Rolling bearing fault diagnosis method and system based on multi-channel adaptive fusion

Abstract

The invention discloses a rolling bearing fault diagnosis method and system based on multi-channel adaptive fusion, and the method comprises the steps: obtaining a one-dimensional vibration acceleration signal, carrying out the normalization, and converting the one-dimensional vibration acceleration signal into a two-dimensional gray-scale map; inputting the two-dimensional grey-scale map into a convolutional neural network model for processing, specifically comprising the following steps: a multi-scale feature extraction module extracts feature information of the two-dimensional grey-scale map, adaptively weights the importance weight of multi-channel information, and outputs a first multi-scale feature map; a double-attention cavity convolution module extracts the spatial position and channel information of the first multi-scale feature map, and outputs a second feature map; and the residual dense module performs deep feature extraction on the second feature map through residual connection and dense connection, and outputs a rolling bearing fault diagnosis classification result. According to the method, the problems of low model fault diagnosis effect and poor generalization ability caused by strong noise interference and large load change of the rolling bearing can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling bearing fault diagnosis, and more particularly to a rolling bearing fault diagnosis method and system based on multi-channel adaptive fusion. Background Art

[0002] Rolling bearings are a commonly used rotating mechanical component used to support the rotation of shafts. In actual working environments, bearings are prone to failures when rotating at high speed for a long time. If a failure is not diagnosed and handled in time, it will have a serious impact on the entire mechanical operation system. Therefore, effective diagnosis of rolling bearing failures is of great practical significance.

[0003] Traditional fault diagnosis methods mainly include support vector machines, K-nearest neighbor methods and BP neural networks. Although these methods have certain feature information extraction capabilities, they cannot obtain higher fault diagnosis accuracy due to their shallow network structure, which makes it difficult to extract deep feature information. Moreover, they ignore the problem of insufficient feature information extraction under noise interference, and the diagnostic accuracy under strong noise and variable load conditions needs to be improved.

[0004] The present invention aims to solve the problems of insufficient feature information extraction and low diagnostic accuracy of convolutional neural networks in the prior art when processing rolling bearing fault diagnosis, especially when diagnosing rolling bearing signals under various loads under variable load and strong noise conditions. Summary of the invention

[0005] In view of this, the present invention provides a rolling bearing fault diagnosis method and system based on multi-channel adaptive fusion, which can solve the problems of low model fault diagnosis effect and poor generalization ability of rolling bearings due to strong noise interference and large load changes.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] The rolling bearing fault diagnosis method based on multi-channel adaptive fusion includes:

[0008] Obtain a one-dimensional vibration acceleration signal and normalize it, and convert it into a two-dimensional grayscale image;

[0009] The two-dimensional grayscale image is input into the convolutional neural network model for processing, including:

[0010] The multi-scale feature extraction module extracts feature information of the two-dimensional grayscale image and performs adaptive weighting on the importance weights of multi-channel information, and outputs a first multi-scale feature map;

[0011] The dual-attention hole convolution module extracts the spatial position and channel information of the first multi-scale feature map and outputs the second feature map;

[0012] The residual dense module performs deep feature extraction on the second feature map through residual connection and dense connection, and outputs the rolling bearing fault diagnosis classification result.

[0013] Preferably, the multi-scale feature extraction module extracts feature information of the two-dimensional grayscale image and adaptively weights the importance weights of multi-channel information to output a first multi-scale feature map, specifically including:

[0014] Obtain the initial features of the two-dimensional grayscale image through a wide convolutional layer;

[0015] The features are further refined through four convolutional layers;

[0016] The results of the convolutional layer are normalized through the BN layer;

[0017] Nonlinear calculations are introduced after normalization through the ReLU activation layer;

[0018] Reduce the feature map size through the maximum pooling layer;

[0019] SK-Net is used to dynamically adjust the importance of features at different scales.

[0020] Preferably, the dual-attention hole convolution module extracts the spatial position and channel information of the first multi-scale feature map and outputs the second feature map, specifically including:

[0021] The dilated convolution layer performs a dilated convolution operation on the first multi-scale feature map;

[0022] The maximum pooling layer and the mean pooling layer perform maximum pooling and mean pooling operations on the feature map after the dilated convolution respectively;

[0023] The BN layer normalizes the feature map after the pooling operation;

[0024] The ReLU activation layer performs ReLU activation on the normalized feature map;

[0025] The feature map after weighted activation of the channel attention layer enters the channel attention layer and is weighted and adjusted according to the importance of each channel; the weighted feature map is merged with the feature maps after the other two dilated convolutions to generate the final second feature map.

[0026] Preferably, the residual dense module performs deep feature extraction on the second feature map through residual connection and dense connection, and outputs the rolling bearing fault diagnosis classification result, which specifically includes:

[0027] Different deep features are extracted through three groups of convolutional layers, and the output of the first group of convolutional layers is connected with the input and output of the third group of convolutional layers through residual connections;

[0028] The rolling bearing fault diagnosis classification results are output through the Softmax layer.

[0029] Preferably, a model training phase is also included, in which a random sampling strategy is adopted:

[0030] Assume there is a one-dimensional data sequence x=[x 1 ,x 2 ,...,x N ], and samples need to be extracted from it, each sample length is L, L = 1024, step length is S, S = 128, then for each sample k, its starting index i k Calculated by the following formula:

[0031] i k = k·S

[0032] Where: k = 0, 1, 2, ...., M-1, M is the total number of samples extracted:

[0033]

[0034] Where: Indicates a round-down operation;

[0035] The kth sample is:

[0036] x k =[x ik+1 ,x ik+2 ,....,x ik+L ].

[0038] Rolling bearing fault diagnosis system based on multi-channel adaptive fusion, including:

[0039] Signal acquisition and preprocessing module: used to obtain and normalize the one-dimensional vibration acceleration signal and convert it into a two-dimensional grayscale image;

[0040] Convolutional neural network model processing module: used to input the two-dimensional grayscale image into the convolutional neural network model for processing, including:

[0041] The multi-scale feature extraction module extracts feature information of the two-dimensional grayscale image and performs adaptive weighting on the importance weights of multi-channel information, and outputs a first multi-scale feature map;

[0042] The dual-attention hole convolution module extracts the spatial position and channel information of the first multi-scale feature map and outputs the second feature map;

[0043] The residual dense module performs deep feature extraction on the second feature map through residual connection and dense connection, and outputs the rolling bearing fault diagnosis classification result.

[0044] Preferably, the multi-scale feature extraction module includes a plurality of parallel CNN modules, each of which includes a wide convolution layer, 4 convolution layers, a maximum pooling layer and an SK-Net in sequence, wherein each convolution layer includes a BN layer and a ReLU activation layer;

[0045] The wide convolutional layer is used to obtain the initial features of the two-dimensional grayscale image;

[0046] Four convolutional layers are used to further refine the features;

[0047] The BN layer is used to normalize the results of the convolutional layer;

[0048] The ReLU activation layer is used to introduce nonlinear calculations after normalization;

[0049] The maximum pooling layer is used to reduce the feature map size;

[0050] SK-Net is used to dynamically adjust the importance of features at different scales.

[0051] Preferably, the dual-attention hole convolution module includes three parallel layers, each of which includes a hole convolution layer, a maximum pooling layer, a mean pooling layer, a BN layer and a ReLU activation layer, and a channel attention layer is arranged after the parallel layer;

[0052] The dilated convolution layer is used to perform a dilated convolution operation on the first multi-scale feature map;

[0053] The maximum pooling layer and the mean pooling layer perform maximum pooling and mean pooling operations on the feature map after the dilated convolution respectively;

[0054] The BN layer normalizes the feature map after the pooling operation;

[0055] The ReLU activation layer performs ReLU activation on the normalized feature map;

[0056] The feature map after weighted activation of the channel attention layer enters the channel attention layer and is weighted and adjusted according to the importance of each channel; the weighted feature map is merged with the feature maps after the other two dilated convolutions to generate the final second feature map.

[0057] Preferably, the residual dense module comprises three groups of convolutional layers, wherein the output of the first group of convolutional layers is residually connected to the input and output of the third group of convolutional layers respectively, and the output of the third group of convolutional layers is connected to the Softmax layer;

[0058] The three groups of convolutional layers extract different deep features respectively, and the output of the first group of convolutional layers is residually connected with the input and output of the third group of convolutional layers respectively;

[0059] The Softmax layer is used to output the rolling bearing fault diagnosis classification results.

[0060] Preferably, a model training module is also included, and a random sampling strategy is adopted in the model training stage:

[0061] Assume there is a one-dimensional data sequence x=[x 1 ,x 2 ,...,x N ], and samples need to be extracted from it, each sample length is L, L = 1024, step length is S, S = 128, then for each sample k, its starting index i k Calculated by the following formula:

[0062] i k = k·S

[0063] Where: k = 0, 1, 2, ...., M-1, M is the total number of samples extracted:

[0064]

[0065] Where: Indicates a round-down operation;

[0066] The kth sample is:

[0067] x k =[x ik+1 ,x ik+2 ,....,x ik+L ].

[0069] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a rolling bearing fault diagnosis method and system based on multi-channel adaptive fusion, which has the following advantages:

[0070] 1) The present invention does not need to consider problems such as gradient vanishing and shallow information loss, as well as the process of manual feature selection. By designing a reasonable network structure, the present invention can automatically learn and extract fault information from the original vibration signal. This not only reduces the workload of manual intervention, but also improves the efficiency and accuracy of fault diagnosis, enabling the model to autonomously mine key information.

[0071] 2) In the model training phase, the present invention adopts a random sampling strategy to randomize the training data. This method greatly improves the system's ability to resist noise interference, and the model can maintain stable fault diagnosis performance even in a high-noise environment. Random sampling technology can effectively enhance data diversity, prevent model overfitting, and improve the model's robustness in dealing with various complex noise environments.

[0072] 3) The present invention adopts dilated convolution and SK-Net to significantly enhance the receptive field of the model. Through dilated convolution, more long-distance dependent feature information can be captured without adding additional computational cost, while SK-Net enables the model to perform adaptive weighted processing according to the importance of each feature channel. The combination of the two greatly improves the robustness and generalization ability of the model under diverse working conditions, enabling it to accurately diagnose fault types under complex working conditions.

[0073] 4) The present invention combines the feature information of maximum pooling and average pooling, which improves the model's ability to perceive local and global feature information and effectively enhances the diversity of feature extraction. In addition, the present invention has made improvements in deep-level feature extraction and uses an optimized residual module for feature learning. By adding batch normalization and ReLU activation function after each convolutional layer, the stability of the network parameters during training is ensured, further optimizing the performance of the model.

[0074] 5) The present invention also has the ability to automatically adjust the distribution of test data, and can still achieve stable and accurate rolling bearing fault diagnosis even under strong noise interference and variable load conditions. Through a reasonably designed data normalization mechanism, the model can adapt to complex working conditions and ensure the stability of the diagnosis results. Whether in high load, low load or strong noise environment, the system can maintain a high diagnostic accuracy, thereby effectively meeting the actual needs for accurate fault diagnosis in industrial equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0076] Figure 1 A flow chart of the rolling bearing fault diagnosis method based on multi-channel adaptive fusion provided by the present invention.

[0077] Figure 2 This is a structural diagram of the convolutional neural network model provided by the present invention.

[0078] Figure 3 This is a schematic diagram of the structure of the multi-scale feature extraction module provided by the present invention.

[0079] Figure 4 This is a schematic diagram of the variable noise fault diagnosis results provided by the present invention.

[0080] Figure 5A schematic diagram for comparing the variable load fault diagnosis results provided by the present invention.

[0081] Figure 6 This is a schematic diagram for comparing the variable operating condition fault diagnosis results provided by the present invention.

[0082] Figure 7a , Figure 7b , Figure 7c , 7d The results of confusion matrix experiments performed on data with a signal-to-noise ratio of 3 dB under a 1HP load by 1DCNN, WDCNN, ECA_ResNet and the method of the present invention are respectively shown.

[0083] Figure 8 This is a corresponding diagnostic result diagram at a speed of 1130r / min in a noisy environment provided by the present invention.

[0084] Fig. 9 This is the fault diagnosis result of adding a 5.1kg load at a speed of 1130r / min in a noisy environment provided by the present invention.

[0085] Fig.10 This is a block diagram of the rolling bearing fault diagnosis system based on multi-channel adaptive fusion provided by the present invention. DETAILED DESCRIPTION

[0086] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0087] The embodiment of the present invention discloses a rolling bearing fault diagnosis method based on multi-channel adaptive fusion, such as Figure 1 As shown, including:

[0088] Obtain a one-dimensional vibration acceleration signal and normalize it, and convert it into a two-dimensional grayscale image;

[0089] The two-dimensional grayscale image is input into the convolutional neural network model for processing, including:

[0090] The multi-scale feature extraction module extracts feature information of the two-dimensional grayscale image and performs adaptive weighting on the importance weights of multi-channel information, and outputs a first multi-scale feature map;

[0091] The dual-attention hole convolution module extracts the spatial position and channel information of the first multi-scale feature map and outputs the second feature map;

[0092] The residual dense module performs deep feature extraction on the second feature map through residual connection and dense connection, and outputs the rolling bearing fault diagnosis classification result.

[0093] The present invention designs a multi-scale feature extraction module, which uses wide convolutions of different scales to capture different feature information of the input signal and suppress high-frequency noise; and uses SK-Net to adaptively weight the importance weights of multi-channel information to selectively learn and utilize important feature information; designs a dual-attention hole convolution module, which uses hole convolution layers with different hole rates to increase the feature information receptive field, and calculates the maximum pooling value and the average pooling value after splicing through channel attention to improve the network's perception of local and global features; constructs a residual dense block with jump connections to solve the problems of feature information loss and model overfitting; finally, the fault category diagnosis result is obtained by the Softmax layer, and the rolling bearing signals under various loads are diagnosed.

[0094] In this embodiment, if Figure 2 As shown, the multi-scale feature extraction module extracts the feature information of the two-dimensional grayscale image and adaptively weights the importance weights of the multi-channel information to output a first multi-scale feature map, specifically including:

[0095] The multi-scale feature extraction module includes multiple parallel CNN modules, each of which includes a wide convolution layer, 4 convolution layers, a maximum pooling layer and an SK-Net in sequence, where each convolution layer includes a BN layer and a ReLU activation layer; the convolution kernel size of the wide convolution layer is (128, 1), the number of channels is 16, the sizes of the four convolution layers are (3, 1), the number of channels are 32, 64, 64, 64, the pooling layer step size is 2, and the output channel is 64; the normalization formula of the BN layer is:

[0096]

[0097] Where: x is the original data point, x min and x max are the maximum and minimum values ​​of the feature in the dataset, respectively.

[0098] Use wide convolutional layers to obtain the initial features of the two-dimensional grayscale image;

[0099] The features are further refined through four convolutional layers;

[0100] The results of the convolutional layer are normalized through the BN layer;

[0101] Introduce nonlinear calculations through the ReLU activation layer;

[0102] Use the max pooling layer to reduce the feature map size;

[0103] SK-Net is used to dynamically adjust the importance of features at different scales.

[0104] SK-Net is used to perform channel weighting to capture key information, and a large-size convolution kernel is used to increase the network's field of view, thereby extracting periodic fault features and suppressing high-frequency noise. The BN method is used to process the features extracted by the convolution layer to reduce the gap between the distribution of training data and test data, speed up the convergence of the model, and reduce the training time of the model. Finally, the ReLU activation function is used to avoid gradient disappearance, so that subsequent steps can obtain more comprehensive information. The ReLU function is:

[0105] f(x)=max(x,0)

[0106] Where: means that for any input x, if x is positive, the output is positive, and if x is negative, the output is 0. Compared with saturated nonlinear functions such as sigmoid or tanh, ReLU is more computationally efficient because it only contains threshold operations. In the positive range, the derivative of ReLU is 1, which means that it does not cause the gradient to approach zero at large or small input values ​​like sigmoid or tanh, which helps the training of deep networks.

[0107] The dual-attention hole convolution module extracts the spatial position and channel information of the first multi-scale feature map and outputs the second feature map, which specifically includes:

[0108] The dual-attention atrous convolution module consists of three parallel layers, each of which includes a atrous convolution layer, a maximum pooling layer, a mean pooling layer, a BN layer, and a ReLU activation layer. A channel attention mechanism is added after the parallel layer to weight important detail feature information. The convolution kernel sizes of the first to third layers are 1×1, 3×3, and 5×5, respectively, and the output channels are 64. The atrous convolution layers of the first to third layers have different atrous rates, which are 6, 4, and 2, respectively.

[0109] The dilated convolution layer is used to perform a dilated convolution operation on the first multi-scale feature map;

[0110] The maximum pooling layer and the mean pooling layer perform maximum pooling and mean pooling operations on the feature map after the dilated convolution respectively;

[0111] The BN layer normalizes the feature map after the pooling operation;

[0112] The ReLU activation layer performs ReLU activation on the normalized feature map;

[0113] The feature map after weighted activation of the channel attention layer enters the channel attention layer and is weighted and adjusted according to the importance of each channel;

[0114] The weighted feature map is combined with the feature maps after the other two dilated convolutions to generate the final second feature map.

[0115] Among them, the channel attention mechanism includes: for CNN with 2D image input, one dimension is the scale space of the image, that is, length and width, and the other dimension is the channel, and different channels often contribute differently to the final task. The channel attention mechanism dynamically adjusts the weight of each channel so that the network can focus on those features that are more critical to the task, thereby improving the overall performance. The core of the channel attention mechanism is to assign a weight to each channel, which reflects the importance of the channel to the final task.

[0116] The dual-attention dilated convolution module of the present invention uses dilated convolutions with dilation rates of 2, 4, and 6 to increase the receptive field, concatenates information through maximum pooling and average pooling, and enhances the perception of local and global feature information; then, the global information is modeled through the channel attention mechanism to generate channel attention weights.

[0117] The residual dense module performs deep feature extraction on the second feature map through residual connection and dense connection, and outputs the rolling bearing fault diagnosis classification results, including:

[0118] The residual dense module includes three groups of convolutional layers, where the output of the first group of convolutional layers is residually connected to the input and output of the third group of convolutional layers, and the output of the third group of convolutional layers is connected to the Softmax layer;

[0119] Different deep features are extracted through three groups of convolutional layers, and the output of the first group of convolutional layers is residually connected with the input and output of the third group of convolutional layers. The output of the last convolutional layer is used as the input of softmax.

[0120] The rolling bearing fault diagnosis classification results are output through the Softmax layer.

[0121] The residual network of the present invention can extract interrelated information, process features generated by convolutional networks, and better capture the associated information of original data. At the same time, compared with a single convolutional network, multi-scale convolution can better work stably under variable load and strong noise interference.

[0122] The Softmax function is an activation function commonly used in multi-classification tasks, especially in the output layer of neural networks. It converts a set of real numbers into a probability distribution so that the sum of the output values ​​is 1 and each element is in the interval [0, 1]. The Softmax function is very suitable for representing class probabilities because its output can be directly interpreted as a probability estimate for each class.

[0123] In this embodiment, it further includes a model training module, specifically including:

[0124] Collect one-dimensional vibration acceleration signals through an acceleration sensor, including signals of bearings with normal bearings, inner ring faults, rolling element faults, and outer ring faults, as well as signals with different fault diameters, ensuring that high-resolution vibration data is obtained, covering various complex working conditions, and can also cover different fault degrees, mild, moderate, and severe;

[0125] Normalize the collected one-dimensional vibration acceleration signals to obtain normalized vibration acceleration signals to eliminate the influence of amplitude differences on model training;

[0126] Convert the normalized acceleration signals into two-dimensional grayscale images of size 32×32;

[0127] Divide the processed two-dimensional grayscale images into a training set and a test set according to a ratio. Specifically, use 1024 sampling points as a sample, and perform sliding overlapping sampling with a step size of 128. Set mixed fault type labels for the processed two-dimensional grayscale images according to the corresponding fault types of each sample, and construct the required data set. Through the above method, sampling can be prevented when the length of the collected data samples is not long enough, so that the sample length of the latter part contains partial information of the previous sample length. Since the method of sliding window is used, there will be an overlapping part between each new sample and the previous sample (when S < L). This helps to capture continuous patterns or features in the data.

[0128] To more precisely describe how to extract overlapping samples through the sliding window method, this embodiment gives a formula to represent this process;

[0129] Suppose there is a one-dimensional data sequence x = [x 1 , x 2 ,..., x N of length N, and samples need to be extracted from it. Each sample has a length of L, L = 1024, and the step size is S, S = 128. Then for each sample k, its starting index i k is calculated by the following formula:

[0130] i k = k·S

[0131] In the formula: k = 0, 1, 2,...., M - 1, M is the total number of samples extracted:

[0132]

[0133] In the formula: represents the floor operation;

[0134] The k-th sample is:

[0135] x k =[x ik+1 ,x ik+2 ,....,x ik+L ]

[0136] The two-dimensional grayscale images in the training set are input together with the corresponding training labels into the constructed convolutional neural network model for training, so that the model can have higher recognition ability under complex working conditions.

[0137] The performance of the model is obtained through the test set, the current fault status of the rolling bearing is diagnosed, and accurate fault classification results are provided.

[0138] In order to better illustrate the technical effect of the present invention, a specific embodiment is used to experimentally verify the present invention. The Case Western Reserve University (CWRU) bearing data set was selected for this experiment. The test bench consists of a motor, a torque sensor and a power tester. The motor drive end bearing type is SKF-6205, the sampling frequency is 12Khz, and pits of different sizes are processed on the rolling bearing by electric sparks. The fault types and labels of the bearing data set are divided into 16 states. In actual industrial environments, rolling bearings often operate under different loads and are interfered by strong noise. Therefore, verifying the noise resistance and generalization ability of the fault diagnosis method is crucial for bearing fault diagnosis. The variable operating condition data set is divided into training samples and test samples in a ratio of 3:1, and the sampling length of each sample is 1024. The number of training data and test data and the fault type are shown in Table 1.

[0139] Table 1 Speed ​​load correspondence table

[0140]

[0141] In order to illustrate the technical performance of the present invention and the comparative method, the fault diagnosis results are evaluated using the accuracy index, and the calculation formula is as follows:

[0142]

[0143] Among them, TP represents the number of true positive samples, FP represents the number of false positive samples, TN represents the number of true negative samples, and FN represents the number of false negative samples.

[0144] Three comparison methods were set up in this experiment.

[0145] They are WDCNN (see the literature Beibei M, Yanxia S, Dinghui W, et al. Three-level inverter fault diagnosis using EMD and support vector machine approach [C] / / 2017 12th IEEE Conference on Industrial Electronics and Applications (ICIEA). IEEE, 2017: 1595-1598.),

[0146] ECA_ResNet (see the literature Van den Hoogen JOD, Bloemheuvel SD, Atzmueller M. Animproved wide-kernel CNN for classifying multivariate signals in faultdiagnosis[C] / / 2020International Conference on Data Mining Workshops(ICDMW).IEEE, 2020:275-283.),

[0147] 1DCNN (see the literature Sun S, Zhang T, Li Q, et al. Fault diagnosis of conventional circuit breaker contact system based on time-frequency analysis and improved AlexNet[J]. IEEE Transactions on Instrumentation and Measurement, 2020,70:1-12.),

[0148] Figure 4 Figure 2 is the accuracy and loss value of fault diagnosis of the proposed method under variable noise conditions. It can be seen that the accuracy of the proposed method is higher than 96% when the signal-to-noise ratio is 0dB, and the loss value converges to 0.009, while the loss value almost converges to 0 under other noise conditions, indicating that the proposed method has strong anti-noise performance.

[0149] The experimental results in a strong noise and multi-load mixed environment are as follows Figure 5As shown, the prediction accuracy of the neural network model proposed in the present invention is the highest. At the same time, it can be found that for tiny Gaussian white noise interference, most methods can make high-accuracy predictions. As the Gaussian white noise increases, only the method proposed in the present invention decays the slowest. Until the Gaussian white noise is 0dB, it has an accuracy rate of 95.00%, proving that its stability is also very good.

[0150] Figure 6 is the experimental result of variable working condition. Figure 6 It can be seen that the accuracy of 1DCNN, WDCNN and ECA_ResNet are 44.34%, 72.94% and 59.72% respectively when the signal-to-noise ratio is 0dB, while the diagnostic accuracy of the proposed method remains above 95%, which verifies that the proposed method has better generalization performance and noise resistance performance. Figure 6 HP stands for horsepower. The corresponding speeds for 0 to 3 horsepower are 1797r / min, 1772r / min, 1750r / min, and 1730r / min respectively.

[0151] Table 2 Comparison of noise resistance performance between the method of the present invention and different comparison methods

[0152]

[0153]

[0154] In Table 2, the accuracy of 1DCNN, WDCNN and ECA_ResNet are 44.34%, 72.94% and 59.72% respectively when the signal-to-noise ratio is 0 dB, while the average accuracy of the method of the present invention is 95.49%, and the diagnostic accuracy remains above 95%, which verifies that the method of the present invention has better generalization performance and noise resistance performance.

[0155] Figure 7a-7d The confusion matrix experiment of 1DCNN, WDCNN, ECA_ResNet and the proposed method on the data with a signal-to-noise ratio of 3dB under 1HP load is shown. The confusion matrix results in a 3dB noise environment show that the method of the present invention accurately identifies most fault types. In addition, compared with 1DCNN, WDCNN and ECA_ResNet, the method of the present invention has fewer samples that are incorrectly identified. The above results show that the fault diagnosis method proposed in the present invention can accurately identify bearing faults and has good noise resistance.

[0156] Figure 8 This is the fault diagnosis result when the speed is 1130r / min in a noisy environment. Figure 8It can be seen that the accuracy of the proposed method reaches more than 97.5% when the speed is 1130r / min and the signal-to-noise ratio is 0dB, which is 15.1% higher than WDCNN, 16.4% higher than ECA-ResNet, and significantly higher than 1DCNN.

[0157] Fig. 9 This is the fault diagnosis result when the speed is 1130r / min and a load of 5.1kg is added in a noisy environment. Fig. 9 It can be seen that the proposed method achieves an accuracy of 96.66% when the signal-to-noise ratio is 0dB. Therefore, it is proved that the proposed method has strong generalization performance.

[0158] In order to evaluate the classification performance of the proposed method, the present invention also introduces four performance indicators, namely accuracy, recall, precision and F1 value, to measure the detection ability of the classification model for different fault types.

[0159] The recall rate represents the proportion of samples that are correctly predicted as positive examples among all samples that are actually positive examples, as shown in the formula:

[0160] Recall=TP / (TP+FN)

[0161] Precision represents the proportion of samples that are actually positive among all samples predicted to be positive, as shown in the formula:

[0162] Precision=TP / (TP+FP)

[0163] The F1 value combines the accuracy and recall rate. It is a comprehensive indicator that represents the harmonic mean of the accuracy and recall rate, as shown in the formula:

[0164] F1=2*precision*recall /

[0165] (precision+recall)

[0166] Among them, TP refers to the number of samples whose true category is positive and predicted to be positive, TN refers to the number of samples whose true category is negative and predicted to be negative, FP refers to the number of samples whose true category is negative but predicted to be positive, and FN refers to the number of samples whose true category is positive but wrongly predicted to be negative.

[0167] Table 3 Performance indicators

[0168]

[0169] Table 3 shows the experimental results of data with a signal-to-noise ratio of 3 dB under a 1HP load. It can be seen that the method of the present invention has better performance.

[0170] In summary, the rolling bearing fault diagnosis method based on multi-channel adaptive fusion of the present invention can solve the problems of low model fault diagnosis effect and poor generalization ability of rolling bearings due to strong noise interference and large load changes.

[0171] The embodiment of the present invention discloses a rolling bearing fault diagnosis method system based on multi-channel adaptive fusion, such as Fig.10 As shown, including:

[0172] Signal acquisition and preprocessing module: used to obtain and normalize the one-dimensional vibration acceleration signal and convert it into a two-dimensional grayscale image;

[0173] Convolutional neural network model processing module: used to input the two-dimensional grayscale image into the convolutional neural network model for processing, including:

[0174] The multi-scale feature extraction module extracts feature information of the two-dimensional grayscale image and performs adaptive weighting on the importance weights of multi-channel information, and outputs a first multi-scale feature map;

[0175] The dual-attention hole convolution module extracts the spatial position and channel information of the first multi-scale feature map and outputs the second feature map;

[0176] The residual dense module performs deep feature extraction on the second feature map through residual connection and dense connection, and outputs the rolling bearing fault diagnosis classification result.

[0177] The specific implementation method of the system of the present invention is consistent with that of the method part of the above embodiment, which will not be repeated here. Please refer to the method part for details.

[0178] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0179] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A rolling bearing fault diagnosis method based on multi-channel adaptive fusion, characterized in that: include: Obtain a one-dimensional vibration acceleration signal and normalize it, and convert it into a two-dimensional grayscale image; The two-dimensional grayscale image is input into the convolutional neural network model for processing, including: The multi-scale feature extraction module extracts feature information of the two-dimensional grayscale image and performs adaptive weighting on the importance weights of multi-channel information, and outputs a first multi-scale feature map; The dual-attention hole convolution module extracts the spatial position and channel information of the first multi-scale feature map and outputs the second feature map; The residual dense module performs deep feature extraction on the second feature map through residual connection and dense connection, and outputs the rolling bearing fault diagnosis classification result.

2. The rolling bearing fault diagnosis method based on multi-channel adaptive fusion according to claim 1 is characterized in that: The multi-scale feature extraction module extracts the feature information of the two-dimensional grayscale image and adaptively weights the importance weights of the multi-channel information to output a first multi-scale feature map, specifically including: Obtain the initial features of the two-dimensional grayscale image through a wide convolutional layer; The features are further refined through four convolutional layers; The results of the convolutional layer are normalized through the BN layer; Nonlinear calculations are introduced after normalization through the ReLU activation layer; Reduce the feature map size through the maximum pooling layer; SK-Net is used to dynamically adjust the importance of features at different scales.

3. The rolling bearing fault diagnosis method based on multi-channel adaptive fusion according to claim 1 is characterized in that: The dual-attention hole convolution module extracts the spatial position and channel information of the first multi-scale feature map and outputs the second feature map, which specifically includes: The dilated convolution layer performs a dilated convolution operation on the first multi-scale feature map; The maximum pooling layer and the mean pooling layer perform maximum pooling and mean pooling operations on the feature map after the dilated convolution respectively; The BN layer normalizes the feature map after the pooling operation; The ReLU activation layer performs ReLU activation on the normalized feature map; The feature map after weighted activation of the channel attention layer is weighted and adjusted according to the importance of each channel; the weighted feature map is merged with the feature maps after the other two dilated convolutions to generate the final second feature map.

4. The rolling bearing fault diagnosis method based on multi-channel adaptive fusion according to claim 1 is characterized in that: The residual dense module performs deep feature extraction on the second feature map through residual connection and dense connection, and outputs the rolling bearing fault diagnosis classification results, including: Different deep features are extracted through three groups of convolutional layers, and the output of the first group of convolutional layers is connected with the input and output of the third group of convolutional layers through residual connections; The rolling bearing fault diagnosis classification results are output through the Softmax layer.

5. The rolling bearing fault diagnosis method based on multi-channel adaptive fusion according to claim 1 is characterized in that: It also includes the model training phase, in which a random sampling strategy is adopted: Suppose there is a one-dimensional data sequence x=[x1,x2,...,x N ], and samples need to be extracted from it, each sample length is L, L = 1024, step length is S, S = 128, then for each sample k, its starting index i k Calculated by the following formula: i k =k·S Where: k = 0, 1, 2, ...., M-1, M is the total number of samples extracted: Where: Indicates a round-down operation; The kth sample is: x k =[x ik+1 ,x ik+2 ,....,x ik+L ]。 6. Rolling bearing fault diagnosis system based on multi-channel adaptive fusion, characterized in that: include: Signal acquisition and preprocessing module: used to obtain and normalize the one-dimensional vibration acceleration signal and convert it into a two-dimensional grayscale image; Convolutional neural network model processing module: used to input the two-dimensional grayscale image into the convolutional neural network model for processing, including: The multi-scale feature extraction module extracts feature information of the two-dimensional grayscale image and performs adaptive weighting on the importance weights of multi-channel information, and outputs a first multi-scale feature map; The dual-attention hole convolution module extracts the spatial position and channel information of the first multi-scale feature map and outputs the second feature map; The residual dense module performs deep feature extraction on the second feature map through residual connection and dense connection, and outputs the rolling bearing fault diagnosis classification result.

7. The rolling bearing fault diagnosis system based on multi-channel adaptive fusion according to claim 6 is characterized in that: The multi-scale feature extraction module includes multiple parallel CNN modules, each of which includes a wide convolution layer, 4 convolution layers, a maximum pooling layer and an SK-Net in sequence, and each convolution layer includes a BN layer and a ReLU activation layer; The wide convolutional layer is used to obtain the initial features of the two-dimensional grayscale image; Four convolutional layers are used to further refine the features; The BN layer is used to normalize the results of the convolutional layer; The ReLU activation layer is used to introduce nonlinear calculations after normalization; The maximum pooling layer is used to reduce the feature map size; SK-Net is used to dynamically adjust the importance of features at different scales.

8. The rolling bearing fault diagnosis system based on multi-channel adaptive fusion according to claim 6 is characterized in that: The dual-attention dilated convolution module consists of three parallel layers, each of which includes a dilated convolution layer, a maximum pooling layer, a mean pooling layer, a BN layer, and a ReLU activation layer. A channel attention layer is added after the parallel layer. The dilated convolution layer is used to perform a dilated convolution operation on the first multi-scale feature map; The maximum pooling layer and the mean pooling layer perform maximum pooling and mean pooling operations on the feature map after the dilated convolution respectively; The BN layer normalizes the feature map after the pooling operation; The ReLU activation layer performs ReLU activation on the normalized feature map; The feature map after weighted activation of the channel attention layer is weighted and adjusted according to the importance of each channel; The weighted feature map is combined with the feature maps after the other two dilated convolutions to generate the final second feature map.

9. The rolling bearing fault diagnosis system based on multi-channel adaptive fusion according to claim 6, characterized in that: The residual dense module includes three groups of convolutional layers, where the output of the first group of convolutional layers is residually connected to the input and output of the third group of convolutional layers, and the output of the third group of convolutional layers is connected to the Softmax layer; The three groups of convolutional layers extract different deep features respectively, and the output of the first group of convolutional layers is residually connected with the input and output of the third group of convolutional layers respectively; The Softmax layer is used to output the rolling bearing fault diagnosis classification results.

10. The rolling bearing fault diagnosis system based on multi-channel adaptive fusion according to claim 6, characterized in that: It also includes a model training module, which adopts a random sampling strategy during the model training phase: Suppose there is a one-dimensional data sequence x=[x1,x2,...,x N ], and samples need to be extracted from it, each sample length is L, L = 1024, step length is S, S = 128, then for each sample k, its starting index i k Calculated by the following formula: i k =k·S Where: k = 0, 1, 2, ...., M-1, M is the total number of samples extracted: Where: Indicates a round-down operation; The kth sample is: x k =[x ik+1 ,x ik+2 ,....,x ik+L ]。

Images