Bearing diagnosis method and system based on adaptive weighted double branch, electronic device and storage medium

Abstract

The application discloses a bearing diagnosis method and system based on adaptive weighted double branches, an electronic device and a storage medium, which comprises the following steps: 1, collecting single-channel original vibration signals of a rolling bearing, and preprocessing the original vibration signals; 2, constructing a double-branch parallel convolution network, wherein the first branch adopts a first receptive field to extract local transient impact features of the signals, and the second branch adopts a second receptive field to extract periodic features and global environmental features of the signals; 3, generating a first weight vector of a first branch output feature vector and a second weight vector of a second branch output feature vector by using an adaptive gating mechanism; 4, performing element-level weighted fusion on the output feature vectors of the first branch and the second branch based on the optimized weight vectors to obtain a fusion feature vector; and 5, inputting the fusion feature vector into a classifier to output a fault type classification result of the rolling bearing.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis technology, and in particular to a bearing diagnosis method, system, electronic device, and storage medium based on adaptive weighted bibranch. Background Technology

[0002] Rolling bearings, as core supporting components of rotating machinery, play a crucial role in reducing friction and supporting rotational motion. However, under continuous mechanical wear and operating stress, rolling bearings are prone to failures such as spalling, pitting, cracking, or indentation. These failures can lead to abnormal equipment operation or even unexpected shutdowns, resulting in production losses. Therefore, accurate detection and diagnosis of bearing failures are of great significance for implementing predictive maintenance, reducing downtime, and lowering maintenance costs.

[0003] Traditional fault diagnosis methods are mainly divided into signal processing-based methods and deep learning-based methods. Signal processing-based methods typically rely on expert knowledge and human experience, and the analysis results are easily affected by subjective factors, limiting their consistency and generalization ability. Deep learning-based methods can automatically extract features and have strong adaptability, driving the transformation of fault diagnosis from traditional methods to data-driven methods. However, in real-world industrial scenarios, strong noise interference (such as environmental noise, electromagnetic interference, and mechanically coupled vibration) often masks the subtle features of early faults, and the signal-to-noise ratio of vibration signals is usually below -5dB or even completely submerged in background noise. This leads to problems such as high misdiagnosis rates and difficulties in feature extraction for traditional diagnostic methods.

[0004] To improve diagnostic performance, multi-sensor signal fusion strategies are widely adopted to enhance feature representation. However, while enriching the feature set, multi-source information also introduces significant redundancy, interfering with the model's focus on key features. More critically, in many practical industrial equipment (such as marine gearboxes and aero engines), the compact mechanical structure, high airtightness requirements, or harsh operating environments severely limit sensor installation space and wiring channels, often allowing only single-channel or a very small number of sensors to be deployed, resulting in extremely limited available monitoring information. Taking marine gearboxes as an example, their narrow internal space and the need for watertight external wiring typically only allow for the placement of single-channel accelerometers on the gearbox surface, making it difficult to construct a multi-sensor fusion diagnostic system. This limitation in sensor placement makes it difficult to implement diagnostic strategies relying on multi-sensor fusion. In recent years, methods based on single-channel signal analysis have shown good performance in specific scenarios, but traditional convolutional neural networks are limited by fixed receptive field designs, making it difficult to capture multi-scale features simultaneously: small receptive fields can accurately capture local details but struggle to model long-range dependencies, while large receptive fields can cover a wider range but easily obscure subtle features. The limitations of single-scale perception are particularly pronounced due to the dual constraints of single-channel information input caused by sensor limitations and strong background noise, which restricts the model's ability to represent the overall signal and thus affects diagnostic accuracy.

[0005] Therefore, there is an urgent need for a rolling bearing fault diagnosis method and system that can effectively extract fault features in a high-noise environment using only a single-channel vibration signal, under conditions where sensor placement is limited, and that possesses strong robustness and high diagnostic accuracy. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a bearing diagnostic method, system, electronic device, and storage medium based on adaptive weighted dual-branch, which effectively solves the problems of low signal-to-noise ratio of single-channel vibration signals and difficulty in extracting multi-scale fault features in industrial scenarios with limited sensor placement (such as marine gearboxes).

[0007] To achieve the aforementioned objectives of the invention, the technical solution adopted to solve its technical problems is as follows: In a first aspect, the present invention provides a bearing diagnosis method based on adaptive weighted dual-branch, comprising the following steps: Step 1: Acquire the raw single-channel vibration signal of the rolling bearing, and preprocess the raw vibration signal to construct a sample set; Step 2: Construct a dual-branch parallel convolutional network, wherein the first branch uses the first receptive field to extract the local transient impact features of the signal, and the second branch uses the second receptive field to extract the periodic features and global environmental features of the signal. The first receptive field is smaller than the second receptive field. Step 3: Use an adaptive gating mechanism to generate the first weight vector of the first branch output feature vector and the second weight vector of the second branch output feature vector; Step 4: Based on the optimized weight vector, perform element-wise weighted fusion of the output feature vectors of the first branch and the second branch to obtain the fused feature vector; Step 5: Input the fused feature vector into the classifier and output the failure type classification result of the rolling bearing.

[0008] Furthermore, in step 1, the preprocessing includes: segmenting the original vibration signal using a sliding window overlapping sampling method, and constructing training sample sets with different signal-to-noise ratios by adding noise.

[0009] Furthermore, in step 2, the network structures of both the first branch and the second branch include a first convolutional layer, a second convolutional layer, a convolutional block attention module, and a multi-head self-attention module connected in sequence, wherein: The convolutional block attention module is set after the second convolutional layer and is used to enhance key features sequentially along the channel dimension and spatial dimension; The multi-head self-attention module is positioned after the convolutional block attention module and is used to capture global dependencies between features.

[0010] Furthermore, the convolutional block attention module includes a channel attention submodule and a spatial attention submodule, wherein: The channel attention submodule aggregates spatial information through average pooling and max pooling, and then generates a channel attention map through a shared multilayer perceptron. The spatial attention submodule performs average pooling and max pooling along the channel dimension, and then generates a spatial attention map through a convolutional layer.

[0011] Furthermore, the multi-head self-attention module includes multiple parallel self-attention heads, each of which generates a query matrix through a linear transformation. Key matrix Sum matrix Calculate the attention score: , in, is the scaling factor, and T is the transpose of the matrix; The outputs of multiple self-attention heads are concatenated and then linearly transformed through a weight matrix to obtain the multi-head self-attention output.

[0012] Furthermore, in step 3, the adaptive gating mechanism includes a gating unit and a robustness optimization module, wherein: The gated unit includes a fully connected layer and a sigmoid activation function. It performs a non-linear mapping on the branch output feature vector to generate an initial weight vector, the calculation formula of which is: , , Where σ is the Sigmoid function, and Dense represents a fully connected layer. and This represents the feature vector output in a two-branch network. and These are the initial weight vectors for the two branches in a two-branch network; The robustness optimization module obtains multiple sets of weight vectors through repeated experiments, removes outliers using a clustering algorithm, and optimizes the mean of the removed weight vectors to obtain the optimized first and second weights.

[0013] Preferably, the clustering algorithm employs at least one of DBSCAN, K-means, K-medoids, GMM, ISODATA, or MeanShift.

[0014] Furthermore, in step 4, the calculation formula for the element-level weighted fusion is: , in, Represents the Hadamard product. and Y is represented as the optimized weight vector, and Y is the fused feature vector. The optimized weight vector is multiplied element-wise with the feature vector of the corresponding branch and then added together to obtain the weighted fusion feature.

[0015] Secondly, the present invention provides a bearing diagnostic system based on adaptive weighted dual-branch, comprising: The signal acquisition module is used to acquire the single-channel raw vibration signal of the rolling bearing; The signal preprocessing module is used to preprocess the original vibration signal and construct a noisy sample set; The network construction and training module is used to construct and train a dual-branch parallel convolutional network, wherein the dual-branch parallel convolutional network includes a first branch and a second branch in parallel, and the receptive field of the first branch is smaller than that of the second branch. The weight generation module is used to generate the first weight of the first branch output feature vector and the second weight of the second branch output feature vector using an adaptive gating mechanism. The feature fusion module is used to perform weighted fusion of the output feature vectors of the first branch and the second branch based on the first weight and the second weight to obtain a fused feature vector; The fault classification output module is used to input the fused feature vector into the classifier and output the fault type classification result of the rolling bearing.

[0016] Furthermore, the adaptive gating mechanism includes a gating unit and a weight optimization module; the gating unit includes a fully connected layer and a sigmoid activation function to generate an initial weight vector; the weight optimization module uses a clustering algorithm to remove outliers and optimize the mean of multiple sets of weight vectors obtained from repeated experiments to obtain an optimized weight vector.

[0017] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method.

[0018] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0019] By employing the above technical solutions, this invention has the following advantages and positive effects compared with the prior art: 1. Lightweight and efficient for sensor-constrained scenarios: The method of this invention is specifically designed for industrial scenarios (such as marine gearboxes) where sensor deployment is limited. By constructing a dual-branch parallel convolutional network, fault features can be effectively extracted using only a single-channel signal without relying on multiple sensor inputs. This not only overcomes the monitoring challenges caused by space constraints but also significantly reduces hardware costs and system complexity, demonstrating extremely high engineering application value.

[0020] 2. Multi-scale feature extraction capability in constrained scenarios: To address the challenge of single-channel signal information having a single dimension, the dual-branch structure adopted in this invention captures local transient impact features and periodic global environmental features respectively through receptive fields of different sizes. This enables the mining of multi-scale feature representations from local to global in limited input information, solving the problem that traditional convolutional networks with fixed receptive fields cannot take into account multi-scale information.

[0021] 3. Feature enhancement capability in strong noise and weak signal environments: It integrates convolutional block attention and multi-head self-attention mechanism. Convolutional block attention enhances key features sequentially along the channel and spatial dimensions, while multi-head self-attention captures the global dependencies between features. The synergistic effect of the two improves the model's ability to identify weak fault features under harsh conditions such as limited sensors and low signal-to-noise ratio.

[0022] 4. Robust adaptive fusion mechanism: The dual-branch feature fusion weights are generated through an adaptive gating mechanism. Combined with clustering algorithms (such as K-means clustering) and mean optimization to remove outliers, the optimized weight vector is obtained. This avoids the limitations of manually setting weights. Especially in the single-channel analysis path, it ensures the stability of feature fusion and classification accuracy.

[0023] 5. Excellent noise resistance and generalization ability: Experimental results show that the present invention can still maintain a diagnostic accuracy of over 99% in a strong noise environment of -30dB. It performs well on multiple public datasets such as CWRU, HUST, and PU. In particular, it still has strong robustness and generalization ability under the limited condition of providing only a single-channel vibration signal, and is suitable for rolling bearing fault diagnosis in complex industrial environments. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A schematic flowchart of the bearing diagnosis method based on adaptive weighted dual branches provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the experimental platform system used in the embodiments of the present invention; Figure 3 This is a flowchart illustrating the partitioning of the multi-condition rolling bearing fault dataset in an embodiment of the present invention. Figure 4 This is a schematic diagram of the structure of a dual-branch parallel convolutional network in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the adaptive gating mechanism and weight optimization in an embodiment of the present invention; Figure 6 This is a schematic diagram of element-level weighted fusion in an embodiment of the present invention; Figure 7 This is a detailed configuration diagram of the dual-branch network structure in an embodiment of the present invention; Figure 8 This is a comparison chart of model performance under different training cycles in the embodiments of the present invention; Figure 9 This is a performance comparison chart of different weight fusion strategies under noise-free conditions in the embodiments of the present invention; Figure 10 This is a performance comparison chart of different weighted fusion strategies at a signal-to-noise ratio of -10dB in the embodiments of the present invention; Figure 11This is a performance comparison chart of different weighted fusion strategies at a signal-to-noise ratio of -15dB in the embodiments of the present invention; Figure 12 This is a performance comparison chart of different weighted fusion strategies at a signal-to-noise ratio of -20dB in the embodiments of the present invention; Figure 13 This is a performance comparison chart of different weighted fusion strategies at a signal-to-noise ratio of -25dB in the embodiments of the present invention; Figure 14 This is a performance comparison chart of different weighted fusion strategies at a signal-to-noise ratio of -30dB in the embodiments of the present invention; Figure 15 This is a performance comparison chart of different signal channel positions in embodiments of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0026] Example 1: Fault Diagnosis Method Reference Figures 1-6 As shown, this embodiment provides a bearing diagnostic method based on adaptive weighted bibranching, used to identify and classify inner ring faults, outer ring faults, and rolling element faults in marine gearbox structures under conditions of limited sensor placement. The method includes the following steps: Step 1: Acquire the raw single-channel vibration signal of the rolling bearing, and preprocess the raw vibration signal to construct a sample set.

[0027] Specifically, in step 1, the preprocessing includes: segmenting the original vibration signal using a sliding window overlapping sampling method, and constructing training sample sets with different signal-to-noise ratios by adding noise, with the signal-to-noise ratio range being -30dB to 0dB.

[0028] Reference Figure 2To further illustrate the application scenarios and data sources of this invention, the experimental system is described in detail below. The experimental platform includes components such as a drive motor, gearbox, eddy current brake, torque-speed sensor, and pump. The drive motor is connected to the input shaft of the gearbox via a coupling, and the output shaft of the gearbox is connected to the eddy current brake via the torque-speed sensor to apply load torque. To comprehensively acquire the vibration response characteristics of the rolling bearing during operation, 12 vibration acceleration sensor channels were deployed at different locations on the experimental platform. For example, the specific arrangement is as follows: Three single-axis accelerometers (models Y2110033, Y2110038, and Y2110093) were installed on the top surface of the gearbox to measure the vibration in the X direction at that location; a triaxial accelerometer (model SN5808445) was installed on the left side of the gearbox to simultaneously measure the vibration in the X, Y, and Z directions at that location; a triaxial accelerometer (model SN5808446) was installed on the right side of the gearbox to simultaneously measure the vibration in the X, Y, and Z directions; and a triaxial accelerometer (model SN6251144) was installed on the base of the experimental platform to measure the overall structural vibration response in the X, Y, and Z directions at the base. All vibration acceleration signals were synchronously acquired and recorded using a DH5922D dynamic signal acquisition system. The acquisition system supports multi-channel parallel acquisition. In this experiment, the sampling frequency was set to 10kHz to fully capture the high-frequency transient impact components caused by rolling bearing failure. During the acquisition process, the signal is conditioned by a charge amplifier before being input into the acquisition system and finally stored in the computer as a voltage signal.

[0029] This embodiment focuses on pitting bearing failures, categorizing them into four states based on fault location: normal state, rolling element failure, inner ring failure, and outer ring failure. Each failure type is further divided into three levels according to wear degree (0.1mm, 0.3mm, 0.5mm), resulting in a total of 10 failure types. To simulate industrial operating conditions, three theoretical speeds (570r / min, 1140r / min, 1710r / min) and two load levels (0A, 0.5A) are set, forming six different operating condition combinations, which are evenly distributed in the sample.

[0030] Regarding sample generation, to fully capture fault characteristics, the sample length is determined based on the minimum theoretical rotational speed (570 r / min) and sampling frequency (10 kHz), calculated using the following formula: Where T is the number of rotation cycles. Here, r is the sampling frequency, and r is the rotational speed (r / min). In this embodiment, the sample length is 2560 points to ensure that the samples cover at least two bearing rotation cycles. A sliding window overlapping sampling method is used for data augmentation, with the step size set to 1 / 4 of the sample length. To simulate an industrial noise environment, a Gaussian noise injection strategy is used to generate training samples with different signal-to-noise ratios (-10dB, -15dB, -20dB, -25dB, -30dB) to verify the robustness of the model under extreme noise conditions.

[0031] Regarding sample partitioning, the total sample size was calculated as “number of samples for each working condition (200) × number of fault types (10) × number of speeds (3) × number of loads (2)”, totaling 12,000 samples. The samples were randomly partitioned without replacement into a training set (3,600 samples), a validation set (1,200 samples), and a test set (7,200 samples) in a ratio of 3:1:6.

[0032] Step 2: Construct a dual-branch parallel convolutional network. The network structure uses two parallel branches. The first branch (branch 1) uses the first receptive field to extract local transient impact features of the signal, such as fault pulses. The second branch (branch 2) uses the second receptive field to extract periodic features and global environmental features of the signal. The first receptive field is smaller than the second receptive field. Figure 4 As shown.

[0033] Reference Figure 7 Both branches use the same network structure, differing only in the kernel size. Each branch consists of: a first convolutional layer (101, 201), a first batch normalization layer (102, 202), a first activation function layer (103, 203), a second convolutional layer (104, 204), a second batch normalization layer (105, 205), and a second activation function layer (106, 206). For branch 1, the first convolutional layer can use a kernel size of 3×1 to 7×1, and the second convolutional layer can use 7×1 to 11×1. In branch 2, the first convolutional layer can use 7×1 to 11×1, with a stride of 2-4. All of these ranges can effectively separate local details from the global periodicity; the specific size can be adjusted according to the sampling frequency and fault feature frequency.

[0034] Specifically, in step 2, the network structures of both the first and second branches include a first convolutional layer, a second convolutional layer, a convolutional block attention module, and a multi-head self-attention module connected in sequence, wherein: The convolutional block attention module is set after the second convolutional layer and is used to enhance key features sequentially along the channel dimension and spatial dimension; The multi-head self-attention module is positioned after the convolutional block attention module and is used to capture global dependencies between features.

[0035] Furthermore, the convolutional block attention module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule aggregates spatial information through average pooling and max pooling to generate two spatial attention maps. After passing through a shared multilayer perceptron, element-wise summation and a sigmoid function are used to generate a channel attention map. The spatial attention submodule performs average pooling and max pooling along the channel dimension, concatenates the results, and then passes them through a convolutional layer to generate a spatial attention map. The channel attention map is multiplied by the input feature map and used as the input to the spatial attention submodule. The spatial attention map is multiplied by an intermediate feature map to obtain the final output.

[0036] Furthermore, the multi-head self-attention module includes multiple parallel self-attention heads, each of which generates a query matrix through a linear transformation. Key matrix Sum matrix Calculate the attention score: , in, is the scaling factor, and T is the transpose of the matrix; The outputs of multiple self-attention heads are concatenated and then linearly transformed through a weight matrix to obtain the multi-head self-attention output.

[0037] Step 3: Use an adaptive gating mechanism to generate the first weight vector of the first branch output feature vector and the second weight vector of the second branch output feature vector, such as... Figure 5 .

[0038] Specifically, in step 3, the adaptive gating mechanism includes a gating unit and a robustness optimization module, wherein: The gated unit includes a fully connected layer and a sigmoid activation function. It performs a non-linear mapping on the branch output feature vector to generate an initial weight vector, the calculation formula of which is: , , Where Dense represents a fully connected layer, and These are the initial weight vectors for the two branches in a two-branch network. and σ represents the feature vector output in the dual-branch network. σ is the Sigmoid function, which constrains the weights within the (0,1) interval. This not only provides a probabilistic interpretation of the weight vector but also ensures gradient smoothing to facilitate network training, thus forming an efficient "soft attention" mechanism.

[0039] The robustness optimization module obtains multiple sets of weight vectors through repeated experiments, removes outliers using a clustering algorithm, and optimizes the mean of the removed weight vectors to obtain the optimized first and second weights. In this embodiment, the clustering algorithm can be at least one of DBSCAN (Density-Based Spatial Clustering of Applications with Noise), K-means, K-medoids, GMM (Gaussian Mixture Model), ISODATA (Iterative Self-organizing Data Analysis Technique Algorithm), or Mean Shift.

[0040] Because the training of the gating unit is affected by random initialization and the order of mini-batch data, some elements in the weight vector may converge to local optima, forming outliers that deviate from the main distribution. These outliers cause the fused weights to fail to accurately reflect the importance of the bi-branch features. In a preferred embodiment, K-means clustering with K=2 is used to naturally divide the weight values ​​into 'credible intervals' and 'outlier intervals'. After removing the latter, the average is taken to obtain stable weights that are insensitive to random fluctuations.

[0041] In this embodiment, specifically, multiple repeated experiments (e.g., 10 times) are conducted, and the weight vector obtained from each experiment is collected. For each element in the weight vector, its values ​​in each experiment are paired to form a data point distribution in a two-dimensional space. The K-means clustering algorithm (K=2) is used to identify the main clusters and outliers, and data points far from the main clusters are identified as outliers and removed. Finally, the arithmetic mean of the remaining data points after removing outliers is calculated as the optimized estimate of the weight element. The above operations are performed sequentially on all weight elements to finally obtain the optimized weight vector. and This optimization module is executed offline during the training phase. The optimized weights are saved as fixed parameters and used directly during the inference phase without adding extra computational overhead to online diagnostics.

[0042] Step 4: Based on the optimized weight vector, perform element-wise weighted fusion of the output feature vectors of the first and second branches to obtain the fused feature vector, such as... Figure 6 .

[0043] Specifically, in step 4, the calculation formula for the element-level weighted fusion is as follows: , in, Represents the Hadamard product. and Y is represented as the optimized weight vector, and Y is the fused feature vector. The optimized weight vector is multiplied element-wise with the feature vector of the corresponding branch and then added together to obtain the weighted fusion feature.

[0044] In this embodiment, this element-level weighting mechanism allows each weight value to serve as the attention coefficient for the corresponding feature, autonomously enhancing information-rich features, suppressing redundant features and noise interference, and achieving fine-grained feature fusion.

[0045] Step 5: Input the fused feature vector into the classifier and output the failure type classification result of the rolling bearing.

[0046] The fused feature vector Y is sequentially input into a batch normalization layer, an activation function layer, a fully connected layer, and a softmax layer to output the probability distribution of rolling bearing fault types, ultimately completing fault classification. Fault types include normal state, rolling element fault, inner ring fault, and outer ring fault. Each fault type is divided into three levels according to the degree of wear (0.1mm, 0.3mm, 0.5mm), for a total of 10 fault types.

[0047] The aforementioned dual-branch parallel convolutional network was trained with the following parameters: the Adam optimizer was used with an initial learning rate of 0.001, employing a piecewise constant learning rate decay strategy, decreasing the learning rate by a factor of 0.97 every 5 training epochs to dynamically adjust the learning rate and accelerate convergence. The mini-batch size was set to 50, and L2 regularization (coefficient of 0.0001) was used to prevent overfitting. At the beginning of each training epoch, the training data was randomly shuffled to enhance the randomness and stability of the training process.

[0048] Reference Figure 8 To determine the optimal number of training epochs, a comparative experiment was conducted under a -30dB strong noise environment. The experimental results show that in the early training phase (10-40 epochs), the model accuracy rapidly increased from approximately 93% to approximately 98.63%; in the middle training phase (40-70 epochs), the average accuracy further increased to approximately 99.2%, and the model output tended to stabilize; after 70 epochs, the model entered a convergence plateau, with the accuracy stabilizing above 99%. Considering accuracy, stability, and training efficiency, 70 epochs were chosen as the training stopping point.

[0049] Reference Figures 9-14Nine fixed weight ratios (from 10%:90% to 90%:10%) were set up for comparison with the improved adaptive gating mechanism proposed in this invention. Experimental results show that the fixed weight strategy exhibits large fluctuations and insufficient stability at different signal-to-noise ratio levels; while the method of this invention maintains a median accuracy of over 99% at all signal-to-noise ratio levels, with a concentrated box plot distribution, demonstrating excellent repeatability and stability, especially at extremely low signal-to-noise ratios of -30dB.

[0050] Reference Figure 15 A systematic analysis of multiple channels on the experimental platform was conducted. Experimental results show that, except for channels 2 and 3, the remaining channels exhibit stable performance, with median accuracy concentrated in the 97%-100% range, and compact box plots with low dispersion. Channel 2, being furthest from the fault source, suffers the most severe signal attenuation and interference, resulting in the lowest performance. Channels 10, 11, and 12, located on the base, capture the overall structural response, exhibiting weaker fault characteristic components and relatively lower performance. This experiment demonstrates that the method of this invention can maintain high-precision diagnosis even when using only single-channel signals that are far from the fault source and have poor signal quality, fully proving its engineering practicality under conditions of limited sensor placement.

[0051] Example 2: Fault Diagnosis System This embodiment provides a bearing diagnostic system based on adaptive weighted dual-branching, including: The signal acquisition module is used to acquire the single-channel raw vibration signal of the rolling bearing; The signal preprocessing module is used to preprocess the original vibration signal and construct a noisy sample set; The network construction and training module is used to construct and train a dual-branch parallel convolutional network, wherein the dual-branch parallel convolutional network includes a first branch and a second branch in parallel, and the receptive field of the first branch is smaller than that of the second branch. The weight generation module is used to generate the first weight of the first branch output feature vector and the second weight of the second branch output feature vector using an adaptive gating mechanism. The feature fusion module is used to perform weighted fusion of the output feature vectors of the first branch and the second branch based on the first weight and the second weight to obtain a fused feature vector; The fault classification output module is used to input the fused feature vector into the classifier and output the fault type classification result of the rolling bearing.

[0052] The above modules are implemented by the processor executing computer programs stored in memory, or integrated into embedded systems, digital signal processors (DSPs) or field-programmable gate arrays (FPGAs) to achieve offline training and online real-time diagnostics.

[0053] Furthermore, the adaptive gating mechanism includes a gating unit and a weight optimization module; the gating unit includes a fully connected layer and a sigmoid activation function to generate an initial weight vector; the weight optimization module uses a clustering algorithm to remove outliers and optimize the mean of multiple sets of weight vectors obtained from repeated experiments to obtain an optimized weight vector.

[0054] Example 3: Electronic Equipment

[0055] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method in Embodiment 1 above.

[0056] Example 4: Computer-readable storage medium This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1. The storage medium includes, but is not limited to, non-transitory storage media such as ROM, RAM, hard disk, solid-state drive, USB flash drive, and CD-ROM.

[0057] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A bearing diagnostic method based on adaptive weighted bibranching, characterized in that, Includes the following steps: Step 1: Acquire the raw single-channel vibration signal of the rolling bearing, and preprocess the raw vibration signal to construct a sample set; Step 2: Construct a dual-branch parallel convolutional network, wherein the first branch uses the first receptive field to extract the local transient impact features of the signal, and the second branch uses the second receptive field to extract the periodic features and global environmental features of the signal. The first receptive field is smaller than the second receptive field. Step 3: Use an adaptive gating mechanism to generate the first weight vector of the first branch output feature vector and the second weight vector of the second branch output feature vector; In step 3, the adaptive gating mechanism includes a gating unit and a robustness optimization module, wherein: The gated unit includes a fully connected layer and a sigmoid activation function. It performs a non-linear mapping on the branch output feature vector to generate an initial weight vector, the calculation formula of which is: (Official 2) (Official 3) Where σ is the Sigmoid function, and Dense represents a fully connected layer. and This represents the feature vector output in a two-branch network. and These are the initial weight vectors for the two branches in a two-branch network; The robust optimization module obtains multiple sets of weight vectors through repeated experiments, removes outliers using a clustering algorithm, and optimizes the mean of the removed weight vectors to obtain the optimized first and second weights. Step 4: Based on the optimized weight vector, perform element-wise weighted fusion of the output feature vectors of the first branch and the second branch to obtain the fused feature vector; Step 5: Input the fused feature vector into the classifier and output the failure type classification result of the rolling bearing.

2. The bearing diagnosis method based on adaptive weighted bi-branching according to claim 1, characterized in that, In step 1, the preprocessing includes: segmenting the original vibration signal using a sliding window overlapping sampling method, and constructing training sample sets with different signal-to-noise ratios by adding noise.

3. The bearing diagnosis method based on adaptive weighted bi-branching according to claim 1, characterized in that, In step 2, the network structures of both the first and second branches include a first convolutional layer, a second convolutional layer, a convolutional block attention module, and a multi-head self-attention module connected in sequence, wherein: The convolutional block attention module is set after the second convolutional layer and is used to enhance key features sequentially along the channel dimension and spatial dimension; The multi-head self-attention module is positioned after the convolutional block attention module and is used to capture global dependencies between features.

4. The bearing diagnosis method based on adaptive weighted bi-branching according to claim 3, characterized in that, The convolutional block attention module includes a channel attention submodule and a spatial attention submodule, wherein: The channel attention submodule aggregates spatial information through average pooling and max pooling, and then generates a channel attention map through a shared multilayer perceptron. The spatial attention submodule performs average pooling and max pooling along the channel dimension, and then generates a spatial attention map through a convolutional layer.

5. The bearing diagnosis method based on adaptive weighted bi-branching according to claim 3, characterized in that, The multi-head self-attention module includes multiple parallel self-attention heads, each of which generates a query matrix through a linear transformation. Key matrix Sum matrix Calculate the attention score: (Official 1) in, is the scaling factor, and T is the transpose of the matrix; The outputs of multiple self-attention heads are concatenated and then linearly transformed through a weight matrix to obtain the multi-head self-attention output.

6. The bearing diagnosis method based on adaptive weighted bi-branching according to claim 1, characterized in that, The clustering algorithm used is at least one of DBSCAN, K-means, K-medoids, GMM, ISODATA, or Mean Shift.

7. The bearing diagnosis method based on adaptive weighted bi-branching according to claim 1, characterized in that, In step 4, the calculation formula for the element-level weighted fusion is as follows: (Official 4) in, Represents the Hadamard product. and Y is represented as the optimized weight vector, and Y is the fused feature vector. The optimized weight vector is multiplied element-wise with the feature vector of the corresponding branch and then added together to obtain the weighted fusion feature.

8. A bearing diagnostic system based on adaptive weighted bi-branching, characterized in that, The bearing diagnostic method described in claim 1 is used for diagnosis, including: The signal acquisition module is used to acquire the single-channel raw vibration signal of the rolling bearing; The signal preprocessing module is used to preprocess the original vibration signal and construct a noisy sample set; The network construction and training module is used to construct and train a dual-branch parallel convolutional network, wherein the dual-branch parallel convolutional network includes a first branch and a second branch in parallel, and the receptive field of the first branch is smaller than that of the second branch. The weight generation module is used to generate the first weight of the first branch output feature vector and the second weight of the second branch output feature vector using an adaptive gating mechanism. The feature fusion module is used to perform weighted fusion of the output feature vectors of the first branch and the second branch based on the first weight and the second weight to obtain a fused feature vector; The fault classification output module is used to input the fused feature vector into the classifier and output the fault type classification result of the rolling bearing.

9. The bearing diagnostic system based on adaptive weighted dual-branch as described in claim 8, characterized in that, The adaptive gating mechanism includes a gating unit and a weight optimization module; the gating unit includes a fully connected layer and a sigmoid activation function to generate an initial weight vector; the weight optimization module uses a clustering algorithm to remove outliers and optimize the mean of multiple sets of weight vectors obtained from repeated experiments to obtain an optimized weight vector.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

Images