A rotating machinery fault diagnosis device and method

Abstract

The present invention discloses a rotating machinery fault diagnosis device and method based on a dual-branch multi-scale fusion network of time-frequency images, including: a data acquisition module for acquiring vibration signals of rotating machinery equipment; a data preprocessing module for dividing and preprocessing the vibration signal data by using a sliding window segmentation function; a data transformation module for converting the vibration signal data into time-frequency images and time-frequency features by using continuous wavelet transform; a feature extraction module including a time-frequency feature extractor and a multi-scale feature extractor, respectively analyzing and processing the time-frequency features to extract numerical branch features and analyzing and processing the time-frequency images to extract multi-scale visual branch features; and a dynamic weighted fusion module for dynamically generating weights to perform weighted fusion on the numerical branch features and the multi-scale visual branch features and outputting a fault diagnosis result. The rotating machinery fault diagnosis device and method of the present invention have high rotating machinery fault diagnosis efficiency and accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and particularly to a rotating machinery fault diagnosis device and method based on a dual-branch multi-scale fusion network of time-frequency images. Background Technique

[0002] Rotating machinery is widely used in various fields such as industry, energy, and transportation, such as equipment like steam turbines, generators, compressors, and pumps. These mechanical devices will inevitably have various faults during long-term operation, such as bearing damage, imbalance, gear wear, and rotor eccentricity. If these faults are not diagnosed and processed in time, it will not only lead to a decline in equipment performance but also may cause serious safety accidents and economic losses. With the development of industrial automation and intelligence, higher requirements are put forward for the real-time monitoring and fault diagnosis of the operating state of rotating machinery.

[0003] Traditional fault diagnosis methods usually rely on experienced technicians to judge the equipment condition by manually observing and analyzing data such as vibration signals. However, this method has problems such as strong subjectivity and low diagnosis efficiency, and it is difficult to meet the needs of the complex production environment of modern industry. Therefore, developing a high-precision and high-efficiency fault diagnosis method has important practical significance for comprehensively preventing equipment failures, minimizing downtime to the greatest extent, and significantly reducing maintenance costs.

[0004] To solve problems such as the unsatisfactory fault diagnosis effect under complex working conditions, some fault diagnosis methods for non-stationary and non-linear signal working condition data sets have been proposed. For example, the Chinese patent document with the publication number CN119494022A discloses a method for constructing a mechanical equipment fault diagnosis model, a diagnosis method, a system, and a device, which relates to the technical field of mechanical fault diagnosis. The fault diagnosis method includes: obtaining vibration signal data of rotating machinery in real time; inputting the vibration signal data into a fault diagnosis model to obtain the fault type of the rotating machinery. The Chinese patent document with the publication number CN113177537A discloses a fault diagnosis method for rotating mechanical equipment, including: extracting time-domain feature data, frequency-domain feature data, and time-frequency domain feature data of the original vibration signal of the equipment to be diagnosed; performing curve secondary coding on the time-frequency domain feature data to obtain a curve secondary coding sequence; calculating the distances between the curve secondary coding sequence and the clustering centers of multiple training samples of known fault types to obtain multi-dimensional distance feature data of the original vibration signal; inputting the time-domain feature data, frequency-domain feature data, and distance feature data of the original vibration signal of the equipment to be diagnosed into a trained fault diagnosis model to obtain the fault type of the equipment to be diagnosed.

[0005] However, directly processing vibration signals has significant deficiencies. Vibration signals are usually one-dimensional and it is difficult to simultaneously express the changing information of time and frequency. In the present invention, a one-dimensional vibration signal is transformed into a two-dimensional time-frequency image by using continuous wavelet transform (CWT). Among them, using the Morlet mother wavelet function can more effectively capture the instantaneous characteristics and dynamic change laws of the signal, especially suitable for non-stationary signals, which makes the ability to capture the characteristics of non-stationary signals weak; at the same time, directly processing vibration signals is more sensitive to noise and easily interferes with the diagnosis results; in addition, it often requires complex artificial feature extraction, relying on expert experience, with low efficiency and strong subjectivity.

[0006] Therefore, it is necessary to find an effective method to solve the problems faced in processing vibration signals in fault diagnosis and improve the efficiency and accuracy of the fault diagnosis method. Summary of the Invention

[0007] The present invention provides a rotating machinery fault diagnosis device and method based on a dual-branch multi-scale fusion network of time-frequency images, which has high rotating machinery fault diagnosis efficiency and accuracy.

[0008] The technical solution of the present invention is as follows:

[0009] A rotating machinery fault diagnosis device based on a dual-branch multi-scale fusion network of time-frequency images, comprising:

[0010] A data acquisition module for acquiring the vibration signal data to be detected of the rotating machinery equipment;

[0011] A data preprocessing module for dividing the vibration signal data to be detected by using a sliding window segmentation function and performing normalization and standardization preprocessing;

[0012] A data transformation module for converting the preprocessed vibration signal data to be detected into a time-frequency image and corresponding time-frequency features by using continuous wavelet transform;

[0013] A feature extraction module, including a parallel time-frequency feature extractor and a multi-scale feature extractor; the time-frequency feature extractor analyzes and processes the time-frequency features to extract numerical branch features, and the multi-scale feature extractor analyzes and processes the time-frequency image to extract multi-scale visual branch features;

[0014] A dynamic weighted fusion module for weighted fusion of the numerical branch features and the multi-scale visual branch features through dynamically generated weights and outputting a fault diagnosis result.

[0015] When capturing the dynamic changes and non-linear characteristics of signals, the present invention uses a dual-branch feature extraction module to process the time-frequency image and time-frequency features respectively. While considering the processing of the time-frequency image, the multi-scale feature branch also takes into account the multi-scale characteristics of vibration signals. Finally, a dynamic weighted fusion module is used for merging, which can capture the dynamic changes and non-linear characteristics of vibration signals more precisely, thereby improving the accuracy of fault diagnosis and the robustness in complex fault patterns and noise environments.

[0016] Preferably, the segment length of the sliding window is 1024 and the overlapping length is 512. Ensure that the segmented signals cover multiple cycles and retain sufficient time resolution.

[0017] Preferably, the continuous wavelet transform uses the Morlet mother wavelet, and the scale range is 1 - 128.

[0018] The preprocessed vibration signal data to be detected is converted into a time-frequency image and corresponding time-frequency features through continuous wavelet transform. The time-frequency image is used for feature extraction in the visual branch of the model, and the feature file is used for feature extraction in the numerical branch of the model. By fusing the information in the time domain and frequency domain, the dynamic characteristics of non-stationary signals can be effectively captured, especially the law of short-time transient changes during fault occurrence, which can more intuitively display the multi-scale characteristics of the signal, help the model automatically extract more comprehensive and accurate features, improve the model's recognition ability for fault patterns under complex working conditions, and thus significantly improve the accuracy and robustness of fault diagnosis.

[0019] The time-frequency feature extractor is constructed based on a convolutional neural network and is used for layer-by-layer extraction and optimization of time-frequency features.

[0020] The time-frequency feature extractor includes a number of convolutional pooling modules; each convolutional pooling module includes an upsampling layer, a convolutional layer, a BatchNorm2d layer, a PReLU activation layer, a Dropout layer, and a one-dimensional MaxPool layer; an attention module is introduced after each layer of the convolutional pooling module.

[0021] The upsampling layer is used to adjust the input to a fixed dimension, adapt to the requirements of the subsequent network structure, and enhance the uniformity of the input. The convolutional layer is used to extract fault features in the time-frequency features. The BatchNorm2d layer is used to accelerate the model convergence, improve the stability of the training process, and enhance the robustness of the model. The Dropout layer is used to improve the training efficiency and generalization ability. The PReLU layer is used to adapt to the dynamic changes of different features. The one-dimensional MaxPool layer is used to downsample the extracted features. The features extracted by different convolutional layers are sent to the attention module respectively. Each attention module generates corresponding attention weights through the compression and reweighting mechanism of global features, and dynamically adjusts the channel importance of the input features, so as to obtain the feature representation after channel attention weighting. Subsequently, the output weighted features are further compressed in size through pooling operations, and after splicing or stacking, a multi-level feature representation is formed.

[0022] The time-frequency feature extractor of the present invention can not only fully extract the deep features in the time-frequency features, but also highlight the role of key features through the dynamic weighting mechanism, significantly improving the model's ability to analyze the fault patterns of non-stationary signals.

[0023] The multi-scale feature extractor is used to extract multi-scale features of the time-frequency image.

[0024] The multi-scale feature extractor includes several multi-convolution kernel modules and a dimensionality reduction module. The multi-convolution kernel module includes multiple convolution kernels of different sizes, which are used to extract the detailed features, medium-scale spatial features, and global range features of the time-frequency image. The dimensionality reduction module includes a global average pooling layer and a fully connected layer, which perform dimensionality reduction processing on the output results of the multi-convolution kernel module.

[0025] Each multi-convolution kernel module includes a 1×1 convolution kernel, a 3×3 convolution kernel, and a 5×5 convolution kernel. The 1×1 convolution kernel is used to capture local detailed features, the 3×3 convolution kernel is used to extract medium-scale spatial features, and the 5×5 convolution kernel is used to obtain the feature representation of the global range, so as to effectively cover the information at different scales in the time-frequency image and avoid the risk of missing key features by a single convolution kernel. After the multi-scale features are extracted, the output results are processed by the dimensionality reduction module, specifically including global average pooling and fully connected operations, mapping the multi-channel features into a low-dimensional unified representation to ensure that the feature dimensions match the subsequent weighted fusion operations.

[0026] The multi-scale feature extractor of the present invention can fully capture the local, detailed, and global characteristics of the time-frequency image, providing comprehensive and accurate feature support for the recognition of complex fault patterns.

[0027] The described dynamic weighted fusion module includes a feature concatenation unit, a weight generation unit, and a weighted fusion unit; the feature concatenation unit concatenates the feature vectors output by the time-frequency feature extractor and the multi-scale feature extractor in the channel dimension to form a joint feature representation; the weight generation unit processes the joint feature through a fully connected network to generate the weights of each branch, and normalizes the weights using the Softmax function; the weighted fusion unit performs weighted summation on the feature vectors of the two branches according to the generated weights to generate a fused feature representation and outputs the fault diagnosis result.

[0028] Preferably, the training methods of the feature extraction module and the dynamic weighted fusion module include:

[0029] Collect the vibration signal generated by the rotating machinery under different working conditions, perform sliding window segmentation processing on the vibration signal, and generate time-frequency images and corresponding time-frequency features through continuous wavelet transform to construct a training data set;

[0030] Use the training data set to train the feature extraction module and the dynamic weighted fusion module. During the training process, the time-frequency feature extractor and the multi-scale feature extractor respectively extract features from the input time-frequency features and time-frequency images. The dynamic weighted fusion module performs weighted fusion on the feature vectors of the two branches and outputs the predicted fault category label;

[0031] Calculate the deviation value between the current output and the expected target through forward propagation, and continuously optimize and adjust the weights and biases of the parameters of each layer through the error backpropagation algorithm.

[0032] Furthermore, the cross-entropy loss function is used as the objective function in the optimization process, and the parameters are updated through the Adam optimizer. The initial learning rate is set to 0.005, and the training is iterated 50 rounds.

[0033] A rotating machinery fault diagnosis method based on a two-branch multi-scale fusion network of time-frequency images includes:

[0034] The data acquisition module obtains the vibration signal data to be detected of the rotating machinery equipment;

[0035] The data preprocessing module uses the sliding window segmentation function to divide the vibration signal data to be detected and performs normalization and standardization preprocessing;

[0036] The data transformation module uses continuous wavelet transform to convert the preprocessed vibration signal data to be detected into time-frequency images and corresponding time-frequency features;

[0037] The feature extraction module includes a time-frequency feature extractor and a multi-scale feature extractor in parallel; the time-frequency feature extractor analyzes and processes the time-frequency features to extract numerical branch features, and the multi-scale feature extractor analyzes and processes the time-frequency images to extract multi-scale visual branch features;

[0038] The dynamic weighted fusion module performs weighted fusion on the numerical branch features and the multi-scale visual branch features through dynamically generated weights, and outputs the fault diagnosis result.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] The present invention uses the continuous wavelet transform module to convert the vibration signal into a time-frequency image and time-frequency features, which can fully combine the information in the time domain and the frequency domain, effectively capture the key dynamic features in the non-stationary signal, especially the law of short-time transient changes when a fault occurs. Compared with directly processing one-dimensional signals, the time-frequency image can more intuitively display the multi-scale characteristics of the signal, helping the model to automatically extract more comprehensive and accurate features, improving the model's recognition ability for fault patterns under complex working conditions, and thus significantly improving the accuracy and robustness of fault diagnosis.

[0041] The present invention combines time-frequency feature analysis and multi-scale feature analysis. The time-frequency feature branch extracts high-dimensional time-frequency features through convolutional layers and pooling layers, which can capture the dynamic change law of the signal; the multi-scale feature branch extracts features at different scales through various scale convolutions and pooling operations and enhances the robustness of the model, making full use of the global features of the time-frequency image and the local detail information of the multi-scale features, improving the fault diagnosis ability for non-stationary and complex signals; the multi-scale feature branch effectively reduces the computational complexity through the dimensionality reduction module while retaining key information; the dynamic weighted fusion module adaptively generates weights for the two branches, can allocate the optimal fusion ratio according to the characteristics of different input data, and significantly improves the accuracy and generalization ability of the diagnosis result.

[0042] The present invention exhibits good performance in feature extraction, fusion, and model robustness, and is applicable to the fault diagnosis of rotating machinery in industrial scenarios with non-stationary and non-linear signals. Description of the Drawings

[0043] Figure 1 It is a schematic structural principle diagram of the rotating machinery fault diagnosis method and system of the double-branch multi-scale fusion network based on time-frequency images of the present invention;

[0044] Figure 2 It is a training flow chart of the diagnosis model of the present invention;

[0045] Figure 3 It is a fault recognition accuracy chart of the fault diagnosis model of the present invention. Detailed Embodiments

[0046] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be noted that the following embodiments are intended to facilitate the understanding of the present invention and do not impose any limitation on it.

[0047] A method and system for rotating machinery fault diagnosis based on a double-branch multi-scale fusion network of time-frequency images of the present invention, as Figure 1 shown, uses a continuous wavelet transform module to convert the vibration data in the dataset into time-frequency images; constructs a double-branch multi-scale convolutional network architecture for feature extraction of the preprocessed time-frequency images, including a time-frequency feature analysis module and a multi-scale feature analysis module; builds a dynamic weighted fusion module for integrated output of data; builds a time-frequency double-branch multi-scale fusion network fault diagnosis model based on the double-branch multi-scale convolutional network architecture and the dynamic weighted fusion module, and trains the model for fault diagnosis.

[0048] The sequence data of the vibration signal is used as the input signal. After standardizing the input signal, the signal is segmented using a sliding window segmentation function, with each segment having a length of 1024 and an overlap length of 512, ensuring that the segmented signal covers multiple cycles and retains sufficient time resolution. Subsequently, the signal data is processed using continuous wavelet transform, selecting the Morlet mother wavelet, and the scale range is set from 1 to 128. The generated time-frequency images and corresponding characteristic files are saved separately. The time-frequency images are used for feature extraction of the visual branch of the model, and the characteristic files are used for feature extraction of the numerical branch of the model. By fusing the information in the time domain and the frequency domain, the dynamic characteristics of non-stationary signals can be effectively captured, especially the laws of short-time transient changes during fault occurrence, which can more intuitively display the multi-scale characteristics of the signal, help the model automatically extract more comprehensive and accurate features, and improve the model's recognition ability for fault patterns under complex working conditions, thus significantly improving the accuracy and robustness of fault diagnosis.

[0049] The time-frequency feature branch is based on a convolutional neural network module, consisting of a convolutional pooling module and an attention module, and is used to extract and optimize time-frequency features layer by layer. The convolutional pooling module includes an upsampling layer, a convolutional layer, a BatchNorm2d layer, a PReLU activation layer, a Dropout layer, and a one-dimensional MaxPool layer. The upsampling layer adjusts the dimension of the input time-frequency features to a fixed size. The convolutional layer extracts local features and global patterns of the signal through a 3×3 convolutional kernel. The BatchNorm2d layer accelerates the model convergence and improves stability. The PReLU activation layer enhances the non-linear expression ability of features. The Dropout layer reduces the risk of overfitting. The MaxPool layer extracts key information by reducing the size of the feature map. In addition, an attention module is introduced after each layer of the convolutional pooling module. The importance of channels is calculated through global average pooling, and dynamic weights are generated through a fully connected network to weight and optimize the feature channels, thereby strengthening the expression ability of key features. This design can not only fully extract deep features in time-frequency features, but also highlight the role of key features through a dynamic weighting mechanism, significantly improving the model's ability to analyze non-stationary signal fault patterns.

[0050] Taking the dual-branch multi-scale fusion network architecture as an example, assuming that the dimension of the time-frequency features after continuous wavelet transform (CWT) is (127, 1024), and the number of classes (num_class) is set to 10, the network parameters of the time-frequency feature branch are as shown in Table 1 below:

[0051]

[0052] The multi-scale feature branch is based on a multi-scale convolutional structure, including a multi-convolutional kernel module composed of convolutional kernels of different sizes (such as 1×1, 3×3, and 5×5) and a dimensionality reduction module of average pooling, and is used to extract multi-scale features of the time-frequency image. The multi-convolutional kernel module captures local detail features through a 1×1 convolutional kernel, extracts mid-scale spatial features through a 3×3 convolutional kernel, and obtains global range feature representations through a 5×5 convolutional kernel, thereby effectively covering information at different scales in the time-frequency image and avoiding the risk of missing key features by a single convolutional kernel. After the multi-scale features are extracted, the output results are processed by a dimensionality reduction module, specifically including global average pooling and fully connected operations, mapping multi-channel features into a low-dimensional unified representation to ensure that the feature dimensions match the subsequent weighted fusion operations. The multi-scale design of this branch can fully capture the local, detailed, and global characteristics of the time-frequency image, providing comprehensive and accurate feature support for the recognition of complex fault patterns.

[0053] Taking the dual-branch multi-scale fusion network architecture as an example, assuming that the dimension of the time-frequency image after continuous wavelet transform (CWT) is (3, 127, 128), and the number of classes (num_class) is set to 10, the network parameters of the multi-scale feature branch are as shown in Table 2 below:

[0054]

[0055] The dynamic weighted fusion module adaptively integrates the outputs of the time-frequency feature branch and the multi-scale feature branch through three steps: feature splicing, weight generation, and weighted fusion. First, the feature vectors of the two branches are spliced in the channel dimension to form a joint feature representation. Then, the joint feature is processed by a fully connected network to generate two fusion weights. The first fully connected layer reduces the dimension to extract important features, and the second fully connected layer outputs the fusion weights. Finally, the Softmax function is used to normalize the weights to ensure that their sum is 1. Finally, the feature vectors of the two branches are weighted and summed according to the generated weights to generate a fusion feature representation and output the fault diagnosis result. By dynamically adjusting the importance of branch features, this module achieves adaptive optimization based on the input data, which not only improves the robustness of the fusion features but also enhances the fault diagnosis ability of the model under complex working conditions.

[0056] As Figure 2 shown, to train the fault diagnosis model of the present invention, vibration signals generated by a rotating machine under different working conditions are collected, such as vibration signals in the normal working state and various fault states, and they are organized into corresponding data sets. Subsequently, the vibration signals are processed by sliding window segmentation, with each segment having a length of 1024 points and an overlap of 512 points, and time-frequency images and corresponding time-frequency characteristic files are generated through continuous wavelet transform. The organized data set is randomly sampled, with 70% as the training set, 15% as the validation set, and 15% as the test set. The training set and the validation set are used to train the model. During the training process of the diagnostic model network, the deviation value between the current output and the expected target is calculated through forward propagation, and the weights and biases of the parameters of each layer are continuously adjusted through the error backpropagation algorithm. During the training process, the time-frequency feature branch and the multi-scale feature branch respectively extract features from the input time-frequency images, and the dynamic weighted fusion module integrates and optimizes the feature vectors of the two branches and outputs the predicted fault class labels. The entire optimization process uses the cross-entropy loss function as the objective function, updates the parameters through the Adam optimizer, sets the initial learning rate to 0.005, and trains for 50 iterations. After multiple rounds of cyclic iteration, the performance of the network continuously improves on the training set and the validation set, and finally converges to the optimal solution, obtaining a model with a relatively high diagnostic accuracy. To verify the performance of the model, the reserved test set is input into the trained diagnostic model. The model extracts and fuses features layer by layer and outputs the corresponding predicted fault type labels. By comparing the model predicted labels with the true labels of the test set, the recognition accuracy of the fault diagnosis is calculated. As Figure 3The dynamic weighted fusion network of the present invention shows excellent fault diagnosis ability in the test set, with an average diagnosis accuracy rate reaching over 95%, demonstrating good robustness and complex working condition adaptation ability.

[0057] Comparison and verification with other models: The diagnostic model (DBMS-FNet) of the present invention is respectively compared with VIbration Analysis ML and LogReg-Diagnosis in the prior art, and the results are shown in Table 3 below:

[0058]

[0059] The results show that the model of the present invention can effectively improve the analysis ability of complex fault modes through the dual-branch multi-scale fusion network architecture and the dynamic weighted fusion module, demonstrating significant advantages. It innovatively combines the extraction and optimization of time-frequency features and multi-scale features, overcoming the problem of insufficient fault diagnosis accuracy of traditional methods under complex working conditions. Through the dynamic weighted mechanism, the adaptive optimization of different feature branches is realized, effectively improving the robustness and applicability of the model. In addition, the model design of the present invention has scalability and can be adapted to the fault diagnosis requirements of other mechanical systems, providing a new solution for industrial intelligent maintenance.

[0060] The above embodiments have detailed the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, supplements, equivalent replacements, etc. made within the scope of the principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A rotating machinery fault diagnosis device based on a dual-branch multi-scale fusion network of time-frequency images, characterized in that, Including: A data acquisition module that acquires vibration signal data to be detected of rotating machinery; A data preprocessing module that divides the vibration signal data to be detected by using a sliding window segmentation function and performs normalization preprocessing; A data transformation module that converts the preprocessed vibration signal data to be detected into a time-frequency image and corresponding time-frequency features by using continuous wavelet transform; A feature extraction module, including a parallel time-frequency feature extractor and a multi-scale feature extractor; The time-frequency feature extractor analyzes and processes the time-frequency features to extract numerical branch features, and the multi-scale feature extractor analyzes and processes the time-frequency image to extract multi-scale visual branch features; the time-frequency feature extractor includes a number of convolutional pooling modules; each convolutional pooling module includes an upsampling layer, a convolutional layer, a BatchNorm2d layer, a PReLU activation layer, a Dropout layer, and a one-dimensional MaxPool layer; the convolutional layer extracts local and global features of the signal through a 3×3 convolutional kernel; an attention module is introduced after each convolutional pooling module; the multi-scale feature extractor includes a number of multi-convolutional kernel modules and a dimensionality reduction module; the multi-convolutional kernel modules include convolutional kernels of multiple different sizes for extracting detailed features, medium-scale spatial features, and global range features of the time-frequency image; The dimensionality reduction module includes a global average pooling layer and a fully connected layer for performing dimensionality reduction processing on the output result of the multi-convolutional kernel module; A dynamic weighted fusion module that dynamically generates weights to perform weighted fusion on the numerical branch features and the multi-scale visual branch features and outputs a fault diagnosis result; the dynamic weighted fusion module includes a feature splicing unit, a weight generation unit, and a weighted fusion unit; the feature splicing unit splices the feature vectors output by the time-frequency feature extractor and the multi-scale feature extractor in the channel dimension to form a joint feature representation; the weight generation unit processes the joint feature through a fully connected network to generate weights for each branch and normalizes the weights using the Softmax function; the weighted fusion unit performs weighted summation on the feature vectors of the two branches according to the generated weights to generate a fused feature representation and outputs a fault diagnosis result.

2. The rotating machinery fault diagnosis device of the dual-branch multi-scale fusion network based on time-frequency images according to claim 1, characterized in that, The segment length of the sliding window is 1024 and the overlapping length is 512.

3. The rotating machinery fault diagnosis device based on the time-frequency image double-branch multi-scale fusion network according to claim 1, characterized in that, The continuous wavelet transform uses a Morlet mother wavelet and the scale range is 1-128.

4. The rotating machinery fault diagnosis device based on the time-frequency image double-branch multi-scale fusion network according to claim 1, characterized in that, Each multi-convolutional kernel module includes a 1×1 convolutional kernel, a 3×3 convolutional kernel, and a 5×5 convolutional kernel. The 1×1 convolutional kernel captures local detailed features, the 3×3 convolutional kernel extracts medium-scale spatial features, and the 5×5 convolutional kernel obtains a global range feature representation.

5. The rotating machinery fault diagnosis device based on the time-frequency image double-branch multi-scale fusion network according to claim 1, characterized in that, The training methods of the feature extraction module and the dynamic weighted fusion module include: Collecting vibration signals generated by rotating machinery under different working conditions, performing sliding window segmentation processing on the vibration signals, and generating a time-frequency image and corresponding time-frequency features through continuous wavelet transform to construct a training data set; The feature extraction module and the dynamic weighted fusion module are trained using a training dataset. During the training process, the time-frequency feature extractor and the multi-scale feature extractor respectively extract features from the input time-frequency features and time-frequency images. The dynamic weighted fusion module weights and fuses the feature vectors of the two branches and outputs the predicted fault class label; Calculate the deviation value between the current output and the expected target through forward propagation, and continuously optimize and adjust the weights and biases of the parameters of each layer through the error backpropagation algorithm.

6. The rotating machinery fault diagnosis device based on the dual-branch multi-scale fusion network of time-frequency images according to claim 5, characterized in that, The cross-entropy loss function is used as the objective function in the optimization process, and the parameters are updated by the Adam optimizer. The initial learning rate is set to 0.005, and the training is iterated 50 rounds.

7. A method for fault diagnosis based on the rotating machinery fault diagnosis device according to any one of claims 1-6, characterized in that, It includes: The data acquisition module obtains the vibration signal data to be detected of the rotating mechanical equipment; The data preprocessing module uses the sliding window segmentation function to divide the vibration signal data to be detected and performs normalization preprocessing; The data transformation module uses continuous wavelet transform to convert the preprocessed vibration signal data to be detected into time-frequency images and corresponding time-frequency features; The feature extraction module includes a parallel time-frequency feature extractor and a multi-scale feature extractor; the time-frequency feature extractor analyzes and processes the time-frequency features to extract numerical branch features, and the multi-scale feature extractor analyzes and processes the time-frequency images to extract multi-scale visual branch features; the time-frequency feature extractor includes several convolutional pooling modules; each convolutional pooling module includes an upsampling layer, a convolutional layer, a BatchNorm2d layer, a PReLU activation layer, a Dropout layer, and a one-dimensional MaxPool layer; the convolutional layer extracts the local and global features of the signal through a 3×3 convolutional kernel; an attention module is introduced after each convolutional pooling module; the multi-scale feature extractor includes several multi-convolutional kernel modules and a dimensionality reduction module; the multi-convolutional kernel module includes multiple convolutional kernels of different sizes for extracting the detailed features, medium-scale spatial features, and global range features of the time-frequency image; The dimensionality reduction module includes a global average pooling layer and a fully connected layer to perform dimensionality reduction processing on the output result of the multi-convolutional kernel module; The dynamic weighted fusion module weights and fuses the numerical branch features and the multi-scale visual branch features through dynamically generated weights and outputs the fault diagnosis result; the dynamic weighted fusion module includes a feature splicing unit, a weight generation unit, and a weighted fusion unit; the feature splicing unit splices the feature vectors output by the time-frequency feature extractor and the multi-scale feature extractor in the channel dimension to form a joint feature representation; the weight generation unit processes the joint feature through a fully connected network to generate the weights of each branch, and normalizes the weights using the Softmax function; the weighted fusion unit sums the feature vectors of the two branches according to the generated weights to generate a fused feature representation and outputs the fault diagnosis result.

Images