Bearing fault diagnosis method based on IDSP and reverse DSC cascade residual error

Abstract

The invention discloses a bearing fault diagnosis method based on IDSP and reverse DSC cascade residual errors, and the method comprises the steps: obtaining one-dimensional vibration signals, from different position sensors, of a bearing operation platform, carrying out the preprocessing, and dividing the one-dimensional vibration signals into a training sample set and a verification sample set; training the fault diagnosis model by using the training sample set, and performing accuracy verification by using the verification sample set to obtain an optimal model; the fault diagnosis model comprises an IDSP shallow feature extraction module, an Inception deep feature extraction module and a reverse DSC cascade residual network module. And preprocessing a bearing one-dimensional vibration signal needing fault diagnosis, and inputting the preprocessed signal into the optimal model to obtain a corresponding fault diagnosis result. According to the method, through multi-scale feature extraction and optimization of a reverse DSC cascade residual network, a bearing fault diagnosis model has higher feature extraction capability, stronger noise immunity and generalization capability, and high diagnosis accuracy can be kept under variable working conditions and variable noise conditions; and meanwhile, the method also has less parameter quantity and higher diagnosis speed.

Description

Technical Field

[0001] The invention relates to the technical field of bearing fault diagnosis, and more particularly to a bearing fault diagnosis method based on IDSP and reverse DSC cascade residual. Background Art

[0002] As one of the most widely used components in rotating machinery, the operating status of rolling bearings directly affects the overall performance, working efficiency and service life of the equipment. The harsh working environment of high-speed variable loads can easily cause various types of bearing failures. In order to avoid the huge economic losses and unpredictable safety accidents caused by them, it is very necessary to carry out bearing fault diagnosis.

[0003] Traditional fault diagnosis methods usually perform fault diagnosis based on the vibration signal generated by the contact between the faulty component and the normal component, mainly including support vector machines, decision trees and random forests. However, these methods not only rely on the prior knowledge of experts and technicians, but also have certain limitations. For example, the performance of support vector machines is highly dependent on kernel selection, and decision trees and random forests are very sensitive to noisy data. In short, traditional fault diagnosis methods are difficult to extract deep network feature information due to shallow structural features and various constraints, so that the fault diagnosis effect is not good.

[0004] Compared with traditional fault diagnosis methods, fault diagnosis methods based on deep learning can automatically extract key fault features without any manual feature selection operations. Typical deep learning algorithms include denoising autoencoders, deep belief networks, sparse autoencoders, and convolutional neural networks. Among them, CNN does not need to introduce an unsupervised pre-training process to train the network model, and has the ability to eliminate the noise in the fault data set. Therefore, CNN models are widely used in the field of bearing fault diagnosis. For example, Liu Jun et al. realized multi-domain feature extraction including the original signal and its spectrum and time domain filtered signal by designing a multi-branch CNN and attention mechanism. Xiao Junqing et al. proposed a method that first extracts nonlinear features from fault information through intrinsic time scale decomposition, and then uses CNN to realize bearing fault diagnosis. Jin et al. proposed a fault diagnosis method based on variational mode decomposition and improved convolutional neural network, applied VMD to decompose and calculate the signal, and applied the signal processed by VMD to the convolutional neural network to realize bearing fault diagnosis.

[0005] Although the above-mentioned deep learning-based fault diagnosis method solves the problem that traditional fault diagnosis methods cannot extract deep and rich features, there is still the problem of complex network structure. As the complexity of the network structure increases, the performance of the model will gradually decline.

[0006] In addition, the above method only considers the data flow process of forward propagation and ignores the importance of backpropagation, resulting in serious information loss and affecting the fault diagnosis effect to a certain extent.

[0007] Therefore, how to solve the problem of complex network structure, reduce the amount of calculation, enhance the ability of reverse optimization, and improve the accuracy of bearing fault diagnosis on the basis of extracting deep features is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0008] In view of this, the present invention provides a bearing fault diagnosis method based on IDSP and reverse DSC cascade residual, which can at least partially solve the problems mentioned in the above background technology.

[0009] In order to achieve the above object, the present invention adopts the following technical solution:

[0010] The embodiment of the present invention provides a bearing fault diagnosis method based on IDSP and reverse DSC cascade residual, comprising the following steps:

[0011] S1, obtaining one-dimensional vibration signals from sensors at different positions of the bearing operating platform;

[0012] S2, preprocessing the one-dimensional vibration signal and dividing it into a training sample set and a verification sample set;

[0013] S3, using the training sample set to train the fault diagnosis model, using the verification sample set to verify the accuracy, and obtaining the optimal model; the fault diagnosis model is constructed based on the IDSP and reverse DSC cascade residual model, including: an IDSP shallow feature extraction module, an Inception deep feature extraction module and a reverse DSC cascade residual network module;

[0014] S4. After preprocessing the one-dimensional vibration signal of the bearing to be diagnosed, the signal is input into the optimal model to obtain a corresponding fault diagnosis result.

[0015] Further, the pretreatment comprises:

[0016] S21, performing a normalization operation on the one-dimensional vibration signal;

[0017] S22, converting the one-dimensional signal after the normalization operation into a two-dimensional grayscale image.

[0018] Further, the step S22 specifically includes:

[0019] A sliding window of a preset size is used to perform equal-length interception on the normalized one-dimensional vibration signal;

[0020] Divide the intercepted data segment into n equal parts, each containing n data points;

[0021] Reconstruct n data segments to obtain a two-dimensional grayscale image.

[0022] Furthermore, the step S3 comprises:

[0023] S31, input the preprocessed training sample set into the IDSP shallow feature extraction module to perform feature extraction;

[0024] S32, input the shallow features extracted by IDSP into the Inception deep feature extraction module, combine the channel attention and information reconstruction methods, use a three-layer convolutional network to perform deeper feature learning on the shallow features, and obtain deep features;

[0025] S33, input the extracted deep features into the reverse DSC cascade residual network, and further optimize and enhance the features through the layer-by-layer residual structure, combined with convolution and upsampling operations; reduce information loss through back propagation and improve feature learning efficiency;

[0026] S34. Introduce the focal loss function in the model training process, give higher weights to minority classes and difficult-to-classify samples, and optimize classification performance; use the Softmax function to output the final diagnosis result;

[0027] S35, updating the model parameters through the Adam optimizer;

[0028] S36, by means of cross-validation, verifying the training sample set and the verification sample set under different noise and variable working conditions respectively;

[0029] S37. When the diagnostic accuracy of the model tends to be stable and the error converges, the model convergence is determined, and the optimal model parameters of the final training are saved to obtain the optimal model.

[0030] Furthermore, the step S31 includes:

[0031] S311, taking the preprocessed training sample set as input, using three different scales of convolution of 3×3, 5×5, and 1×1 to perform multi-position and multi-level feature extraction operations;

[0032] S312, normalizing the convolutions of different scales and performing ReLU activation function processing;

[0033] S313, using the channel attention mechanism to increase the weight of key information and reduce the weight of non-key information, so that the channel focuses on key information and eliminates redundant information;

[0034] S314: perform fusion processing on the channel attention output features.

[0035] Furthermore, the step S313 includes:

[0036] Perform global pooling operations on the input feature map, including global average pooling and global maximum pooling, to obtain global statistical information for each channel;

[0037] Input the statistical information into a small fully connected neural network, including two fully connected layers and an activation function Relu, where the activation function Relu is used to generate the attention weight of each channel;

[0038] The attention weights generated by the fully connected network will undergo a normalization operation to ensure that they are between 0 and 1;

[0039] The normalized weights are multiplied element-wise by each channel of the original feature map, thereby adjusting the contribution of each channel, enhancing the feature representation of important channels, and suppressing unimportant channels.

[0040] Furthermore, the step S33 includes:

[0041] Use reverse DSC to reduce the parameters of the cascaded residual structure, and use the DSC step-by-step calculation method to extract channel-by-channel feature information;

[0042] The convolution operation and upsampling operation are combined by using the reverse DSC cascade residual network structure, so that the detail information can be retained in the upsampled feature map, and the information loss is reduced through back propagation, thereby improving the feature learning efficiency.

[0043] Further, the step 35 comprises:

[0044] Initialize the initial value of the Adam optimizer;

[0045] The training samples are input into the model based on IDSP and reverse DSC cascade residuals to learn and diagnose fault characteristics. During the learning process, the Adam optimizer is used to adjust the learning rate of each training process and continuously update the loss value and accuracy.

[0046] It can be seen from the above technical solution that compared with the prior art, the present invention has the following advantages:

[0047] This method uses multi-scale feature extraction, channel attention mechanism, and optimization of the reverse DSC cascade residual network to enable the bearing fault diagnosis model to have higher feature extraction capabilities, stronger noise resistance and generalization capabilities, ensuring that the bearing can maintain high diagnostic accuracy even in complex working conditions and noisy environments; at the same time, it also has fewer parameters and faster diagnostic speed. In addition, by introducing the focal loss function and optimizer, the classification performance of the model in minority class samples and difficult-to-distinguish samples is improved, and the diagnostic efficiency and model stability are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] 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.

[0049] Figure 1 This is the overall framework diagram of the bearing fault diagnosis based on IDSP and reverse DSC cascade residual provided by the present invention.

[0050] Figure 2 A flow chart of a bearing fault diagnosis method based on IDSP and reverse DSC cascade residual provided by the present invention.

[0051] Figure 3 A schematic diagram of an IDSP shallow feature extraction module provided in an embodiment of the present invention.

[0052] Figure 4 A schematic diagram of a reverse DSC cascade residual structure provided in an embodiment of the present invention.

[0053] Figure 5 This is a graph of test data accuracy results under different noise conditions provided by an embodiment of the present invention.

[0054] Figure 6 This is a graph of test data loss values ​​under different noise conditions provided by an embodiment of the present invention.

[0055] Figure 7a A confusion matrix diagram for comparing ablation experiment results provided in an embodiment of the present invention.

[0056] Figure 7b An experimental loss diagram for comparing ablation experiment results provided in an embodiment of the present invention.

[0057] Figure 8a A graph showing the anti-noise performance experimental results of adding 3dB noise to the original vibration signal of a gearbox data set provided by an embodiment of the present invention;

[0058] Figure 8bA graph showing the anti-noise performance experimental results of adding 6dB noise to the original vibration signal of the gearbox data set provided by an embodiment of the present invention;

[0059] Figure 8c A graph showing the anti-noise performance experimental results of adding 9dB noise to the original vibration signal of the gearbox data set provided by an embodiment of the present invention;

[0060] Figure 8d A graph showing the anti-noise performance experimental results of adding 12dB noise to the original vibration signal of the gearbox data set provided by an embodiment of the present invention;

[0061] Figure 9a A schematic diagram of a disorganized and interdependent data set of a gearbox provided by an embodiment of the present invention.

[0062] Figure 9b This is a T-SNE classification result diagram of the gearbox dataset provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] 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.

[0064] The embodiment of the present invention discloses a bearing fault diagnosis method based on IDSP and reverse DSC cascade residual. Figure 1-2 As shown, the following steps are included:

[0065] S1, obtaining one-dimensional vibration signals from sensors at different positions of the bearing operating platform;

[0066] S2. Preprocessing the one-dimensional vibration signal to divide it into a training sample set and a verification sample set;

[0067] S3, using the training sample set to train the fault diagnosis model, using the verification sample set to verify the accuracy, and obtaining the optimal model; the fault diagnosis model is constructed based on the IDSP and reverse DSC cascade residual model, including: an IDSP shallow feature extraction module, an Inception deep feature extraction module and a reverse DSC cascade residual network module;

[0068] S4. After preprocessing the one-dimensional vibration signal of the bearing to be diagnosed, the signal is input into the optimal model to obtain a corresponding fault diagnosis result.

[0069] This fault diagnosis method can achieve a high fault diagnosis accuracy under conditions of variable noise and variable operating conditions, while also having a smaller number of parameters and a faster diagnosis speed.

[0070] The following is a detailed description of each of the above steps:

[0071] S1. Obtain one-dimensional vibration signals from sensors at different positions on the bearing operating platform:

[0072] This step aims to collect raw data, that is, the one-dimensional vibration signals generated during the operation of the bearing. These signals are usually collected by sensors (such as accelerometers) installed at different positions to reflect the operating status of the bearing.

[0073] The data collected by different sensors can capture various characteristic information when a bearing fault occurs, including vibration intensity, frequency, and time domain signal changes.

[0074] S2. Preprocess the one-dimensional vibration signal and divide it into a training sample set and a validation sample set to construct the input data set of the diagnostic model:

[0075] Preprocessing includes normalizing the original one-dimensional vibration signal, that is, standardizing the data to the same numerical range to reduce the impact of amplitude differences in the signal on the model.

[0076] The specific expression is:

[0077]

[0078] Where, X represents the one-dimensional vibration signal after normalization; X max represents the maximum value of the one-dimensional vibration signal in the diagnostic model input data set; X min It represents the minimum value of the one-dimensional vibration signal in the diagnostic model input data set;

[0079] Then, the normalized one-dimensional vibration signal is converted into a two-dimensional grayscale image; specifically, a sliding window of a preset size is used to perform equal-length interception on the normalized one-dimensional vibration signal; the intercepted data segment is divided into n parts, each containing n data points; the n data segments are reconstructed to obtain a two-dimensional grayscale image. For example, first, a sliding window of size 784 is used to perform equal-length interception on the normalized one-dimensional vibration signal, and then the intercepted data is divided into 28 parts, each containing 28 data points, and finally the data segment containing 28 data points is reconstructed to obtain a two-dimensional grayscale image [28, 28]. That is, the features of each time window are taken as a row or column of the matrix, and the values ​​of the two-dimensional matrix are mapped to the grayscale level.

[0080] Finally, the preprocessed one-dimensional signal is divided into a training sample set and a validation sample set. The training sample set is used for model training, and the validation sample set is used for verification and adjustment of model performance.

[0081] S3. Use the training sample set to train the fault diagnosis model, use the verification sample set to verify the accuracy, and obtain the optimal model: wherein the fault diagnosis model includes: IDSP shallow feature extraction module, Inception deep feature extraction module and reverse DSC cascade residual network module.

[0082] Specifically include:

[0083] S31: Input the preprocessed training sample set to the IDSP shallow feature extraction module for feature extraction. The IDSP (Inception-based Deep Shallow Pathway) module can extract key fault features in the input signal through multi-scale convolution.

[0084] S32: Input the shallow features extracted by IDSP into the Inception deep feature extraction module. The Inception module uses a multi-layer convolutional network to perform deeper feature extraction on the shallow features and extract more complex and detailed fault features.

[0085] S33: The extracted deep features are input into the reverse DSC (Depthwise Separable Convolution) cascade residual network, and further optimized through the layer-by-layer residual structure. Combined with convolution and upsampling operations, the feature learning efficiency is optimized and information loss is prevented.

[0086] S34: Introduce the focal loss function in the model training process, adjust the weights of minority classes and difficult-to-classify samples, so that the model can better learn samples with higher classification difficulty and improve classification performance; use the Softmax function to output the final diagnosis result;

[0087] S35: Use the Adam optimizer to update the model parameters. The Adam optimizer can dynamically adjust the learning rate, accelerate the model training process and prevent it from falling into the local optimum.

[0088] S36: Through cross-validation, the training sample set and the validation sample set are validated under different noise and variable working conditions to ensure that the model can perform well in a variety of working environments.

[0089] S37: When the diagnostic accuracy of the model tends to be stable and the error converges, the optimal model parameters after training are saved as the final diagnostic model.

[0090] In this embodiment, the reverse DSC cascade residual network structure is used to improve the back propagation capability while ensuring the lightweight of the network; the introduction of focal loss increases the weights of difficult samples and minority class samples, thereby enhancing the network performance from the learning criterion aspect.

[0091] In the above step S31, the IDSP module structure is as follows Figure 3 As shown in the figure, the structure is mainly divided into two parts. The first part is to extract features from the input fault data through three layers of convolutions of different scales. In order to speed up the training speed of the network and optimize the fault data, the output of each convolution layer is batch normalized and processed by the Relu activation function. Then, the output feature information is subjected to the channel attention mechanism to eliminate redundant information.

[0092] The second part reconstructs the input information through convolution and deconvolution. Information reconstruction can not only reduce noise interference and improve signal quality, but also extract fault signals with stronger time-varying and periodicity, optimize the organizational structure of information, and improve the performance of bearing fault diagnosis.

[0093] In step S32, the reconstructed data and the output of the Inception layer are combined and sent to the global maximum pooling layer, and the global maximum pooling layer reduces the dimension and outputs all the feature information extracted by the IDSP shallow feature extraction module.

[0094] In the above step S33, a cascaded residual network structure is also designed, and a multi-level residual connection method is used to reduce the computational load of the network. DSC is used to reduce the parameters of the cascaded residual structure, and the unique step-by-step calculation method of DSC is used to extract channel-by-channel feature information. The improved structure not only ensures the information fusion between layers, but also the feature parameter amount will not increase sharply to cause model degradation.

[0095] However, the forward residual structure will lose some information features during the forward propagation process, affecting the diagnostic effect. Therefore, the present invention further designs a reverse DSC cascade residual network structure based on the DSC cascade residual structure, which combines the convolution operation with the upsampling operation, so that the detail information can be retained in the upsampled feature map, that is, the information in the original feature map can be restored more accurately, minimizing the loss of information.

[0096] The reverse DSC cascade residual structure is as follows Figure 4 As shown in the figure, this structure can not only map the input data into different feature representations during forward propagation, but also gradually extract higher-level feature representations from different hidden layers, thereby achieving more efficient learning and improving the stability of the network. At the same time, back propagation through the shortcut path of the residual structure is more conducive to reverse learning and avoids the gradient vanishing and gradient exploding problems that are easily caused during back propagation.

[0097] In the above step S34, the model is optimized using the focus loss function. The loss function is proposed by the present invention to solve the problem of low classification accuracy of low-frequency categories in the training process caused by uneven samples and imbalanced categories.

[0098] The cross entropy loss function only obtains the average cross entropy loss of the entire training set by summing or averaging the cross entropy loss of each sample. The focus loss function attenuates the weight of each sample through two adjustable parameters. The degree of attenuation depends on the confidence of the classifier. That is, the greater the confidence of the correctly classified sample, the smaller the weight of the loss function and the smaller the impact on the model; and the lower the confidence of the misclassified sample, the greater the weight of the loss function and the greater the impact on the model.

[0099] The focal loss function makes the model pay more attention to samples that are difficult to classify, thereby improving the classification and generalization capabilities of the model. The specific expression is as follows:

[0100]

[0101] in: is the probability that the i-th sample is predicted to be of category j; n is the number of samples in the training set, is the coefficient of the sample; γ ≥ 0 is a controllable hyperparameter used to balance the loss weight between easy and difficult classification samples.

[0102] In step S35, the initial value of the Adam optimizer is initialized.

[0103] The training samples are input into the model based on IDSP and reverse DSC cascade residuals to learn and diagnose fault characteristics. During the learning process, the Adam optimizer is used to adjust the learning rate of each training process and continuously update the loss value and accuracy.

[0104] In steps S36-S37, the training sample set and the verification sample set are verified under different noise and variable operating conditions by cross-validation; if the diagnostic accuracy of the IDSP and reverse DSC cascade residual model reaches a stable level with the increase of the number of training iterations of the IDSP and reverse DSC cascade residual model, it is determined that the IDSP and reverse DSC cascade residual model has converged; at this time, the training of the IDSP and reverse DSC cascade residual model is terminated, and the optimal diagnostic model parameters are saved.

[0105] In an embodiment of the present invention, the training process flow chart of the bearing fault diagnosis model based on IDSP and reverse DSC cascade residual network is shown in Figure 2 shown.

[0106] Through layer-by-layer extraction and optimization of each step, the model can accurately extract the fault characteristics of the bearing during operation, especially in the case of noise interference and variable working conditions, and can maintain a high diagnostic accuracy. Through the optimization of the focus loss function and the dynamic learning rate adjustment of the Adam optimizer, the classification ability of the model has been significantly improved, and the generalization performance of the model has been ensured through cross-validation.

[0107] S4. After preprocessing the one-dimensional vibration signal of the bearing to be diagnosed, input it into the optimal model to obtain the corresponding fault diagnosis result:

[0108] In this step, the vibration signal of the new bearing to be diagnosed is obtained and preprocessed in the same way as in the training stage (normalization, signal segmentation, etc.). The preprocessed signal is then input into the previously trained optimal model, which extracts and analyzes fault features through IDSP, Inception, and reverse DSC modules, and outputs the corresponding fault diagnosis results.

[0109] When the present invention is implemented, for example, firstly obtain the one-dimensional vibration signal of the bearing operating table from sensors at different positions, construct a data set, select 3000 samples as the training data set, and select 1000 samples as the verification data set, that is, the sample ratio is 3:1.

[0110] After dividing the data set, the training set data is sent to the fault diagnosis model for model training, that is, learning the bearing fault characteristics. As the iterative batches increase, the fault diagnosis accuracy and loss value of the model will change. Through cross-validation, the training sample set and the verification sample set are verified under different noise and variable working conditions respectively; when the fault diagnosis accuracy and loss value converge, the model will immediately end the training and save the best model as the final model.

[0111] After the training is completed, the one-dimensional vibration signal of the bearing that needs to be diagnosed is preprocessed and sent to the saved model for fault diagnosis, and the diagnosis result is output by the Softmax function.

[0112] In order to verify the effectiveness of the method provided by the embodiment of the present invention in terms of fault diagnosis performance, two specific embodiments are used for demonstration. In addition, fault diagnosis experiments under variable working conditions and noise environments are carried out in these two specific embodiments, and the experimental environment is Tensorflow-python.

[0113] Embodiment 1:

[0114] 1.1. Dataset Description:

[0115] The CWRU dataset includes 4 different fault states, records the running data of the motor at 4 speeds (1797, 1772, 1750 and 1730r / min), and covers different load conditions of 0HP, 1HP, 2HP and 3HP. The damage positions of the simulated bearing in actual operation are the bearing rolling element, inner ring, outer ring 3 o'clock, outer ring 6 o'clock, and outer ring 12 o'clock, and the damage diameters are 0.1778, 0.3556, 0.5334 and 0.7112mm respectively. According to the damage state of the bearing, the dataset is divided into 16 states, and the corresponding labels of the damage degree of the normal state and the fault state are shown in Table 1. In this experiment, the data collected from the dataset are divided into training sample data and verification sample data according to a ratio of 3:1.

[0116] Table 1 is a description of the fault types in this dataset.

[0117]

[0118]

[0119] 1.2 Model structure parameters

[0120] This experiment was conducted on a computer with Windows 10 system, Intel Core i5-8300H processor and 8GB memory. The programming language was Python and the deep learning framework was Tensorflow. The value is set to 0.25, the γ value is set to 2 based on experience, the learning rate is 0.001, and to prevent overfitting, the dropout layer weight value is set to 0.5 and the decay rate is set to 0.9. The Adam optimization algorithm is used in the experiment to update the network training parameters in real time. The number of iterations is set to 2200 and the batch number is 64.

[0121] 1.3 Comparison Method

[0122] In order to verify whether the present invention has a good fault diagnosis effect in experiments such as variable noise and variable working conditions, it will be compared with 4 existing excellent methods, namely IRC method, CNN method, ResNet method and VST method. The IRC method consists of an input layer, an improved data pooling layer, a hole residual block, an ordinary residual block, a fully connected layer and a classification layer, the CNN method consists of an input layer, a feedforward denoising CNN, a pooling layer, a fully connected layer and a classification layer, etc., the ResNet method consists of an input layer, a residual structure and a classification layer, etc., and the VST method consists of a VGG-16 network and a Swin-Transformer network, wherein the VGG-16 network is used to extract local feature information, and the Swin-Transformer method is used to extract global feature information, and the features extracted by the two are fused by cross attention.

[0123] 1.4. Variable noise fault diagnosis results and analysis

[0124] The working environment of rolling bearings is complex. The strong noise generated by the friction and vibration between the components and the noise of the surrounding environment seriously affects the health of rolling bearings. Only a fault diagnosis method with extremely strong noise resistance can achieve efficient bearing fault classification. To this end, the present invention conducts a variable noise experiment on the proposed method. The rolling bearing data samples under 1HP load are selected as training samples and verification samples, and different degrees of Gaussian white noise are introduced into the verification samples for interference, and variable noise verification samples with signal-to-noise ratios of 3, 6, 9, and 12dB are obtained respectively.

[0125] Then, the present invention tests the same batch of test data under different noises, and the test accuracy is shown in Figure 5 As shown in Figure 1, when the signal-to-noise ratio is 3dB, the accuracy is 95.13%; when the signal-to-noise ratio is 15dB, the accuracy is 99.42%. Figure 6 As shown, when the iteration batch reaches about 200, the loss value is close to 0. The present invention has a fast convergence speed and a high accuracy, which can indicate that the present invention has a strong anti-noise ability in a variable noise environment.

[0126] 1.5. Variable operating condition fault diagnosis results and analysis

[0127] In actual work, rolling bearings need to frequently change working conditions to meet work requirements, so the bearing diagnosis method needs to have a strong ability to change working conditions. In order to fully verify the superiority of the present invention, the present invention conducts a variable working condition comparison experiment, in which the data samples of 0HP, 1HP, 2HP and 3HP are used as sub-data sets, for example, 0-1, 0-2, 0-3 represent the 0HP sub-data set as the training sample, and the sub-data sets under 1HP, 2HP, 3HP are used as the verification samples, and the experimental numbers under other working conditions are analogous.

[0128] For the same batch of test data under different noises, the variable working condition fault diagnosis results of the six methods are shown in Table 2. It can be clearly seen that the fault diagnosis rate of the method proposed in the present invention is significantly higher than that of the other five methods, and the average fault diagnosis rate reaches 94.31%, which is 2.57, 9.33, 4.01, 2.84, and 1.87 percentage points higher than the five comparison methods, respectively. Compared with the other methods, the diagnostic accuracy of the present invention under different working conditions tends to change smoothly, indicating that the present invention has generalization performance under variable working conditions. The maximum and minimum values ​​of other methods differ greatly, indicating that the diagnostic effect of the model is not stable when facing different working conditions, and the generalization ability is poor.

[0129] Table 2 Fault diagnosis results of various methods under variable working conditions

[0130]

[0131]

[0132] 1.6. Results and analysis of different characteristic signals

[0133] In related technologies, many scholars have converted one-dimensional vibration time series signals into two-dimensional time-frequency images as the input of the model. These conversion methods include continuous wavelet transform (CWT), short-time Fourier transform (STFT), random projection (RP), Gram angular field (GADF), etc. However, this conversion method has two problems: on the one hand, they pay too much attention to the accuracy of the model and ignore the spatial complexity and training time of the model, which is not conducive to the practical application of the model algorithm; on the other hand, some data information will be lost during the conversion process.

[0134] The present invention conducts comparative experiments on 5 different feature forms as inputs of the model of the present invention, and the comparative results are shown in Table 3. It can be seen from the table that while ensuring the fault diagnosis effect, the training time of the one-dimensional vibration time series data completely crushes the training time of the other two-dimensional feature maps as inputs, that is, the method proposed in the present invention is more applicable.

[0135] Table 3 Comparison of training time for different feature map input models

[0136]

[0137]

[0138] 1.7 Ablation Experiment Results and Analysis

[0139] In order to further verify the effectiveness of the main innovation of the method proposed in the present invention for bearing fault diagnosis, the channel attention and information reconstruction module in the IDSP shallow feature extraction module, the reverse DSC cascade residual network structure, and the traditional residual structure are set as variable experiments.

[0140] The fault diagnosis accuracy of the variable noise in the ablation experiment is shown in Table 4. The average fault diagnosis accuracy of the reverse DSC cascade residual structure is only 91.47%. The confusion matrix results are shown in Figure 7a As shown in , it can be seen that the channel attention and information reconstruction modules are important for the extraction of optimized key feature information. When the reverse DSC cascade residual structure is replaced by the traditional residual structure, the fault diagnosis rates in four noise environments are 77.7%, 90.8%, 96.40%, and 97.14%, respectively, which are all inferior to the reverse DSC cascade residual structure; the loss comparison results are shown in Figure 7b As shown, the traditional residual structure has non-convergence and large fluctuations at 3dB and 6dB, while the reverse DSC cascade residual structure has converged to close to 0 after an iteration batch of about 200. Experiments show that the reverse DSC cascade residual structure designed by the present invention has a significant improvement in the overall fault diagnosis effect of the present invention.

[0141] Table 4. Comparison of fault diagnosis accuracy in ablation experiments

[0142]

[0143] Embodiment 2:

[0144] 2.1. Gearbox dataset description:

[0145] In order to further verify the effectiveness of the method proposed in this invention, the experiment also uses the bearing data in the gearbox data set of Southeast University for verification. The data set contains a total of four bearing vibration signals under different operating conditions, namely normal, inner ring fault, outer ring fault, rolling element fault, and inner and outer ring combined fault. Two operating conditions, 20Hz-0V and 30Hz-2V, are set, and a total of 8 channels of data are collected. This experiment uses 2 channels of data.

[0146] This experiment selects 640 samples from each fault type and divides them into training set and validation set in a ratio of 3:1. The distribution details of the dataset are shown in Table 5.

[0147]

[0148] 2.2 Fault diagnosis results and analysis

[0149] The experiments in this part are different from those in the previous one only in the data set, and all other network model parameters are the same.

[0150] In order to further verify the fault diagnosis ability of the present invention in a strong noise environment, this section uses the Southeast University data set to test the model, adding 3, 6, 9, and 12 dB of white noise to the original vibration signal, and the confusion matrix results are shown in Figures 8a-8d As shown in the figure, it can be seen that the performance indicators of the model proposed in the present invention can still maintain a high fault diagnosis rate in a strong noise environment. When the noise intensity is 3dB, the fault diagnosis accuracy, precision, recall rate, and F1-score are 96.50%, 96.57%, 96.50%, and 96.47%, respectively. When the noise intensity reaches 12dB, the four performance indicators are all above 99%, indicating that the model has strong noise resistance and generalization performance.

[0151] 2.3 Visualization of diagnostic results and analysis

[0152] In order to clearly view the fault classification of the present invention, the present invention performs a T-SNE visualization experiment on the test results, such as Figure 9a-9b As shown in the figure, it can be seen that after the feature extraction and classification of the present invention, the disorganized and interdependent data are separated from each other and the same type is aggregated. This shows that the present invention has a strong fault classification ability, and the final classification accuracy rate reaches 99.37%.

[0153] In summary, the present invention proposes a rolling bearing fault diagnosis method based on IDSP and reverse DSC cascade residual, which can solve the problem of low model fault diagnosis effect and poor generalization ability of rolling bearings due to strong noise interference and large changes in operating conditions. It has the following advantages:

[0154] (1) An IDSP shallow feature extraction module was designed to adaptively enhance effective information and suppress interference information from the input two-dimensional data based on the channel attention mechanism, information reconstruction and global maximum pooling layer. The information reconstruction module not only optimized the data structure, but also enhanced the readability, comprehensibility and applicability of the data, avoided information duplication and redundancy, and improved the value of fault information.

[0155] (2) A reverse DSC cascade residual network structure is designed, which not only ensures the lightweight of the network, but also makes the reverse residual structure easier to backpropagate without information loss. The cascade residual structure ensures that each layer of convolution can learn the information features before the input, thereby enhancing the feature learning efficiency of the entire structure.

[0156] (3) The focus loss function is used to measure the prediction accuracy of the classifier. Two variable parameters are added on the basis of the cross entropy loss function. By adjusting the size of the parameters, the focus factor belonging to the network is obtained, the weights of difficult-to-classify samples and minority-class samples are increased, and the weights of easy-to-classify samples are reduced, thereby better handling the problem of class imbalance, optimizing the model parameters, and enhancing the classification effect.

[0157] (4) Multiple experimental verifications were carried out using the CWRU dataset and the Southeast University bearing dataset. The final experimental results show that the fault diagnosis effect of the method proposed in the present invention is good not only in environments with strong noise and variable working conditions, but also has strong generalization performance.

[0158] 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.

[0159] 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 bearing fault diagnosis method based on IDSP and reverse DSC cascade residual, characterized in that: The following steps are involved: S1, obtaining one-dimensional vibration signals from sensors at different positions of the bearing operating platform; S2, preprocessing the one-dimensional vibration signal and dividing it into a training sample set and a verification sample set; S3, using the training sample set to train the fault diagnosis model, using the verification sample set to verify the accuracy, and obtaining the optimal model; the fault diagnosis model is constructed based on the IDSP and reverse DSC cascade residual model, including: an IDSP shallow feature extraction module, an Inception deep feature extraction module and a reverse DSC cascade residual network module; S4. After preprocessing the one-dimensional vibration signal of the bearing to be diagnosed, the signal is input into the optimal model to obtain a corresponding fault diagnosis result.

2. A bearing fault diagnosis method based on IDSP and reverse DSC cascade residual according to claim 1, characterized in that: The pretreatment comprises: S21, performing a normalization operation on the one-dimensional vibration signal; S22, converting the one-dimensional signal after the normalization operation into a two-dimensional grayscale image.

3. A bearing fault diagnosis method based on IDSP and reverse DSC cascade residual according to claim 2, characterized in that: The step S22 specifically includes: A sliding window of a preset size is used to perform equal-length interception on the normalized one-dimensional vibration signal; Divide the intercepted data segment into n equal parts, each containing n data points; Reconstruct n data segments to obtain a two-dimensional grayscale image.

4. A bearing fault diagnosis method based on IDSP and reverse DSC cascade residual according to claim 1, characterized in that: The step S3 comprises: S31, input the preprocessed training sample set into the IDSP shallow feature extraction module to perform feature extraction; S32, input the shallow features extracted by IDSP into the Inception deep feature extraction module, and use a three-layer convolutional network to perform deeper feature learning on the shallow features to obtain deep features; S33, input the extracted deep features into the reverse DSC cascade residual network, and further optimize and enhance the features through the layer-by-layer residual structure, combined with convolution and upsampling operations; reduce information loss through back propagation and improve feature learning efficiency; S34. Introduce the focal loss function in the model training process, give higher weights to minority classes and difficult-to-classify samples, and optimize classification performance; use the Softmax function to output the final diagnosis result; S35, updating the model parameters through the Adam optimizer; S36, by means of cross-validation, verifying the training sample set and the verification sample set under different noise and variable working conditions respectively; S37. When the diagnostic accuracy of the model tends to be stable and the error converges, the model convergence is determined, and the optimal model parameters of the final training are saved to obtain the optimal model.

5. A bearing fault diagnosis method based on IDSP and reverse DSC cascade residual according to claim 4, characterized in that: The step S31 comprises: S311, taking the preprocessed training sample set as input, using three different scales of convolution of 3×3, 5×5, and 1×1 to perform multi-position and multi-level feature extraction operations; S312, normalizing the convolutions of different scales and performing ReLU activation function processing; S313, using the channel attention mechanism to increase the weight of key information and reduce the weight of non-key information, so that the channel focuses on key information and eliminates redundant information; S314: perform fusion processing on the channel attention output features.

6. A bearing fault diagnosis method based on IDSP and reverse DSC cascade residual according to claim 5, characterized in that: The step S313 includes: Perform global pooling operations on the input feature map, including global average pooling and global maximum pooling, to obtain global statistical information for each channel; Input the statistical information into a small fully connected neural network, including two fully connected layers and an activation function Relu, where the activation function Relu is used to generate the attention weight of each channel; The attention weights generated by the fully connected network will undergo a normalization operation to ensure that they are between 0 and 1; The normalized weights are multiplied element-wise by each channel of the original feature map, thereby adjusting the contribution of each channel, enhancing the feature representation of important channels, and suppressing unimportant channels.

7. A bearing fault diagnosis method based on IDSP and reverse DSC cascade residual according to claim 4, characterized in that: The step S33 comprises: Use reverse DSC to reduce the parameters of the cascaded residual structure, and use the DSC step-by-step calculation method to extract channel-by-channel feature information; The convolution operation and upsampling operation are combined by using the reverse DSC cascade residual network structure, so that the detail information can be retained in the upsampled feature map, and the information loss is reduced through back propagation, thereby improving the feature learning efficiency.

8. The bearing fault diagnosis method based on IDSP and reverse DSC cascade residual according to claim 4 is characterized in that: The step 35 comprises: Initialize the initial value of the Adam optimizer; The training samples are input into the model based on IDSP and reverse DSC cascade residuals to learn and diagnose fault characteristics. During the learning process, the Adam optimizer is used to adjust the learning rate of each training process and continuously update the loss value and accuracy.

Images