Fault diagnosis method for multi-sensor time-frequency fusion equipment under quasi-unbalance condition

Abstract

The invention provides a multi-sensor time-frequency fusion equipment fault diagnosis method under a quasi-unbalance condition, and belongs to the technical field of equipment intelligent fault diagnosis. The method comprises the following steps: based on an original class imbalance sensor data set, constructing an equipment class balance data set through a de-noising diffusion probability model; training an equipment fault identification model through the equipment class balance data set; and receiving multi-source sensor data by using the pre-trained equipment fault identification model, and outputting an equipment fault type. According to the method, the unbalanced data set is expanded to the balanced state by generating the high-quality virtual data, so that the training effect of the equipment fault recognition model is optimized; and a frequency domain feature extraction branch is introduced into the fault identification model, and a multi-scale gating convolution module and an efficient adaptive fusion module are designed, so that adaptive fusion of time-frequency domain features of the multi-source sensor is realized, and the prediction precision of the model is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault diagnosis of equipment, and particularly to a multi-sensor time-frequency fusion equipment fault diagnosis method under the condition of class imbalance. Background Technique

[0002] With the acceleration of the industrial intelligentization process, fault prediction and health management technology plays a core role in ensuring the reliable operation of key equipment. The multi-sensor information fusion technology can effectively overcome the defects of limited signal coverage and weak anti-interference ability of single sensors by integrating monitoring data from different sources (such as vibration, temperature, and acoustic emission signals), and significantly improve the accuracy and robustness of fault diagnosis. In recent years, intelligent diagnosis methods based on deep learning have further promoted the development of this field and become an important research direction for realizing accurate perception of equipment status and early fault warning.

[0003] Currently, multi-sensor information fusion fault diagnosis methods mainly rely on time-domain signal features to construct classification models, and achieve multi-source data fusion through simple feature addition or tensor splicing. Some studies use convolutional neural networks or long short-term memory networks to extract time-series signal features, or introduce attention mechanisms in the fusion stage to optimize feature weights.

[0004] However, the existing methods still face three significant limitations: First, the existing model training uses real datasets, but in actual industrial scenarios, fault samples are scarce and the data distribution is highly unbalanced, resulting in a sharp decline in the recognition performance of the model for minority-class fault types; Second, feature extraction is overly concentrated in the time domain, ignoring the fault features contained in the frequency-domain signals, and is prone to misjudgment in a strong noise environment; Finally, traditional fusion strategies such as splicing lack the exploration of deep associations between features, and redundant parameters increase the computational burden, limiting the applicability of the model in industrial scenarios with high real-time requirements.

[0005] Therefore, a multi-sensor time-frequency fusion equipment fault diagnosis method under the condition of class imbalance is needed. Summary of the Invention

[0006] In view of this, the present invention provides a multi-sensor time-frequency fusion equipment fault diagnosis method under the condition of class imbalance, which constructs a device class-balanced dataset through a denoising diffusion probability model for training a device fault recognition model; the device fault recognition model fuses the time-domain and frequency-domain diversity features of multi-sensor data to improve its diagnosis performance, and solves the problems of insufficient equipment fault recognition accuracy caused by unbalanced training data in industrial practice and the prediction model ignoring frequency-domain features in the prior art.

[0007] To this end, the present invention provides the following technical solutions: A multi-sensor time-frequency fusion equipment fault diagnosis method under the condition of class imbalance, comprising: Based on the original class-imbalanced sensor dataset, construct a device class-balanced dataset through a denoising diffusion probability model; Train a device fault identification model with the device class-balanced dataset; Use the pre-trained device fault identification model to receive multi-source sensor data and output the device fault type.

[0008] Furthermore, the device fault identification model includes: Receive the time-domain data of multiple sensors, and use Fourier transform to convert the time-domain data of each sensor into frequency-domain data; Through a multi-branch backbone network, based on the time-domain data and frequency-domain data of each sensor, extract the dual-domain features of each sensor's data; Use the second-stage efficient adaptive fusion module to fuse the dual-domain features of each sensor to obtain the multi-source features of the device.

[0009] Furthermore, each branch backbone network includes: a time-domain network, a frequency-domain network, and a first-stage efficient adaptive fusion module; The time-domain network extracts time-domain features; The frequency-domain network extracts frequency-domain features; The first-stage efficient adaptive fusion module fuses the time-domain features and frequency-domain features to obtain the dual-domain features of the sensor data.

[0010] Furthermore, the construction of the device class-balanced dataset based on the original class-imbalanced sensor dataset through a denoising diffusion probability model includes: Collect the original class-imbalanced sensor data of the device; Through the denoising diffusion probability model, gradually add Gaussian noise to the original class-imbalanced sensor data, and after a preset number of steps, convert it into pure noise to form a Markov chain, which destroys the original data distribution; Through the neural network to learn the reverse denoising path, reconstruct the original data distribution, and generate sensor data with various balanced distributions.

[0011] Furthermore, the neural network includes: An improved DiffWave network; The improvement to the DiffWave network includes: Based on the DiffWave network, introduce efficient multi-scale attention to enhance local multi-scale feature extraction; And add two linear layers to suppress high-frequency noise.

[0012] Furthermore, the time-domain network includes: Several downsampling modules and multi-scale gated convolution modules; The downsampling module and the multi-scale gated convolution module are arranged alternately; The downsampling module extracts the transformation size and outputs multi-scale features; The multi-scale gated convolution module fuses the multi-scale features and outputs multi-scale fused features.

[0013] Furthermore, the time-domain network has the same structure as the frequency-domain network.

[0014] Furthermore, the first-stage efficient adaptive fusion module includes: Concatenate the time-domain features and the frequency-domain features, and perform a grouping operation to obtain fused features; Perform convolution and pooling on the fused features respectively to obtain local features and global features; Integrate the local features and the global features to obtain intermediate features; Use the intermediate features to determine the corresponding weights of the time-domain features and the frequency-domain features; Based on the corresponding weights, perform weighted summation on the time-domain features and the frequency-domain features to obtain dual-domain features.

[0015] Furthermore, the multi-scale gated convolution module includes: Obtain the corresponding weights of each scale feature through a gating mechanism; Based on the weights corresponding to each scale feature, perform weighted fusion on the multi-scale features to obtain multi-scale fused features; Use the multi-scale fused features to output the output features of the multi-scale gated convolution module through a linear layer.

[0016] Furthermore, the output features of the last multi-scale gated convolution module in the time-domain network are used as the time-domain features; The output features of the last multi-scale gated convolution module in the frequency-domain network are used as the frequency-domain features.

[0017] Advantages and positive effects of the present invention: The present invention expands the imbalanced dataset to a balanced state by generating high-quality virtual data, thereby optimizing the training effect of the device fault recognition model; and introduces frequency-domain features into the fault recognition model, and through the multi-scale gated convolution module and the efficient adaptive fusion module, realizes the adaptive fusion of multi-source sensor time-frequency domain features, significantly improving the prediction accuracy of the model.

[0018] 1) The present invention introduces a denoising diffusion probability model for data augmentation, and expands the imbalanced dataset to a balanced state by generating virtual data. Based on the existing denoising diffusion probability model, the DiffWave network is improved. The improved DiffWave network introduces a linear layer to filter high-frequency noise, and adopts a dual-branch architecture to adjust the dynamic global features of the bidirectional dilated convolution, thereby enhancing the attention to key information and improving the quality of the generated virtual data.

[0019] 2) The device fault identification model of the present invention adds a frequency-domain feature branch to enhance the recognition accuracy of the classification model. And through the multi-scale gated convolution module, parallel large-kernel multi-scale depth convolution is used to adjust the linear features, and adaptive gating is adopted for weighted multi-scale fusion; the problems of limited diversity of the generated virtual data and inconsistent distribution and duration of key events in the time domain and frequency domain are solved.

[0020] 3) An efficient adaptive fusion module is set in the device fault identification model of the present invention, which is used to realize the two-stage fusion of time-domain and frequency-domain features of multiple sensors; and an efficient grouping mechanism and cross-space learning are combined in the attention feature fusion technology, which can save computing resources and enhance local-global information integration at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0022] Figure 1 It is a flowchart of the multi-sensor time-frequency fusion device fault diagnosis method under the class imbalance condition in Embodiment 1; Figure 2 It is a structural diagram of the multi-source multi-scale time-frequency fusion convolutional neural network in Embodiment 1; Figure 3 It is a flowchart of the downsampling module and the multi-scale gated convolution module in Embodiment 2; Figure 4 It is a structural diagram of the efficient adaptive fusion module in Embodiment 2; Figure 5 It is a structural diagram of the improved DiffWave network in Embodiment 3; Figure 6 It is a module diagram of the rotating equipment test bench in the experiment; Figure 7 It is the data generated by Sensor 1 based on the denoising diffusion probability model in the experiment; Figure 8Data generated by sensor 2 based on the denoising diffusion probability model in the experiment; Figure 9 Graphs showing the model classification results based on different sensor data before and after the denoising diffusion probability model generates data in the experiment. Detailed implementation manner

[0023] To enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0024] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0025] The present invention provides a method for fault diagnosis of multi-sensor time-frequency fusion equipment under class imbalance conditions. The concept is as follows: Use a denoising diffusion probability model to perform data augmentation on the minority class samples of multi-sensors to suppress data imbalance. This model injects progressive Gaussian noise during the forward diffusion process, and then uses an improved DiffWave (DiffusionWaveform Network) to learn the reverse generation process. In terms of the classification network: 1) Through a multi-scale gated convolution module, enhance the network's ability to extract diverse features from time-domain and frequency-domain data. 2) Construct an efficient adaptive feature fusion module to assign learnable weights to the features of different branches in the channel and position dimensions to achieve feature-level data fusion. 3) The constructed multi-source multi-scale time-frequency fusion convolutional neural network uses a two-stage efficient adaptive feature fusion module to perform data fusion in the deep layer of the network, and comprehensively uses the time-domain and frequency-domain information of different sensor data for fault diagnosis.

[0026] Embodiment 1 Combined with Figure 1As shown in the figure, the method for fault diagnosis of multi-sensor time-frequency fusion equipment under class imbalance conditions in this embodiment includes the following steps: S1. Based on the original class-imbalanced sensor data set, construct a class-balanced data set for the equipment through a denoising diffusion probability model; S2. Train an equipment fault recognition model through the class-balanced data set for the equipment; S3. Utilize the pre-trained fault recognition model to receive multi-source sensor data and output the equipment fault type.

[0027] The equipment fault recognition model in this embodiment is a multi-source multi-scale time-frequency fusion convolutional neural network, which improves the diagnostic performance by fusing the time-domain and frequency-domain diversity features of multi-sensor data. The steps include: 1) Receive the time-domain data of multiple sensors, and use Fourier transform to convert the time-domain data of each sensor into frequency-domain data; 2) Through a multi-branch backbone network, based on the time-domain data and frequency-domain data of each sensor, extract the dual-domain features of each sensor's data; Each backbone network includes: a time-domain network, a frequency-domain network, and a first-stage efficient adaptive fusion module; the time-domain network extracts time-domain features; the frequency-domain network extracts frequency-domain features; The time-domain network and the frequency-domain network have the same structure, both including: multiple multi-scale gated convolutional modules and downsampling modules; abstract features are extracted through the multi-scale gated convolutional modules, and the transformation size is extracted through the downsampling modules.

[0028] From the perspective of the change trend of data, there are differences in the occurrence time and duration of key events between time-domain data and frequency-domain data. Using a convolutional kernel with a fixed size to obtain local information may be limited; therefore, the multi-scale gated convolutional module uses multiple different per-channel convolutions to obtain spatial information at different scales, and combines the gated unit mechanism to assign learnable weights to information at different scales.

[0029] 3) Through the first-stage efficient adaptive fusion module, integrate the time-domain features and frequency-domain features of the data of a single sensor; achieve feature-level data fusion by assigning learnable weights to the time-domain features and frequency-domain features of a single sensor; 4) Through the second-stage efficient adaptive fusion module, integrate the time-domain features and frequency-domain features of the data of multiple sensors; assign learnable weights to the data of each sensor to achieve multi-source data fusion.

[0030] Embodiment 2 Combined with Figure 2 As shown in the figure, the backbone network is further described: In this embodiment, preferably, receive the data of 2 sensors; Multi-scale features are obtained through downsampling, and the ability to extract features from time-frequency information is enhanced through a multi-scale gated convolution module. The structure is as shown in Figure 3 ; 1) Downsampling module; Multi-scale features are obtained through downsampling, and the formula is expressed as:

[0031]

[0032]

[0033] Among them, represents the sensor time-domain data or frequency-domain data, and the size is expressed as B×C×L, which is 7×11×15 in this embodiment; , and are features of three scales, is a one-dimensional channel-wise convolution with a size of 7, 11, 15, is a linear layer.

[0034] 2) Multi-scale gated convolution module; The weights corresponding to the features of each scale are obtained by using the gating mechanism:

[0035]

[0036] Among them, Stack is a stacking operation, is a linear layer, represents the pre-fused features, represents the weights corresponding to the multi-scale features; 3 represents the number of multi-scale features; represents the summation index variable.

[0037] The features of the three scales are respectively subjected to Hadamard product with the weights and added to obtain the multi-scale fusion features and the output features of the multi-scale gated convolution module. The formula is expressed as:

[0038]

[0039] Among them, is the output feature of the multi-scale gated convolution module, is a linear layer, is the Hadamard product, and A represents the multi-scale fusion features.

[0040] The output features of the last multi-scale gated convolution module in the time-domain network are used as time-domain features; the output features of the last multi-scale gated convolution module in the frequency-domain network are used as frequency-domain features.

[0041] 3) Combine Figure 4 As shown, the efficient adaptive fusion module in the method of the present invention combines an efficient grouping mechanism and cross-space learning, which not only saves computing resources but also improves the ability of the attention feature to fuse local and global information.

[0042] Taking the first-stage efficient adaptive fusion module as an example to fuse the time-domain features and frequency-domain features of the same sensor, the efficient adaptive fusion module is further described as follows: First, feature fusion, which is expressed by the formula:

[0043] Among them, Group represents the grouping operation, which incorporates part of the channel dimension into the batch dimension to reduce the computational cost; ; In this embodiment, represents the time-domain feature, represents the frequency-domain feature; P is the fused feature; Next, local and global information is obtained, and cross-space interaction is performed, which is expressed by the formula:

[0044]

[0045]

[0046] Among them, represents the local feature, represents the global feature, is a one-dimensional per-channel convolution with a size of 7, GAP is the global average pooling operation, represents a linear layer, is the Sigmoid function, is the intermediate feature integrating local-global information.

[0047] Finally, two weights are used to reallocate the input features:

[0048]

[0049] Among them, Reshape is the operation to restore the channel and batch dimension sizes, , is the learnable weight, is the output of the first-stage efficient adaptive fusion module.

[0050] In this implementation, the output of the first-stage efficient adaptive fusion module is the dual-domain features of a single sensor.

[0051] Similarly, the structure of the second-stage efficient adaptive fusion module is the same, as shown in Figure 3 Figure; that is, Input 1 and Input 2 are the dual-domain features of two sensors, and the output is the multi-source features of the device.

[0052] Embodiment 3 Based on the original class-imbalanced sensor dataset, a device class-balanced dataset is constructed through a denoising diffusion probability model; In this embodiment, the vibration time-domain signals of the device under various fault conditions are simultaneously collected by multiple sensors to construct the original class-imbalanced sensor dataset.

[0053] Existing data augmentation methods are difficult to generate high-quality fault samples with a real distribution.

[0054] Therefore, in this embodiment, a denoising diffusion probability model is used to construct a device class-balanced dataset by generating virtual data.

[0055] The DiffWave network in the existing denoising diffusion probability model has deficiencies in noise suppression and feature preservation; Therefore, in this embodiment, an improved DiffWave network is used in the reverse process of the denoising diffusion probability model to generate a balanced virtual-real mixed dataset for training the fault recognition model.

[0056] The denoising diffusion probability model includes two stages: diffusion and inverse diffusion: In the first stage, Gaussian noise is gradually added to the original data through a Markov chain. After a preset number of steps, the data degenerates into pure noise, forming a forward diffusion chain.

[0057] In the second stage, a neural network is trained to learn the reverse denoising path. The neural network model gradually reconstructs the original data distribution from random noise by predicting the noise distribution at the current time step.

[0058] In this embodiment, the neural network is an improved DiffWave network, as shown in Figure 4 Figure. The improvements include: introducing efficient multi-scale attention to enhance local multi-scale feature extraction, and adding two linear layers to suppress high-frequency noise.

[0059] The formulas of the two linear layers are expressed as:

[0060] Among them, Linear represents the linear layer, and BiDilatedConv represents the bidirectional dilated convolution; represents the output of the efficient multi-scale attention module, y represents the output processed by the convolutional modulation technique.

[0061] Different from the traditional training using a dataset composed entirely of real data, in this embodiment, a virtual-real hybrid balanced dataset after data augmentation is used for training to achieve multi-sensor collaborative fault diagnosis under unbalanced conditions.

[0062] The effect of the present invention is further verified by experiments: S1. Collect data of the gearbox under various operating conditions through a rotating equipment test bench.

[0063] Such as Figure 6 The shown test bench includes: a controller, a brake, a brake controller, a gearbox, a motor, and two acceleration sensors. The two acceleration sensors respectively collect vibration signals in the vertical and horizontal directions. Five operating conditions are simulated in the experiment, and the samples of the faulty operation are set as few samples, and the balance ratio is set to 2%. The detailed classification of the class labels and samples is shown in Table 1.

[0064] Table 1

[0065] S2. Set the sample length to 1024. In the denoising diffusion probability model, the time step is set to 4000, the learning rate is set to 0.0002, and the step size T is set to 200. In the equipment fault recognition model, the number of cycles is set to 50, the Adam optimizer is used, and the learning rate is set to 0.001 and decreases as the number of iterations decreases.

[0066] S3. Experimental process: First, use the denoising diffusion probability model to generate data, and restore the dataset with a balance ratio of 2% to a virtual-real hybrid dataset with a balance ratio of 100%. At this time, the number of training samples for each category is 500. Next, use the virtual-real hybrid dataset to train the multi-source multi-scale time-frequency fusion convolutional neural network. Finally, use the real samples to test the model performance.

[0067] Experimental results: Figure 7 shows the generation results of the denoising diffusion probability model for sensor 1, Figure 8 shows the generation results of the denoising diffusion probability model for sensor 2; where the red and blue lines respectively represent the generated data and the real data. It can be seen from the time and frequency domain graphs that the generated virtual samples are very close to the real samples, but still show different diversities from the real samples. This is preliminary evidence of the effectiveness of the denoising diffusion probability.

[0068] Figure 9Model classification results based on different sensor data before and after generating data for the denoising diffusion probabilistic model. When the balance ratio is 2%, the diagnostic accuracies based on the data of sensor 1 and sensor 2 are 89.7% and 87.3% respectively, and the accuracy based on the data of sensor 1+2 is 96.4%. When the balance ratio is 100%, these results are improved to 97.9%, 97.1% and 99.1% respectively. The denoising diffusion probabilistic model significantly improves the diagnostic performance. On the unbalanced dataset and the balanced virtual-real mixed dataset, the accuracy of the multi-source multi-scale time-frequency fusion convolutional neural network based on the data of sensor 1+2 is higher than that based on the single-sensor data. Therefore, the experimental results verify the effectiveness of the method of the present invention.

[0069] S4. Further, through ablation experiments, verify the effectiveness and rationality of the key steps in the method of the present invention, and the experimental results are shown in Table 2.

[0070] To verify the effectiveness of combining frequency domain features, Method A was constructed, which removed the frequency domain network of the equipment fault identification model.

[0071] To verify the role of the multi-scale gated convolution, Method B was constructed, which removed the multi-scale convolution and gated mechanism of the equipment fault identification model.

[0072] To verify the role of the efficient adaptive fusion module, Method C was constructed, which removed the efficient adaptive module of the equipment fault identification model and only used the conventional addition operation for fusion.

[0073] To verify the role of the improved DiffWave network in the denoising diffusion probabilistic model, Method D was constructed, which used the original DiffWave network to generate data.

[0074] Table 2 Results of ablation experiments

[0075] It can be seen that the proposed method is superior to other methods, and the accuracy of the methods after removing several key operations has decreased significantly. The improved DiffWave network can improve the diagnostic performance by 2.69%. In addition, the experimental results prove the effectiveness of the proposed multi-scale gated convolution, efficient adaptive fusion module and introducing the frequency domain branch.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements 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 method for fault diagnosis of multi-sensor time-frequency fusion equipment under quasi-unbalanced conditions, characterized in that: include: Based on the original class-imbalanced sensor dataset, a device-class balanced dataset is constructed through a denoising diffusion probability model; Train the equipment fault identification model through the equipment class balanced dataset; Use the pre-trained equipment fault identification model to receive multi-source sensor data and output the equipment fault type; The equipment fault identification model is a multi-source multi-scale time-frequency fusion convolutional neural network, including: Receive time domain data from multiple sensors and convert the time domain data of each sensor into frequency domain data using Fourier transform; Through a multi-branch backbone network, dual-domain features of each sensor data are extracted based on the time domain data and frequency domain data of each sensor; The second-stage efficient adaptive fusion module is used to fuse the dual-domain features of each sensor to obtain the multi-source features of the device.

2. According to the method for fault diagnosis of multi-sensor time-frequency fusion equipment under quasi-unbalanced conditions described in claim 1, it is characterized in that: Each branch backbone network includes: a time domain network, a frequency domain network and a first-stage efficient adaptive fusion module; The time domain network extracts time domain features; The frequency domain network extracts frequency domain features; The first-stage efficient adaptive fusion module fuses the time domain features and the frequency domain features to obtain the dual-domain features of the sensor data.

3. According to the method for diagnosing faults of multi-sensor time-frequency fusion equipment under quasi-unbalanced conditions described in claim 1, it is characterized in that: The method of constructing a device class balanced data set based on the original class unbalanced sensor data set by using a denoising diffusion probability model includes: Collect the original class-unbalanced sensor data of the device; Through the denoising diffusion probability model, Gaussian noise is gradually added to the original class-imbalanced sensor data, which is converted into pure noise after a preset step to form a Markov chain, destroying the original data distribution; The reverse denoising path is learned through the neural network, the original data distribution is reconstructed, and various types of balanced distributed sensor data are generated.

4. According to the method for diagnosing faults of multi-sensor time-frequency fusion equipment under quasi-unbalanced conditions as described in claim 3, it is characterized in that: The neural network comprises: Improved DiffWave network; The improvements to the DiffWave network include: Based on the DiffWave network, efficient multi-scale attention is introduced to enhance local multi-scale feature extraction; And two linear layers are added to suppress high-frequency noise.

5. According to the method for diagnosing faults of multi-sensor time-frequency fusion equipment under quasi-unbalanced conditions as described in claim 2, it is characterized in that: The time domain network comprises: Several downsampling modules and multi-scale gated convolution modules; The downsampling modules and the multi-scale gated convolution modules are arranged alternately; The downsampling module extracts the transformation size and outputs multi-scale features; The multi-scale gated convolution module fuses multi-scale features and outputs multi-scale fused features.

6. According to the method for diagnosing faults of multi-sensor time-frequency fusion equipment under quasi-unbalanced conditions as described in claim 5, it is characterized in that: The time domain network has the same structure as the frequency domain network.

7. According to the method for diagnosing faults of multi-sensor time-frequency fusion equipment under quasi-unbalanced conditions as described in claim 2, it is characterized in that: The first-stage efficient adaptive fusion module includes: splicing the time domain features and the frequency domain features, and performing grouping operations to obtain fusion features; Convolution and pooling are performed on the fused features to obtain local features and global features; Integrate local features and global features to obtain intermediate features; Use the intermediate features to determine the corresponding weights of time domain features and frequency domain features; Based on the corresponding weights, the time domain features and the frequency domain features are weighted and summed to obtain dual-domain features.

8. According to the method for diagnosing faults of multi-sensor time-frequency fusion equipment under quasi-unbalanced conditions as described in claim 5, it is characterized in that: The multi-scale gated convolution module comprises: The corresponding weights of each scale feature are obtained through the gating mechanism; Based on the weights corresponding to the features of each scale, the multi-scale features are weightedly fused to obtain multi-scale fused features; The multi-scale fusion features are used to output the multi-scale gated convolution module output features through a linear layer.

9. According to the method for diagnosing faults of multi-sensor time-frequency fusion equipment under quasi-unbalanced conditions as described in claim 8, it is characterized in that: The output feature of the last multi-scale gated convolution module in the time domain network is used as the time domain feature; The output features of the last multi-scale gated convolution module in the frequency domain network are used as the frequency domain features.

Images