A bearing fault diagnosis method based on multi-source time-frequency feature fusion

Abstract

The application discloses a bearing fault diagnosis method based on multi-source time-frequency feature fusion, and comprises the following steps: acquiring a bearing vibration signal and a rotating speed signal; constructing a bearing fault identification model based on a multi-source time-frequency feature extraction and fusion framework, which is used for identifying the bearing fault; training the bearing fault identification model to obtain a trained bearing fault identification model; and inputting the bearing vibration signal and the rotating speed signal into the trained bearing fault identification model to realize bearing fault classification. Through the combination of multi-source information fusion and an attention mechanism, the application improves feature expression capability and fault identification precision under the premise of ensuring model calculation efficiency, and provides an effective implementation method for bearing fault diagnosis under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of signal and information processing technology, and relates to a bearing fault diagnosis method based on multi-source time-frequency feature fusion. Background Technology

[0002] Currently, bearing fault diagnosis methods mainly include traditional signal processing-based methods and data-driven intelligent diagnostic methods. Traditional methods typically rely on Fourier transform, wavelet analysis, and other techniques to process vibration signals and manually extract features to achieve fault identification. While these methods have some physical meaning, their ability to express signal features under complex operating conditions is limited, and they are highly dependent on professional experience. With the development of deep learning, intelligent diagnostic methods based on models such as convolutional neural networks have gradually become a research hotspot. These methods can automatically extract features from raw signals and achieve end-to-end fault identification, improving diagnostic accuracy to some extent. However, most existing methods rely on a single signal source, making it difficult to fully utilize the complementary features contained in multi-source information. Regarding the utilization of multi-source information, although some studies have attempted to jointly analyze vibration signals and rotational speed signals, the fusion effect still needs improvement due to scale differences and information redundancy between different signals. Furthermore, existing deep learning models mainly rely on local information during feature extraction, failing to adequately utilize the correlations between different features, thus limiting feature representation capabilities. Furthermore, to enhance feature representation capabilities, some studies have introduced attention mechanisms for weighted feature processing. However, while these methods improve model performance, they often come at the cost of increased computational complexity and more complex model structures, hindering practical engineering applications. Therefore, existing technologies still have shortcomings in terms of effective fusion of multi-source information, enhanced feature representation capabilities, and improved model efficiency, and require further improvement.

[0003] Rotating machinery is widely used in industrial production, and bearings, as crucial components, directly affect the safety and stability of the equipment. During long-term operation, bearings are susceptible to complex operating conditions, such as load variations, speed fluctuations, and environmental interference, increasing the probability of failure. Once a bearing malfunctions, it not only affects the normal operation of the equipment but may also trigger serious safety accidents, causing significant economic losses. With the continuous development of industrial field monitoring methods, various operating data, such as vibration and speed signals, can be collected in real time. These signals reflect the equipment's operating status from different perspectives. However, in practical applications, due to complex operating conditions, significant differences in signal characteristics, and numerous external interferences, the acquired data suffers from problems such as information dispersion and inconsistent representation, thus affecting the accurate judgment of the equipment's operating status. Against this backdrop, how to fully utilize multi-source monitoring information under complex operating conditions to improve the perception of bearing operating status and achieve accurate fault identification and timely early warning is of great significance for ensuring safe equipment operation, improving production efficiency, and reducing maintenance costs. Summary of the Invention

[0004] To address the aforementioned problems, the technical solution adopted by this invention is: a bearing fault diagnosis method based on multi-source time-frequency feature fusion, comprising the following steps:

[0005] Acquire the vibration and rotational speed signals of the bearing; Based on a multi-source time-frequency feature extraction and fusion framework, a bearing fault identification model is constructed to identify bearing faults. The bearing fault identification model is trained to obtain a trained bearing fault identification model; The vibration and rotational speed signals of the bearing are input into a trained bearing fault identification model to classify bearing faults.

[0006] Furthermore: the bearing fault identification model includes: Time-frequency convolution module: used to process the input vibration signal and rotation speed signal to extract the physically meaningful time-frequency features of different signal sources; Fusion module: Based on the physically meaningful time-frequency features of different signal sources output by the time-frequency convolution module, the features are fused along the channel dimension to construct a unified multi-source feature representation. High-efficiency channel attention module: Based on the fused features output by the fusion module, a high-efficiency channel attention mechanism is introduced to adaptively weight the fused features. By modeling the local interaction relationships between channels, the expressive power of key features is enhanced, resulting in attention-enhanced features. Fault classification module: Based on the attention-enhanced features output by the fusion module, it gradually extracts high-level discriminative features through multi-layer convolution and pooling operations, and completes the identification of fault categories, thereby achieving classification of various operating conditions of bearings.

[0007] Furthermore, the process of processing the input vibration signal and rotation speed signal to extract physically meaningful time-frequency features from different signal sources is as follows: Vibration signals With speed signal Based on the real and imaginary part convolution kernels, the input signal is convolved to obtain the real and imaginary responses, respectively. Then, the final time-frequency representation is output through modulus operation. The calculation process is as follows: (1) in, For convolution kernel function, and These represent the real part convolution kernel and the imaginary part convolution kernel, respectively. These are trainable parameters used to control the frequency characteristics of the convolution kernel function; (2) in, For convolution operators, The channel number represents the convolution kernel. For input signal, and They represent the first The real and imaginary outputs of each channel. For the first The final amplitude of each channel; The kernel function is constructed using STTF, and the calculation formula is as follows: (3) in, For window width parameters, The number of channels in the TFconv layer. This is the sample index, and its value range is... , It is the frequency factor.

[0008] Furthermore: Based on the fused features output by the fusion module, an efficient channel attention mechanism is introduced to adaptively weight the fused features. By modeling the local interaction relationships between channels, the expressive power of key features is enhanced, resulting in attention-enhanced features. The vibration signal and rotational speed signal are obtained after passing through a time-frequency convolution module. and The two feature streams are concatenated along the channel dimension to form a 64-channel joint feature representation. ; Global average pooling is performed on the concatenated time-frequency features to obtain the global description vector for each channel. ; (4) in, Indicates the number of frequency bands; Then, one-dimensional convolution with an adaptive kernel size is used to capture local cross-channel dependencies, with the kernel size... Determined by the following formula: (5) in, Indicates and The closest odd number, and It is a constant; Subsequently, channel attention weights are generated using the Sigmoid function. The calculation formula is as follows: (6) in, This represents the Sigmoid function. Indicates the kernel size as One-dimensional convolution operation; Finally, the weights are multiplied channel by channel of the original time-frequency representation to enhance the information channels and suppress redundant information, resulting in a refined time-frequency representation.

[0009] Furthermore: the fault classification module includes: Convolutional feature extraction submodule: used to capture transient impact features of bearing failures using large convolutional kernels, and to progressively expand the channel dimension using small convolutional kernels; Classification output submodule: Based on the high-level feature representation output by the convolutional feature extraction submodule, it performs feature compression and mapping, and maps it to the category space through a fully connected layer to achieve classification and discrimination of different fault types.

[0010] Furthermore: the convolutional feature extraction submodule includes a first convolutional layer, a second convolutional layer, a third convolutional layer, and a fourth convolutional layer connected in sequence; The first convolutional layer has 64 input channels, 16 output channels, and 15 convolutional kernels; The second convolutional layer has 32 output channels, 3 convolutional kernels, and is followed by max pooling with a stride of 2. The third convolutional layer has 64 output channels and 3 convolutional kernels. The fourth convolutional layer has 128 output channels and 3 convolutional kernels; Then, adaptive max pooling is used to compress the features into a fixed length, which is finally mapped to a 5-dimensional output space, corresponding to five bearing operating states.

[0011] This invention provides a bearing fault diagnosis method based on multi-source time-frequency feature fusion. This method comprehensively considers the differences and complementarities between vibration signals and speed signals in representing equipment operating states. Through joint modeling of multi-source information, it effectively improves the expressive ability of bearing fault features. By constructing a time-frequency convolutional feature extraction module, joint modeling of signal time-domain and frequency-domain information is achieved. This ensures that the extracted features not only contain rich spectral information but also possess a certain degree of physical interpretability, thereby improving the model's ability to identify complex fault modes. Furthermore, an efficient channel attention mechanism is introduced. By adaptively weighting channel features, it highlights key features crucial for fault diagnosis while suppressing redundant and interfering information. Compared with traditional attention methods, this mechanism reduces the number of model parameters and computational complexity while avoiding information compression, thus improving performance while maintaining computational efficiency. In addition, the constructed multi-layer convolutional classification structure can abstract and discriminate the fused features layer by layer, achieving accurate classification of various bearing operating states. Through the synergistic effect of multi-source information fusion and the attention mechanism, this invention maintains high diagnostic accuracy and stability even under complex operating conditions and speed variations. Experimental results show that the present invention can achieve better diagnostic results in bearing fault identification tasks, effectively improve the accuracy and reliability of fault identification, and verify the feasibility and practical value of the method in actual engineering applications. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of the method; Figure 2 It is a time-frequency convolution module; Figure 3 The results are t-sne results trained using only vibration data; Figure 4 The results are t-sne results trained using only rotational speed data; Figure 5 t-sne results trained with vibration and rotational speed data. Detailed Implementation

[0014] It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Figure 1 This is a flowchart of the method; A bearing fault diagnosis method based on multi-source time-frequency feature fusion includes the following steps: Acquire the vibration and rotational speed signals of the bearing; Based on a multi-source time-frequency feature extraction and fusion framework, a bearing fault identification model is constructed to identify bearing faults. The bearing fault identification model is trained to obtain a trained bearing fault identification model; The vibration and rotational speed signals of the bearing are input into a trained bearing fault identification model to classify bearing faults.

[0017] Furthermore: the bearing fault identification model includes: Time-frequency convolution module: used to process the input vibration signal and rotation speed signal to extract the physically meaningful time-frequency features of different signal sources; Figure 2 It is a time-frequency convolution module; Fusion module: Based on the physically meaningful time-frequency features of different signal sources output by the time-frequency convolution module, the features are fused along the channel dimension to construct a unified multi-source feature representation. High-efficiency channel attention module: Based on the fused features output by the fusion module, a high-efficiency channel attention mechanism is introduced to adaptively weight the fused features. By modeling the local interaction relationships between channels, the expressive power of key features is enhanced, while suppressing redundant information, resulting in attention-enhanced features. Fault classification module: Based on the attention-enhanced features output by the fusion module, it gradually extracts high-level discriminative features through multi-layer convolution and pooling operations, and completes the identification of fault categories, thereby achieving classification of various operating conditions of bearings.

[0018] The bearing's various operating conditions include classifications such as normal condition, rolling element failure, inner ring failure, outer ring failure, and combined failure. Furthermore, the process of processing the input vibration signal and rotation speed signal to extract physically meaningful time-frequency features from different signal sources is as follows: Vibration signals With speed signal Based on the real and imaginary part convolution kernels, the input signal is convolved to obtain the real and imaginary responses, respectively. Then, the final time-frequency representation is output through modulus operation. The calculation process is as follows: (1) in, For convolution kernel function, and These represent the real part convolution kernel and the imaginary part convolution kernel, respectively. These are trainable parameters used to control the frequency characteristics of the convolution kernel function; (2) in, For convolution operators, The channel number represents the convolution kernel. For input signal, and They represent the first The real and imaginary outputs of each channel. For the first The final amplitude of each channel; The kernel function is constructed using STTF, and the calculation formula is as follows: (3) in, For window width parameters, The number of channels in the TFconv layer. This is the sample index, and its value range is... , It is the frequency factor.

[0019] Furthermore: Based on the fused features output by the fusion module, an efficient channel attention mechanism is introduced to adaptively weight the fused features. By modeling the local interaction relationships between channels, the expressive power of key features is enhanced, resulting in attention-enhanced features. The vibration signal and rotational speed signal are obtained after passing through a time-frequency convolution module. and The two feature streams are concatenated along the channel dimension to form a 64-channel joint feature representation. ; Global average pooling is performed on the concatenated time-frequency features to obtain the global description vector for each channel. ; (4) in, Indicates the number of frequency bands; Then, one-dimensional convolution with an adaptive kernel size is used to capture local cross-channel dependencies, with the kernel size... Determined by the following formula: (5) in, Indicates and The closest odd number, and It is a constant; Subsequently, channel attention weights are generated using the Sigmoid function. The calculation formula is as follows: (6) in, This represents the Sigmoid function. Indicates the kernel size as One-dimensional convolution operation; Finally, the weights are multiplied channel by channel of the original time-frequency representation to enhance the information channels and suppress redundant information, resulting in a refined time-frequency representation.

[0020] Furthermore: the fault classification module includes: Convolutional feature extraction submodule: By using large convolution to capture the transient impact features of bearing failure, and then gradually expanding the channel dimension through small convolution kernel (3), the computational efficiency is maintained while improving the feature expression capability.

[0021] Classification Output Submodule: The fifth layer is used to perform feature compression and mapping based on the high-level feature representation output by the convolutional feature extraction submodule, and then map it to the category space through a fully connected layer to achieve classification and discrimination of different fault types.

[0022] Furthermore: the convolutional feature extraction submodule includes a first convolutional layer, a second convolutional layer, a third convolutional layer, and a fourth convolutional layer connected in sequence; The first convolutional layer has 64 input channels, 16 output channels, and 15 convolutional kernels; The second convolutional layer has 32 output channels, 3 convolutional kernels, and is followed by max pooling with a stride of 2. The third convolutional layer has 64 output channels and 3 convolutional kernels. The fourth convolutional layer has 128 output channels and 3 convolutional kernels; Then, adaptive max pooling is used to compress the features into a fixed length; after flattening, they are passed through fully connected layers from 512-dimensional to 256-dimensional and from 25-dimensional to 64-dimensional, and finally mapped to a 5-dimensional output space, corresponding to five bearing operating states.

[0023] The first convolutional layer uses a large convolutional kernel to capture the transient impact features of bearing failure, while the second to fourth layers gradually expand the channel dimension through small convolutional kernels, improving feature representation while maintaining computational efficiency. The first convolutional layer is constructed using multiple one-dimensional convolutions, batch normalization, and ReLU activation layers. Max pooling is introduced after the second layer to gradually reduce the feature dimension and expand the receptive field. After deep feature extraction, adaptive max pooling is used to unify the output dimension to ensure consistency of feature representation. Finally, the pooled features are flattened and passed through multiple fully connected layers before the output layer determines the fault category.

[0024] In the training process of the bearing fault identification model, the cross-entropy loss function was used as the classification loss function to measure the difference between the model's predicted class distribution and the true label distribution. The optimizer was Adam, with an initial learning rate of 0.001, a batch size of 64, and a total of 50 training epochs. To ensure the reproducibility of the experimental results, the random seed was fixed at 999.

[0025] The method of this invention was implemented in the Ubuntu 20.04.4 LTS operating system environment, using Python 3.10.12 and the PyTorch framework for model building and training. Experimental results show that this method can effectively fuse complementary information from vibration and rotational speed signals, and achieve high fault identification accuracy while ensuring a lightweight model.

[0026] Figure 3 The results are t-sne results trained using only vibration data; Figure 4 The results are t-sne results trained using only rotational speed data; Figure 5 t-sne results trained with vibration and rotational speed data.

[0027] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A bearing fault diagnosis method based on multi-source time-frequency feature fusion, characterized in that: Includes the following steps Acquire the vibration and rotational speed signals of the bearing; Based on a multi-source time-frequency feature extraction and fusion framework, a bearing fault identification model is constructed to identify bearing faults. The bearing fault identification model is trained to obtain a trained bearing fault identification model; The vibration and rotational speed signals of the bearing are input into a trained bearing fault identification model to classify bearing faults.

2. The bearing fault diagnosis method based on multi-source time-frequency feature fusion according to claim 1, characterized in that: The bearing fault identification model includes: Time-frequency convolution module: used to process the input vibration signal and rotation speed signal to extract the physically meaningful time-frequency features of different signal sources; Fusion module: Based on the physically meaningful time-frequency features of different signal sources output by the time-frequency convolution module, the features are fused along the channel dimension to construct a unified multi-source feature representation. High-efficiency channel attention module: Based on the fused features output by the fusion module, a high-efficiency channel attention mechanism is introduced to adaptively weight the fused features. By modeling the local interaction relationships between channels, the expressive power of key features is enhanced, resulting in attention-enhanced features. Fault classification module: Based on the attention-enhanced features output by the fusion module, it gradually extracts high-level discriminative features through multi-layer convolution and pooling operations, and completes the identification of fault categories, thereby achieving classification of various operating conditions of bearings.

3. The bearing fault diagnosis method based on multi-source time-frequency feature fusion according to claim 1, characterized in that... The process of processing the input vibration and rotation speed signals to extract physically meaningful time-frequency features from different signal sources is as follows: Vibration signals With speed signal Based on the real and imaginary part convolution kernels, the input signal is convolved to obtain the real and imaginary responses, respectively. Then, the final time-frequency representation is output through modulus operation. The calculation process is as follows: (1) in, For convolution kernel function, and These represent the real part convolution kernel and the imaginary part convolution kernel, respectively. These are trainable parameters used to control the frequency characteristics of the convolution kernel function; （2） in, For convolution operators, The channel number represents the convolution kernel. For input signal, and They represent the first The real and imaginary outputs of each channel. For the first The final amplitude of each channel; The kernel function is constructed using STTF, and the calculation formula is as follows: （3） in, For window width parameters, The number of channels in the TFconv layer. This is the sample index, and its value range is... , It is the frequency factor.

4. The bearing fault diagnosis method based on multi-source time-frequency feature fusion according to claim 1, characterized in that... Based on the fused features output by the fusion module, an efficient channel attention mechanism is introduced to adaptively weight the fused features. By modeling the local interaction relationships between channels, the expressive power of key features is enhanced, resulting in attention-enhanced features. The vibration signal and rotational speed signal are obtained after passing through a time-frequency convolution module. and The two feature streams are concatenated along the channel dimension to form a 64-channel joint feature representation. ; Global average pooling is performed on the concatenated time-frequency features to obtain the global description vector for each channel. ; (4) in, Indicates the number of frequency bands; Then, one-dimensional convolution with an adaptive kernel size is used to capture local cross-channel dependencies, with the kernel size... Determined by the following formula: (5) in, Indicates and The closest odd number, and It is a constant; Subsequently, channel attention weights are generated using the Sigmoid function. The calculation formula is as follows: (6) in, This represents the Sigmoid function. Indicates the kernel size as One-dimensional convolution operation; Finally, the weights are multiplied channel by channel of the original time-frequency representation to enhance the information channels and suppress redundant information, resulting in a refined time-frequency representation.

5. The bearing fault diagnosis method based on multi-source time-frequency feature fusion according to claim 1, characterized in that... The fault classification module includes: Convolutional feature extraction submodule: used to capture transient impact features of bearing failures using large convolutional kernels, and to progressively expand the channel dimension using small convolutional kernels; Classification output submodule: Based on the high-level feature representation output by the convolutional feature extraction submodule, it performs feature compression and mapping, and maps it to the category space through a fully connected layer to achieve classification and discrimination of different fault types.

6. The bearing fault diagnosis method based on multi-source time-frequency feature fusion according to claim 1, characterized in that: The convolutional feature extraction submodule includes a first convolutional layer, a second convolutional layer, a third convolutional layer, and a fourth convolutional layer connected in sequence; The first convolutional layer has 64 input channels, 16 output channels, and 15 convolutional kernels; The second convolutional layer has 32 output channels, 3 convolutional kernels, and is followed by max pooling with a stride of 2. The third convolutional layer has 64 output channels and 3 convolutional kernels. The fourth convolutional layer has 128 output channels and 3 convolutional kernels; Then, adaptive max pooling is used to compress the features into a fixed length, which is finally mapped to a 5-dimensional output space, corresponding to five bearing operating states.

Images