feature fusion and serves as a supervisory signal for network training. The fused time-frequency and physical features are mapped to classification probabilities, and the error is calculated using a loss function compared with the one-hot labels, which in turn guides the network to adjust fusion weights and feature representation. This label vector provides a clear category semantic standard for feature fusion, promoting the effective integration of multimodal features and accurate classification of fault categories, thereby improving the performance of the diagnostic model. The ten fault categories described in this embodiment include one normal state and nine typical bearing fault states with different location / fault severity combinations, specifically including normal state, 7mm / 14mm / 21mm inner race faults, 7mm / 14mm / 21mm outer race faults, and 7mm / 14mm / 21mm rolling element faults.

[0063] Specifically, in S105, a two-branch deep neural network is constructed. One branch extracts deep features from the time-frequency spectrum based on ResNet18, inputs a standardized time-frequency spectrum (size 3×224×224), and generates a 512-dimensional feature vector through convolution, batch normalization, and residual connections to capture the dynamic pattern of the vibration signal. The other branch uses Transformer to process the physical feature vector, and models the complex interactions between features through a two-layer encoder (including 8 self-attention and feedforward networks), outputting a 512-dimensional feature vector. Subsequently, the features of the two branches are adaptively adjusted in channel weights through the Efficient Channel Attention (ECA) attention mechanism to highlight key features related to the fault. The feature fusion layer integrates the 512-dimensional features of the two branches through weighted concatenation to form a 1024-dimensional fused feature, which is then compressed to 512 dimensions through a fully connected layer. Finally, the fully connected layer maps the fused feature to fault categories, which are then mapped to 10-dimensional fault category probabilities through a softmax layer, enabling accurate diagnosis of bearing faults under complex working conditions. In the test, the classification accuracy of 50 samples reached 96%.

[0064] ResNet18 is a deep residual network with 18 convolutional layers. It uses residual connections to alleviate the vanishing gradient problem of deep networks. The core formula of its residual block is:

[0065] y=F(x,W i )+x;

[0066] Among them, x is the input feature, F(x,W i ) is the residual function, and y is the output feature. If the input and output dimensions are different, use 1×1 convolution to adjust the x dimension:

[0067] y=F(x,W i )+W s x;

[0068] Among them, W s The convolution weights are dimensionally adjusted. The ResNet18 architecture includes an initial 7×7 convolutional layer with 64 channels and a stride of 2; four groups of residual blocks, each with two 3×3 convolutional layers, with 64, 128, 256, and 512 channels, respectively; global average pooling, and a fully connected layer. In this paper, ResNet18 is used to extract deep features from time-frequency spectrograms. Combining the ECA mechanism with the MLP module achieves a deep fusion of time-frequency and physical features.

[0069] In step S5, the ECA (Efficient Channel Attention) attention mechanism optimizes feature extraction by channel attention, and its core formula is as follows:

[0070]

[0071] Among them, z c is the description of the cth channel, x c (i, j) is the value of the c-th channel feature map at position (i, j), H and W are the height and width of the feature map, Conv1D is a one-dimensional convolution, the convolution kernel size k is adaptively selected, σ is the Sigmoid activation function, w c is the channel weight, is the weighted feature map. The ECA mechanism enhances the expression of fault-related features through low-complexity channel attention.

[0072] Specifically, in S106, 2100 samples were used for training: a training set of 1470 samples, a validation set of 420 samples, and a test set of 210 samples. The optimizer was Adam, with a learning rate of 0.001, a batch size of 32, and L2 regularization and dropout. After 100 training iterations, the test set accuracy reached 98.5%, with an F1 score of 0.98, verifying the model's practicality. Model interpretability was analyzed using feature weight heatmaps and t-SNE dimensionality reduction visualization. The high-weight region for inner-circuit fault samples was concentrated at 162Hz, consistent with theoretical values. t-SNE visualization showed clear separation of fault categories, with a correlation weight of 0.85 between rotational speed and fault frequency. Grid search was used to optimize hyperparameters, resulting in a learning rate of 0.001, a batch size of 32, a dropout rate of 0.3, and an L2 regularization coefficient of 0.01. Data augmentation and early stopping strategies were introduced.

[0073] See also Figure 3 FIG. 1 is a structural diagram of a device according to an embodiment of the present invention, comprising:

[0074] A data acquisition module 301 collects and pre-processes vibration signals of the bearing under different operating conditions, including normal conditions and various fault modes;

[0075] The time-frequency feature acquisition module 302 performs time-frequency analysis on the pre-processed vibration signal using continuous wavelet transform, extracts time-frequency features to generate a time-frequency spectrum, and performs standardization on the time-frequency spectrum;

[0076] The physical feature acquisition module 303 extracts fault-related physical features based on the physical model and operating parameters of the bearing, including fault feature frequency, speed, load, temperature, and bearing geometric parameters. Enhanced physical features are obtained through z-score normalization and principal component analysis dimensionality reduction.

[0077] The label acquisition module 304 establishes an association mapping between enhanced physical features and fault modes, constructs a semantic representation of the physical features by analyzing the corresponding relationship between the fault feature frequency and the vibration signal, and generates a structured physical feature vector;

[0078] Prediction model construction module 305 constructs a deep neural network that integrates time-frequency features and physical information and trains it using the labeled fault data set to generate a final diagnostic model. The deep neural network adopts a dual-branch architecture. The first branch receives the standardized time-frequency spectrum and extracts deep time-frequency features based on ResNet18. The second branch receives the physical feature vector and uses Transformer to interact between features to obtain a second physical feature vector consistent with the dimension of the time-frequency feature. The deep time-frequency features and the second physical feature vector are adaptively adjusted in channel weights through the ECA attention mechanism, and then enter the feature fusion layer to obtain fused features through weighted splicing and integration. The fully connected layer maps the fused features to fault categories and outputs classification probabilities.

[0079] The bearing fault diagnosis module 306 uses the final diagnosis model to achieve accurate classification and status assessment of bearing faults.

[0080] The present invention achieves high-precision, robust and interpretable bearing fault diagnosis by fusing time-frequency features with physical information, and is suitable for complex industrial scenarios.

[0081] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A bearing fault diagnosis method based on time-frequency characteristics and physical information neural network, characterized in that: The steps include: Collect and pre-process bearing vibration signals under different operating conditions, including normal conditions and various failure modes; The pre-processed vibration signal is subjected to time-frequency analysis using continuous wavelet transform, and the time-frequency features are extracted to generate a time-frequency spectrum, which is then standardized. Based on the bearing's physical model and operating parameters, fault-related physical features are extracted, including fault characteristic frequency, speed, load, temperature, and bearing geometric parameters. Enhanced physical features are obtained through z-score normalization and principal component analysis dimensionality reduction. Establish an association mapping between enhanced physical features and fault modes, construct a semantic representation of physical features by analyzing the correspondence between fault feature frequencies and vibration signals, and generate a structured physical feature vector. A deep neural network that integrates time-frequency features and physical information is constructed and trained using a labeled fault dataset to generate the final diagnostic model. The deep neural network adopts a two-branch architecture. The first branch receives the standardized time-frequency spectrum and extracts deep time-frequency features based on ResNet18. The second branch receives the physical feature vector and uses the Transformer to interact between features to obtain a second physical feature vector with the same dimension as the time-frequency feature. The time-frequency deep features and the second physical feature vector are adaptively weighted using the ECA attention mechanism, and then enter the feature fusion layer to obtain fused features through weighted splicing and integration. The fully connected layer maps the fused features to fault categories and outputs the classification probability. The final diagnostic model is used to achieve accurate classification and condition assessment of bearing faults.

2. The bearing fault diagnosis method based on time-frequency characteristics and physical information neural network according to claim 1 is characterized in that: The multiple failure modes include: inner ring damage, outer ring damage, rolling element damage and cage damage.

3. The bearing fault diagnosis method based on time-frequency characteristics and physical information neural network according to claim 1 is characterized in that: The preprocessing includes denoising, signal segmentation and data enhancement.

4. The bearing fault diagnosis method based on time-frequency characteristics and physical information neural network according to claim 1 is characterized in that: The continuous wavelet transform is used to perform time-frequency analysis on the preprocessed vibration signal, and the time-frequency features are extracted to generate a time-frequency spectrum. Specifically, the Morlet wavelet is used to obtain the time-frequency features, and then a high-resolution time-frequency spectrum is generated by continuously changing the scale and translation parameters.

5. The bearing fault diagnosis method based on time-frequency characteristics and physical information neural network according to claim 1 is characterized in that: The bearing geometric parameters include rolling element diameter, pitch circle diameter, and contact angle.

6. The bearing fault diagnosis method based on time-frequency characteristics and physical information neural network according to claim 1 is characterized in that: The bearing fault characteristic frequencies include the inner ring fault frequency BPFI, the outer ring fault frequency BPFO, the rolling element fault frequency BSF and the cage fault frequency FTF, which are respectively expressed as: Among them, f r is the bearing speed (unit: rpm), n is the number of rolling elements, d is the rolling element diameter, D is the bearing pitch circle diameter, and θ is the contact angle.

7. The bearing fault diagnosis method based on time-frequency characteristics and physical information neural network according to claim 1 is characterized in that: The Transformer models the complex interactions between features through a two-layer encoder, each of which includes an 8-head self-attention and feedforward network.

8. The bearing fault diagnosis method based on time-frequency characteristics and physical information neural network according to claim 1 is characterized in that: The ECA attention mechanism is expressed as: In c =σ(Conv1D(z,k)); Among them, z c is the description of the cth channel, x c (i, j) is the value of the c-th channel feature map at position (i, j), H and W are the height and width of the feature map, Conv1D is a one-dimensional convolution, the convolution kernel size k is adaptively selected, σ is the Sigmoid activation function, w c is the channel weight, is the weighted feature map. The ECA mechanism enhances the expression of fault-related features through low-complexity channel attention.

9. The bearing fault diagnosis method based on time-frequency characteristics and physical information neural network according to claim 1 is characterized in that: The deep neural network integrating time-frequency features and physical information is constructed and trained using the labeled fault data set to generate the final diagnosis model. The training includes the following steps: The cross-entropy loss function is used to optimize the model, and the network weights are initialized through the pre-trained model in combination with the transfer learning strategy to enhance the generalization ability in small sample scenarios; Verify the diagnostic performance of the model using the test set and evaluate the precision, recall, and F1 score; The model's interpretability is revealed through feature weight heatmaps and t-SNE dimensionality reduction visualization. The contribution of time-frequency and physical features to fault classification is analyzed, and the relationship between fault features and physical mechanisms is clarified, providing guidance for model optimization. Hyperparameter tuning is performed through grid search and Bayesian optimization, combined with L2 regularization and Dropout technology to reduce the risk of overfitting, optimize the model's convergence speed and generalization performance, and generate the final diagnostic model.

10. A bearing fault diagnosis device based on time-frequency characteristics and physical information neural network, characterized in that: include: A data acquisition module that collects and pre-processes vibration signals of the bearing under different operating conditions, including normal conditions and various fault modes; The time-frequency feature acquisition module uses continuous wavelet transform to perform time-frequency analysis on the pre-processed vibration signal, extracts time-frequency features to generate a time-frequency spectrum, and performs standardization on the time-frequency spectrum; The physical feature acquisition module extracts fault-related physical features based on the bearing's physical model and operating parameters, including fault characteristic frequency, speed, load, temperature, and bearing geometric parameters. Enhanced physical features are obtained through z-score normalization and principal component analysis dimensionality reduction. The label acquisition module establishes an association mapping between enhanced physical features and fault modes. By analyzing the corresponding relationship between fault feature frequencies and vibration signals, it constructs a semantic representation of physical features and generates a structured physical feature vector. The prediction model construction module constructs a deep neural network that integrates time-frequency features and physical information and trains it using a labeled fault dataset to generate the final diagnostic model. The deep neural network adopts a two-branch architecture. The first branch receives the standardized time-frequency spectrum and extracts deep time-frequency features based on ResNet18. The second branch receives the physical feature vector and uses the Transformer to interact between features to obtain a second physical feature vector with the same dimension as the time-frequency feature. The time-frequency deep features and the second physical feature vector are adaptively weighted using the ECA attention mechanism, and then enter the feature fusion layer to obtain fused features through weighted splicing and integration. The fully connected layer maps the fused features to fault categories and outputs the classification probability. The bearing fault diagnosis module uses the final diagnostic model to achieve accurate classification and status assessment of bearing faults.

Images