A bearing fault diagnosis method based on multi-sensor multi-scale feature fusion

Abstract

A bearing fault diagnosis method based on multi-sensor, multi-scale feature fusion is proposed. The method involves collecting current and vibration signals to create a dataset, which is then augmented and divided into training, validation, and test sets. A multi-scale feature fusion network model is constructed. The signal data in the training set is transformed using Fast Fourier Transform (FFT) to extract features. An attention mechanism is then introduced to adjust the weights. The fused features are processed alternately using self-calibrating convolutions and dilated convolutions, followed by pooling, feature flattening, and fully connected layers. Finally, a classifier outputs the fault classification result. The validation set is input into the model to fine-tune its parameters. Finally, the test set is used for testing, and the results are output. By using multi-source data such as vibration and current signals as input, and leveraging the ability of neural networks to discover hidden patterns in training samples, this method aims to extract multi-scale features from multi-source data and fuse these features to complete the intelligent fault diagnosis process for bearings.

Description

Technical Field

[0001] This invention relates to the field of intelligent bearing fault diagnosis methods, specifically a bearing fault diagnosis method based on multi-sensor multi-scale feature fusion. Background Technology

[0002] In modern industry, bearings play a crucial role as key components in the production process, and their operational stability and reliability directly affect production efficiency and equipment lifespan. Bearings are widely used in numerous fields such as power generation, manufacturing, and transportation, including generators, wind turbines, and water pumps. However, due to their unique working environment and high-frequency movement, these devices are susceptible to various external and internal factors, such as vibration, friction, and wear, leading to potential failure risks. Therefore, timely and accurate fault diagnosis of bearings is essential, becoming key to ensuring production efficiency, extending equipment lifespan, and reducing maintenance costs.

[0003] Bearing fault diagnosis has always been a complex and challenging task. Firstly, traditional fault diagnosis processes primarily focus on the analysis of single signal sources, which has significant limitations. These include frequency domain analysis and time domain analysis based on vibration signals, as well as current spectrum analysis based on current signals. Secondly, the use of a single signal source cannot comprehensively and accurately reflect the operating status of the equipment, especially when faced with multiple complex fault modes. Thirdly, due to the complex working environment of mechanical equipment and significant noise interference, the acquired signals often contain a large amount of irrelevant information, leading to unreliable analysis results. While these methods have improved the efficiency of fault diagnosis to some extent, they still have shortcomings in addressing changes in equipment operating conditions, identifying faults at different frequencies, resolving multi-signal fusion, and identifying fault modes in complex environments.

[0004] Secondly, conventional fault diagnosis methods mainly involve three stages: 1. Initializing the signal by denoising and decomposing it; among numerous research methods, signal processing methods are the most common, such as Fourier transform and empirical mode decomposition; 2. Using feature extraction methods to obtain the time-domain and frequency-domain features of the preprocessed signal. Feature extraction methods include wavelet transform, synchronous extraction, and empirical wavelet transform. These methods can further filter out useless features in the signal, making the required fault features more prominent; 3. Inputting the finally extracted features into a machine learning classifier for training, thereby classifying the fault type. However, these methods mainly rely on advanced signal processing techniques and require rich engineering experience and professional knowledge in feature extraction and selection. Furthermore, they still suffer from a separation between fault feature extraction and diagnostic identification. In addition, many traditional machine learning models have relatively simple structures, limiting the classifier's ability to learn complex nonlinear relationships in fault diagnosis problems.

[0005] In the context of today's industrial intelligent and digital transformation, advanced technologies such as "Internet+", big data, artificial intelligence, and deep learning are bringing unprecedented changes to the industrial sector. The integration of these technologies provides entirely new ideas and solutions for bearing fault diagnosis, driving the improvement of industrial intelligence. With the increasing national focus on industry, "Internet+Industry" has become a crucial project, and the rise and development of theoretical technologies such as big data, artificial intelligence, and deep learning have provided new research approaches for modern bearing fault diagnosis. Artificial intelligence plays an increasingly important role in bearing fault diagnosis, and deep learning, as a branch of artificial intelligence, can better handle large-scale unstructured data through neural network structures. However, most existing traditional deep learning models have relatively simple structures, which limits the classifier's ability to handle complex nonlinear relationships when dealing with fault diagnosis problems. In the complex operating environment of equipment, how to utilize multi-source data for feature fusion to improve the accuracy of fault diagnosis is a primary problem that urgently needs to be solved in intelligent bearing fault diagnosis. Summary of the Invention

[0006] The purpose of this invention is to propose a bearing fault diagnosis method based on multi-sensor, multi-scale feature fusion. This scheme uses multi-source data such as vibration signals and current signals as inputs, and employs a deep network model to achieve feature fusion of different types of signals. At the same time, an attention mechanism network is introduced to enhance the connection between key features and improve the classification performance of the model. Furthermore, this scheme comprehensively considers the information from different signal sources and achieves end-to-end learning of multi-scale features through deep learning, making the diagnosis of diverse and complex faults more comprehensive and accurate.

[0007] The technical solution adopted in this invention is: a bearing fault diagnosis method based on multi-sensor multi-scale feature fusion, comprising the following steps: S1. Collect current and vibration signals of the bearing under operating conditions through sensors to create a dataset. Use the BEGAN network model to enhance the dataset samples and divide the enhanced samples into training set, validation set and test set. S2. Build a multi-scale feature fusion network model. Convert the current signal and vibration signal data in the training set into the frequency domain through fast Fourier transform, and then feed them into the multi-scale network model to extract the features of the training set. Introduce the extracted features into the attention mechanism to adjust the weights, and then fuse the features extracted by the two branch networks. S3. The fused features are processed alternately using self-correcting convolution and dilated convolution to improve the model's representation ability and generalization performance. After pooling, feature flattening and full connection, the fault classification result is finally output through the classifier. S4. Input the validation set into the multi-scale network model trained in step S2, evaluate the model's performance, and fine-tune the model's parameters based on the validation results. S5. Determine whether the number of training iterations M of the network model has reached the preset number of iterations N. If yes, proceed to the next step; otherwise, repeat step S2. S6. Input the test set into the network model validated in step S4 to test the performance of the validated network model. Finally, use the test set to test the trained model and output the results. As a preferred embodiment, in step S1, the BEGAN network model comprises two parts: a generator model and a discriminator model. The generator updates its own parameters through the loss function passed by the discriminator, and the discriminator adjusts its own parameters through the parameters passed by the generator. The two are trained and reach a balance through a game between them. As a preferred embodiment, the loss functions of the BEGAN network model for the generator and discriminator models are as follows: L D =L(x; θ) D )-L(G(z D ;θ G );θ D ) L G =L(G(z) D ;θ G );G D ) Where X is the real sample, θ D and θ G The parameters of the generator and discriminator are minimized by L. D and L G Update, z D It is a sample from z; The objective function of the BEGAN network model is: L D =L(x)-k t L(G(z D )) L G =L(G(z) G )) k t+1 =k t +β k (γL(x)-L(G(z G ))) Where: L(v)=|vD(v)| α ,and α∈{1,2}; k tThe value is between [0, 1], and it is used to regulate the use of L(G(z) G The strength of gradient descent at time k0 = 0, β k Let L(v) represent the proportional gain of k, and L(v) represent the reconstruction loss between data v and the data generated by the autoencoder D. D It represents the difference between the reconstruction loss of the real data x and the data x after passing through the autoencoder D, and the reconstruction loss of the data G(z) generated by the generative model and the data D(G(z)) after passing through D. As a preferred option, in step S1, the enhanced samples are divided into a training set, a validation set, and a test set in a ratio of 8:1:1. As a preferred embodiment, in step S2, the current signal and vibration signal data in the training set are converted into the frequency domain using Fast Fourier Transform (FFT), where the definition formula for FFT is: Where N is the time series length; W is the weight function of the DFT. As a preferred option, the specific process in step S2 is as follows: S2.1 The current signal and vibration signal after being transformed by Fast Fourier Transform are respectively input into the first convolutional layer. The input data will be convolved through this convolutional layer to extract the initial shallow features, and then the deep features will be extracted through multi-scale convolutional kernels. S2.2. The extracted multi-scale features are fused and introduced into the channel spatial attention mechanism to model and adjust channel correlation. S2.3 Perform grouped convolutions, and finally fuse the features extracted by the two branch networks. As a preferred embodiment, in step S2.1, firstly, a parallel convolutional filtering operation is performed using three sets of convolutional kernels of different scales, with kernel sizes of 1×3, 1×5, and 1×7 respectively. Then, the result from the previous layer is input again into the next layer of multi-scale convolutional kernels, where the kernel size is the same as the previous layer. The multi-scale convolution operation formula is: As a preferred option, in step S2.3, during the grouped convolution process, the convolution kernel is divided into multiple groups, and the convolution kernel of each group only processes a subset of the input feature map. The size of the convolution kernel of each group is 1×1. As a preferred approach, the specific steps of the self-correcting convolution in step S3 are as follows: S3.1, Input Features Figure X Given a size of C×H××W, it is split into two equal parts according to the number of channels, denoted as... S3.2. Split the convolution kernel K into four parts of the same dimension, each with a different function. Each part has a dimension of C / 2 × C / 2 × H / W, denoted as... S3.3 Process the original scale feature space by performing a K4 convolution on feature X2 to obtain Y2; S3.4 Process the self-correcting scale space by inputting feature X1 into two different resolution channels. When the width of the input feature is reduced by r times to reduce the resolution, average pooling is used to downsample feature X1, and then convolution and upsampling are performed on it. Self-correction is performed by calculating the output and the features extracted by K2. Finally, self-correcting feature Y1 is output by convolution with K3. S3.5. Merge feature Y1 and feature Y2 to output the final feature Y. As a preferred option, in step S3, the formula for calculating dilated convolution is: Where Y(·) represents the input feature with length B, W(i) represents the size of the convolution kernel, r represents the dilation rate, and Z(x) represents the output feature after dilated convolution. The beneficial effects of this invention are:

[0008] To address the shortcomings of existing technologies, this invention provides a bearing fault diagnosis method based on multi-sensor, multi-scale feature fusion. Utilizing modern computer technology, the invention analyzes vibration and current signals acquired by sensors, extracts multi-scale feature information from two different signal types, fuses and classifies them to achieve fault diagnosis. Simultaneously, a deep network model is employed to fuse features from different signal types, and an attention mechanism network is introduced to enhance the connections between key features and improve the model's classification performance. This method comprehensively considers information from different signal sources and achieves end-to-end learning of multi-scale features through deep learning, making the diagnosis of diverse and complex faults more comprehensive and accurate. This solution will provide key theoretical and technical support for the intelligent development of bearing fault diagnosis in rotating machinery, enabling comprehensive monitoring and accurate diagnosis of equipment status, providing more reliable equipment protection for industrial production, and promoting the intelligent and digital transformation of the manufacturing industry. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1A framework diagram for bearing fault diagnosis based on multi-sensor, multi-scale feature fusion; Figure 2 This is a framework diagram of the BEGAN network model; Figure 3 This is a flowchart of the generator model's processing flow. Figure 4 This is a flowchart of the discriminator model's processing flow. Figure 5 The diagram shows the original data signal and the generated data signal. Figure 6 The spectrum diagrams of the current signal and the vibration signal after FFT conversion; Figure 7 This is a schematic diagram of the network model parameters for the feature extraction and fusion module; Figure 8 This is a flowchart of the self-calibrating convolution module; Figure 9 A schematic diagram of the network model parameters for the fault classification module; Figure 10 The flowchart shows a bearing fault diagnosis method based on multi-sensor multi-scale feature fusion. Figure 11 A schematic diagram of the operating parameters used to verify the intelligent fault diagnosis performance of this solution; Figure 12 This is a schematic diagram comparing the accuracy of this solution with existing technologies under varying operating conditions. Detailed Implementation

[0010] The present invention will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments.

[0011] It should be noted that, unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0012] To more clearly describe the specific process of this bearing fault diagnosis method based on multi-sensor multi-scale feature fusion, the appendix is ​​provided. Figure 1-10 The description is as follows: like Figure 1 As shown, the system used in the rotating machinery fault diagnosis method based on multi-sensor multi-scale feature fusion includes a data augmentation module, a feature extraction and fusion module, and a fault classification module. The composition and main function of each module are described in detail below. Regarding the data augmentation module, in order to improve the generalization ability of the model and alleviate the overfitting problem to a certain extent, data augmentation operations are performed on the data collected by the sensors. Here, an improved BEGAN network model is used to generate data to solve the data imbalance problem, so that the fault diagnosis model can achieve better results.

[0013] In this embodiment, the framework of the BEGAN network model Figure 2 As shown, the BEGAN network model consists of a generator and a discriminator, trained through a game-like interaction between them. The generator updates its parameters using the loss function passed from the discriminator, thus generating more data that is difficult for the discriminator to identify. Simultaneously, the discriminator adjusts its parameters using the parameters passed from the generator to determine the authenticity of the input data. Ultimately, the two models reach an equilibrium during training, at which point training ends. The generated data does not originate from the original data samples but is entirely new data generated using the parameters passed from the discriminator. The improved BEGAN network can better balance the training of the generator and discriminator in generation tasks and can generate more realistic samples. The loss functions for the generator and discriminator are as follows: L D =L(x; θ) D )-L(G(z D ;θ G );θ D ) L G =L(G(z) D ;θ G );θ D ) Where x is the real sample, θ D and θ G The parameters of the generator and discriminator are minimized by L. D and L G Update, z D It is a sample from z. The objective function of the improved BEGAN is: L D =L(x)-k t L(G(z D )) L G =L(G(z) G )) k t+1 =k t +β k (γL(x)-L(G(z G ))) Where: L(v)=|vD(v)| α ,and α∈{1,2}; k t The value is between [0, 1], and it is used to regulate the use of L(G(z) G The strength of gradient descent at time k0 = 0, β k Let L(v) represent the proportional gain of k, which is the learning rate of k; the discriminator is an autoencoder, so L(v) represents the reconstruction loss between data v and the data generated by the autoencoder D. D It represents the difference between the reconstruction loss of the real data x and the data x after passing through the autoencoder D, and the reconstruction loss of the data G(z) generated by the generative model and the data D(G(z)) after passing through D.

[0014] To better control the category of the generated data, the data category c is introduced into the BEGAN model, and its generator and discriminator model structure is as follows: Figure 3-4 As shown: Figure 3 As a generator model, in order to better control the category of the generated data, the random noise z is first concatenated with the features of the data category c as input. Then, after passing through a fully connected layer, it is converted into a two-dimensional tensor. Finally, it is output after passing through 4 convolutional layers and 3 upsampling layers. Figure 4 The discriminator model consists of an encoder and a decoder. First, the encoder takes a one-dimensional tensor as input and passes it through three downsampling convolutional layers to obtain a two-dimensional tensor. Then, it outputs a one-dimensional feature vector y through a fully connected layer. This feature vector is then concatenated with the feature vector of the encoded data category c. The decoder uses the same structure as the generator model. Adam is used to optimize the model parameters, balancing the different data types, and this optimized model serves as the input to the feature extraction module. Visualizations of the original and generated data signal graphs are shown below. Figure 5 As shown;

[0015] Regarding the feature extraction and fusion module, by Figure 1 As can be seen, the enhanced current signal and vibration signal data are transformed by fast Fourier transform, and then fed into a multi-scale network model to extract features. Then, an attention mechanism is introduced to adjust the weights, and then feature dimensionality reduction is performed separately for feature fusion. The specific process is described as follows: Since the impact caused by the fault in bearing fault signals is usually periodic, the spectral characteristics obtained after Fourier transform can generally reflect the fault information well. However, bearing signals are often discrete, while Fourier transform is typically suitable for continuous signals. Therefore, Discrete Fourier Transform (DFT) can be used for signal analysis. However, this method is computationally intensive and difficult to implement in real time. When calculating a one-dimensional time series of length N, N operations are required. 2 The DFT algorithm involves N complex multiplications and N(N-1) complex additions. Therefore, when the time series is long or the system has high real-time requirements, the DFT algorithm usually cannot meet the speed requirements. The Fast Fourier Transform (FFT) is an improved algorithm of DFT. The FFT uses a weight function... By leveraging the periodicity and symmetry of the data, the DFT algorithm is used to decompose and combine one-dimensional sequences of length N, transforming the original DFT calculation process into a process of superimposing multiple expressions. This greatly simplifies the computation of the DFT, enabling efficient signal analysis and processing. The definition formula for FFT is: n = 0, 1, 2, 3, ..., N-1, where N is the time series length; W is the weight function of the DFT; the vibration signal and current signal data are converted into a spectrum using FFT and displayed as follows. Figure 6 As shown;

[0016] After the enhanced current signal and vibration signal data are converted to the frequency domain by Fast Fourier Transform, the subsequent process is as follows: First, the current signal and vibration signal, after FFT transformation, are input into the first convolutional layer with a kernel size of 1×3. This convolutional layer performs convolution operations on the input data to extract initial shallow features. Then, multi-scale convolutional kernels are used to extract deeper features. The advantages of multi-scale convolutional kernel feature extraction are: it can capture features at different scales, the network can learn feature representations at different levels simultaneously, it can share some parameters, reducing the number of network parameters and computational cost, and improving feature extraction capabilities. This multi-scale feature extraction module first uses parallel convolutional filtering operations, processing convolutional layers with three sets of convolutional kernels of different scales. These kernel sizes are 1×3, 1×5, and 1×7, respectively. Next, the result processed by the previous layer is input into the next multi-scale convolutional kernel layer, where the kernel size is the same as the previous layer. The multi-scale convolution operation formula is: Among them, M n，l+1，q W represents the n deep feature maps output at layer l+1 after multi-scale convolution operations; i，n，l+1，q(q = 3, 5, 7) is used to represent the parameters of three sets of convolutional kernels of different sizes in the multi-scale feature extraction module; Secondly, the extracted multi-scale features are fused and input into a channel-space attention mechanism to model and adjust the channel correlation of the input features, thereby enhancing the model's ability to learn the importance of features from different channels. By introducing the channel-space attention mechanism, the model can automatically learn the weight of each channel, focusing more attention on channels that have a significant impact on the current task and reducing attention to irrelevant channels, thus improving model performance and generalization ability. Finally, grouped convolutions are performed, dividing the convolutional kernels into multiple groups, with each group's kernel processing only a subset of the input feature map. Here, the grouped convolutions divide the kernels into four groups, each with a 1×1 kernel size. This reduces computational complexity while enhancing the model's representational power. Finally, the features extracted by the two branch networks are fused.

[0017] For specific network model parameters, please refer to the attached instruction manual. Figure 7 The parameters of each network structure are as follows: the parameters of the convolutional layer are [1, 16, 1×3], the parameters of the multi-scale convolutional layer 1 are [16, 32, 1×3, 1], [16, 32, 1×5, 2] and [16, 32, 1×7, 3], the parameters of the multi-scale convolutional layer 2 are [32, 64, 1×3, 1], [32, 64, 1×5, 2] and [32, 64, 1×7, 3], the parameters of the channel spatial attention mechanism are [192, 192], and the parameters of the grouped convolutional layer are [192, 32, 1×1]group=4;

[0018] The fault classification module mainly consists of two parts: self-correcting convolution and dilated convolution, as described in detail below: Regarding self-calibrating convolution, since the output of the feature extraction and fusion module has multi-scale features, self-calibrating convolution (SCConv) can perform convolutional feature transformation between the original scale space and the self-calibrating space with smaller resolution obtained by downsampling without introducing any parameters or model complexity, thereby improving feature extraction capabilities. The operation of self-correcting convolution is as follows: Figure 8 As shown, the specific steps are as follows: Step 1: Input Features Figure X Given a size of C×H×W, it is split into two equal parts according to the number of channels, denoted as... Step 2: Split the convolution kernel K into four parts of the same dimension, each with a different function. Each part has a dimension of C / 2 × C / 2 × H × W, denoted as... Step 3: Process the original scale feature space by performing a K4 convolution on feature X2 to obtain Y2; Step 4: Process the self-correcting scale space. Input feature X1 into two different resolution channels. Reduce the width of the input feature by a factor of r to lower the resolution. Use average pooling to downsample feature X1 to correct the convolution transformation of the high-resolution part with low dimension. This allows for the extraction of important features at a smaller resolution scale while reducing interference from useless information. Then, convolve and upsample it, calculate the output using the Sigmoid function, and perform self-correction with the features extracted by K2. Finally, output the self-correcting feature Y1 through K3 convolution. The specific calculation operations are as follows: T = Avgpool r (X1) In the above calculation formula, Avgpool r (·) represents average pooling; r represents the downsampling rate; Up(·) represents upsampling; σ(·) represents the Sigmoid function; Step 5: Fuse feature Y1 and feature Y2 to output the final feature Y. It should be noted that self-calibrating convolution focuses only on the feature information of each spatial location, adaptively encodes information over a long range of regions, avoids interference from irrelevant feature information, reduces the dimensionality of features, increases the receptive field of the model, and improves the feature extraction capability.

[0019] Dilated convolution is a convolution method that increases the receptive field by adding a certain dilation rate between the elements of the convolution kernel. It extracts features by skipping a stride, thus extracting more features without increasing the kernel parameters. The formula for calculating dilated convolution is: Where Y(·) represents the input feature with length B, W(i) represents the size of the convolution kernel, r represents the dilation rate, and Z(x) represents the output feature after dilated convolution. Alternating between self-calibrating convolutions and dilated convolutions allows for flexible expansion of the receptive field without introducing additional parameters, while maintaining the resolution of the feature map, thereby optimizing the computational efficiency of the model. After passing through pooling layers, feature flattening, and fully connected layers, the model finally outputs the fault classification result through a Softmax classifier.

[0020] Specific model parameters are as follows Figure 9As shown: The parameters of each network structure are as follows: Self-calibrating convolution 1 parameters are [64, 128], dilated convolution 1 parameters are [128, 128] with dilation = 2, self-calibrating convolution 2 parameters are [128, 256], dilated convolution 2 parameters are [256, 256] with dilation = 2, and fully connected layer parameters are [256, 9].

[0021] like Figure 10 The flowchart shown is for a bearing fault diagnosis method based on multi-sensor multi-scale feature fusion. The detailed steps are as follows: Step 1: Collect current and vibration signals from the bearing during its operation using sensors to create a dataset; Step 2: Use the BEGAN network model to augment the dataset samples and divide them into training, validation and test sets in an 8:1:1 ratio; Step 3: Build a multi-scale feature fusion network model and set the model parameters; Step 4: After converting the current signal and vibration signal data in the training set from Step 2 into the frequency domain using Fast Fourier Transform (FFT), the data are fed into the multi-scale network model to extract features from the training set. Step 5: Adjust the attention weights of each feature vector using the attention mechanism layer for the features extracted in Step 4 through the multi-scale network model; Step 6: Finally, perform grouped convolution, dividing the convolution kernel into multiple groups. Each group's convolution kernel only processes a subset of the input feature map. Here, the grouped convolution divides the convolution kernel into 4 groups, and the size of each group's convolution kernel is 1×1. Finally, the features extracted by the two branch networks are fused. Step 7: The fused features of the two branches are used alternately with self-correcting convolution and dilated convolution to improve the model's representation ability and generalization performance. Then, after pooling, feature flattening and full connection, the fault classification result is finally output through the Softmax classifier. Step 8: Input the validation set into the multi-scale network model trained in Step 6, evaluate the model's performance, and fine-tune the model's parameters based on the validation results; Step 9: Determine if the number of training iterations M of the network model has reached the preset number of iterations N. If yes, proceed to the next step; otherwise, repeat step 4. It should be noted that the following factors need to be considered when setting the number of training iterations in this step: First, problem complexity: Complex problems may require more training and iterations. If the problem has highly nonlinear or abstract characteristics, more iterations may be needed to capture this complexity. Second, dataset size: Larger datasets generally require more training iterations. More data helps the model generalize better and reduces the risk of overfitting. Third, early stopping strategy: Stop training when performance on the validation set no longer improves. This helps prevent overfitting and allows finding an appropriate number of training iterations without wasting computational resources. Fourth, iteration efficiency: Balance the number of training iterations based on the efficiency of computational resource utilization. With limited resources, fewer training iterations may be needed to ensure the model makes good progress within a given time. Step 10: Input the test set into the network model validated in Step 8 to test the performance of the validated network model; Step 11: Finally, use the test set to perform performance testing on the trained model, calculate the evaluation metrics, and output the results.

[0022] It should be noted that the current signal acquisition described in this solution is accomplished using a current sensor or current clamp. The process is as follows: the sensor is clamped to the power line of the bearing to monitor changes in current and record the real-time current waveform. An appropriate sampling frequency can ensure that features related to the bearing condition are captured.

[0023] To demonstrate the effectiveness of the bearing fault diagnosis method described in this invention, an analysis of the bearing dataset from the University of Paderborn was conducted, as follows: This dataset was collected under different operating conditions. Each operating condition contains four bearing states: normal, inner ring failure, outer ring failure, and combined failure. Each failure type includes failure severity with a damage diameter of less than or equal to 2 inches and a damage diameter greater than 2 inches. Inner ring failure and combined failure also include damage diameters greater than 4.5 inches. To verify the intelligent fault diagnosis performance of this invention, A, B, and C are used to represent the following three working conditions, such as... Figure 11 As shown, A, B, and C have the same rotational speed parameters, A and B have the same load torque parameters, and A and C have the same radial force loading parameters. Nine different fault states are set for each operating condition, and each fault state includes 1,000 samples, for a total of 9,000. In scenario B, the existing methods WDCNN and ECNN achieve accuracies of 98.54% and 88.82%, respectively. This proposed method, however, achieves an accuracy of 99.86%. Under varying operating conditions, the model's accuracy can, for example... Figure 12 As shown, the accuracy obtained by using the present invention is far superior to the accuracy of the existing methods WDCNN and ECNN.

[0024] The parts not described in detail in this embodiment are existing technologies.

[0025] It should be noted that although the present invention has been described through the above embodiments, the present invention may have many other embodiments. Without departing from the spirit and scope of the present invention, those skilled in the art can obviously make various corresponding changes and modifications to the present invention, but all such changes and modifications should fall within the scope of protection of the appended claims and their equivalents.

Claims

1. A bearing fault diagnosis method based on multi-sensor multi-scale feature fusion, characterized in that: Includes the following steps: S1. Collect current and vibration signals of the bearing under operating conditions through sensors to create a dataset. Use the BEGAN network model to enhance the dataset samples and divide the enhanced samples into training set, validation set and test set. S2. Construct a multi-scale feature fusion network model. Convert the current signal and vibration signal data in the training set into the frequency domain through Fast Fourier Transform, and then feed them into the multi-scale network model to extract the features of the training set. Introduce the extracted features into the attention mechanism to adjust the weights, and then fuse the features extracted by the two branch networks. The specific process in step S2 is as follows: S2.

1. The current signal and vibration signal after Fast Fourier Transform are input into the first layer of convolution. The convolution layer will perform convolution operation on the input data to extract the initial shallow features, and then use multi-scale convolution kernels to extract deep features. S2.

2. The extracted multi-scale features are fused and introduced into the channel spatial attention mechanism to model and adjust channel correlation. S2.3 Perform grouped convolutions, and finally fuse the features extracted by the two branch networks; In step S2.1, firstly, a parallel convolutional filtering operation is performed using three sets of convolutional kernels of different scales for convolutional layer processing. The sizes of these convolutional kernels are 1×3, 1×5, and 1×7, respectively. Then, the result after processing in the previous layer is input into the next layer of multi-scale convolutional kernel convolutional layer, and the size of the convolutional kernel in this layer is the same as that in the previous layer. In step S2.3, during the grouped convolution process, the convolution kernel is divided into multiple groups. The convolution kernel of each group only processes a subset of the input feature map, and the size of the convolution kernel of each group is 1×1. S3. The fused features are processed alternately using self-correcting convolution and dilated convolution to improve the model's representation ability and generalization performance. After pooling, feature flattening and full connection, the fault classification result is finally output through the classifier. S4. Input the validation set into the multi-scale network model trained in step S2, evaluate the model's performance, and fine-tune the model's parameters based on the validation results. S5. Determine whether the number of training iterations M of the network model has reached the preset number of iterations N. If yes, proceed to the next step. Otherwise, repeat step S2; S6. Input the test set into the network model validated in step S4 to test the performance of the validated network model. Finally, use the test set to test the trained model and output the results.

2. The bearing fault diagnosis method based on multi-sensor multi-scale feature fusion according to claim 1, characterized in that: In step S1, the BEGAN network model consists of two parts: a generator model and a discriminator model. The generator updates its own parameters through the loss function passed by the discriminator, and the discriminator adjusts its own parameters through the parameters passed by the generator. The two are trained and reach an equilibrium state through a game between them.

3. The bearing fault diagnosis method based on multi-sensor multi-scale feature fusion according to claim 2, characterized in that: The aforementioned BEGAN network model, wherein the loss functions for the generator model and the discriminator model are: in, For real samples, and The parameters of the generator and discriminator are determined by minimizing the loss. and renew, It is a sample from z; The objective function of the BEGAN network model is: in: ,and ; ; Values ​​in Between, used for regulation in use The strength of gradient descent, initial value , This represents the proportional gain of k. This represents the reconstruction loss between data v and the data generated by the autoencoder D. This represents the reconstruction loss of real data x and the data after x passes through the autoencoder D, and the data generated by the generative model. and the data after D The difference in reconstruction loss.

4. The bearing fault diagnosis method based on multi-sensor multi-scale feature fusion according to claim 1, characterized in that: In step S1, the enhanced samples are divided into training set, validation set and test set in a ratio of 8:1:

1.

5. The bearing fault diagnosis method based on multi-sensor multi-scale feature fusion according to claim 1, characterized in that: In step S3, the specific steps of the self-correcting convolution are as follows: S3.1 Input Feature Map for The size is divided into two equal parts according to the number of channels, denoted as . ; S3.

2. Split the convolution kernel K into four parts of the same dimension, each with a different function and the same dimension size. , recorded as ; S3.

3. Process the original scale feature space, and convert the features... go through Convolution yields ; S3.

4. Process the self-correcting scale space, and convert the features... When the input is fed into two different resolution channels, the width of the input feature decreases. To reduce the resolution, average pooling is used for the features. Perform downsampling, then convolution and upsampling, and calculate the output and... The extracted features are self-corrected, and finally... Convolutional output self-correcting features ; S3.5, Features With features The components are fused together, and the final feature Y is output.

6. The bearing fault diagnosis method based on multi-sensor multi-scale feature fusion according to claim 1, characterized in that: In step S3, the formula for calculating dilated convolution is: in, Represents the input features, with a length of , To represent the size of the convolution kernel, Indicates the void ratio, This indicates the output features after dilated convolution.

Images