Bearing fault diagnosis method and system based on GASF-SSM-MSR-FusionNet

Abstract

The invention discloses a bearing fault diagnosis method and system based on GASF-SSM-MSR-FusionNet, belongs to the technical field of fault diagnosis, and solves the problems of weak generalization ability, single feature extraction and low accuracy of a traditional model. The method comprises the steps of obtaining an original vibration signal, dividing a data set, performing multi-dimensional data preprocessing through GASF and SSM, and respectively generating a phase correlation feature matrix and a time structure similarity matrix; the MS-CNN channel extracts multi-granularity features of an SSM matrix through multi-scale convolution, and the ResNet channel utilizes a residual network to mine deep abstract features of a GASF matrix; and after feature fusion, probability distribution of fault categories is output through a full connection layer, and accurate diagnosis is realized. According to the scheme, through multi-modal feature complementation, multi-scale convolution and deep residual learning, the characterization capability of bearing fault features is remarkably improved, the diagnosis accuracy and generalization performance can be effectively improved under complex working conditions, and the method is suitable for the field of intelligent maintenance of industrial equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular relates to a bearing fault diagnosis method and system based on GASF-SSM-MSR-FusionNet. Background Art

[0002] Traditional fault diagnosis methods primarily rely on manual inspections and classic signal processing techniques, such as time-domain analysis and frequency-domain analysis. While these methods can identify faults to a certain extent, they are limited by their ability to extract signal features under complex operating conditions. Consequently, diagnostic results often fall short of ideal performance, especially when faced with nonlinear interference and complex industrial environments, where accuracy and adaptability are limited.

[0003] With the continuous advancement of artificial intelligence technology, especially the successful application of deep learning in various data processing tasks, data-driven fault diagnosis methods are gaining widespread attention. By collecting multidimensional sensor signals from bearings during operation and combining them with the powerful feature extraction and modeling capabilities of deep neural networks, these methods can deeply explore the underlying structural information in complex data, thereby improving the ability to identify fault types, development trends, and potential risks. Data-driven intelligent diagnostic technology not only expands the application boundaries of traditional methods but also provides strong support for achieving higher levels of intelligent operation and maintenance. Summary of the Invention

[0004] The present invention provides a bearing fault diagnosis method and system based on GASF-SSM-MSR-FusionNet to solve the technical problems existing in the prior art, such as low generalization ability, relatively simple feature extraction and low accuracy of traditional bearing fault diagnosis models.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet includes the following steps: Obtain the original vibration signal and divide the data set; The data sets were preprocessed with GASF and SSM respectively; Construct an MSR-FusionNet feature extraction model. The MSR-FusionNet feature extraction model has dual data input channels. The preprocessed GASF and SSM data are input into the dual data input channels of the MSR-FusionNet feature extraction model for feature extraction. Perform feature fusion on the features extracted from the dual data input channels; Perform probability distribution on the categories of the fused features and output the evaluation results based on the probability distribution.

[0006] The above-mentioned GASF data preprocessing of the data set is specifically as follows: normalize the numerical range of the data set to [-1, 1] to form time series data , and the time series data is converted into the angle in the polar coordinate system and the radius . According to the angle in the polar coordinate system , calculate the cosine similarity of each pair of points in the time series to generate the elements in GASF . According to , construct the Gram sum angle field matrix ; the conversion formulas for the angle and the radius in the polar coordinate system are as follows:

[0007]

[0008] In the formula, is the time step, is the total length of the taken time series; the generation formula for the element in GASF is as follows: , in the formula, and are the angles at the corresponding time steps.

[0009] The above-mentioned SSM data preprocessing of the data set is specifically as follows: use a window with a fixed size to divide the data into several windows, calculate the similarity between the windows using cosine similarity, and construct a self-similarity matrix according to the similarity ; the calculation method of cosine similarity is as follows:

[0010] In the formula, is the cosine similarity, and are two adjacent window vectors, is the inner product of the two vectors, and respectively represent the norms of the vectors.

[0011] The above-mentioned MSR-FusionNet feature extraction model includes dual-channel data input, a fusion layer, and a fully connected layer; the dual-channel data input is: the MS-CNN channel and the ResNet channel. The data after data preprocessing is respectively used for feature extraction through the MS-CNN channel and the ResNet channel. The fusion layer fuses the data with features extracted from the two channels respectively, and then the fully connected layer performs the probability distribution of the fusion features in terms of categories, and outputs the evaluation result according to the probability distribution.

[0012] The MS-CNN channel includes a three-layer structure of large-scale convolution layer, multi-scale convolution layer and feature fusion layer to capture feature information of different granularities. The large-scale convolution layer is mainly divided into large-scale two-dimensional convolution, batch normalization and pooling operations; the multi-scale convolution layer contains three types of small-scale convolution branches, medium-scale convolution branches and large-scale convolution branches, which perform three-scale branch parallel calculations on the output of the large-scale convolution layer; the feature fusion layer combines the features of different scales output by the three channels into a single layer. , and Splice and get the fused vector .

[0013] The large-scale convolution layer is used to perform large-scale two-dimensional convolution on the self-similar matrix output after preprocessing the SSM data. Perform preliminary feature extraction and output , the output calculation formula of the large-scale convolutional layer is as follows:

[0014] Where, is the input matrix of the initial large convolutional layer, is the convolution kernel weight, is the bias, m, n are the positions corresponding to the weight matrix; the ReLU activation function is used to activate the output of the large-scale convolution layer Perform batch normalization and activation, output after batch normalization and activation as follows:

[0015] Finally, the output after batch normalization and activation The maximum pooling operation is used, and its calculation method is as follows:

[0016] Where, is the defined output of the pooling operation, and is the width and height of the pooling window.

[0017] The multi-scale convolution layer includes three types of small-scale convolution branches, medium-scale convolution branches and large-scale convolution branches. The three-scale branches are composed of three convolution blocks. Each convolution block includes a layer of convolution, a layer of pooling and an activation operation in sequence. The output of the large-scale convolution layer is After parallel computation of three-scale convolution branches, features of different scales are output. , and .

[0018] The feature fusion layer integrates feature information at different scales. The feature fusion layer combines the features of different scales output in three channels , and to obtain a fused vector .

[0019] The ResNet channel is mainly divided into three layers, namely the initial convolutional layer, the residual layer, and the global average pooling layer; the data preprocessed by GASF enters the large convolutional kernel of the initial convolutional layer, and its output channel number is 64, which is used for preliminary feature extraction. Then, batch normalization and activation are performed on the output of the initial convolutional layer, and finally, it is the max pooling layer to further reduce the size of the feature map; the residual layer contains 6 residual blocks, each residual block contains two convolutional layers with a convolutional kernel size of 3×3, and the ReLU activation function is selected. The output channel numbers are 64, 128, 256, 512, 512, and 512 in sequence. By gradually increasing the channel number layer by layer, deep features are extracted; the global average pooling layer flattens the deep feature data extracted by the residual layer into .

[0020] A bearing fault diagnosis system based on GASF-SSM-MSR-FusionNet includes a data processing module, a feature extraction module, a feature fusion module, and an evaluation module; The data processing module is used to divide the original vibration signal into a data set, and perform GASF data preprocessing and SSM data preprocessing on the data set respectively; The feature extraction module is used to construct an MSR-FusionNet feature extraction model. The MSR-FusionNet feature extraction model has a dual-data input channel. The data preprocessed by GASF and SSM are respectively input into the dual-data input channel of the MSR-FusionNet feature extraction model for feature extraction; The feature fusion module is used to fuse the features extracted by the dual-data input channels respectively; The fault diagnosis module is used to perform a probability distribution on the fused features in terms of categories, and output an evaluation result according to the probability distribution to generate fault diagnosis data.

[0021] Compared with the prior art, the present invention has the following beneficial effects: A bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet disclosed by the present invention adopts GASF preprocessing to convert the time series into angle and radius information in the polar coordinate system, constructs a matrix through cosine similarity, captures the dynamic correlation and periodic characteristics of the signal in the phase space, and retains the time series dependence and non-linear structure. The normalization operation therein maps the numerical range to [-1, 1], avoids the interference of data scale differences on the model, and improves the feature stability. SSM is used for data preprocessing. By fixed window segmentation and cosine similarity calculation, the similarity of different time windows is quantified, the long-term dependence relationship and structural repeatability of the signal are captured, and the gradual change characteristics or periodic fault modes of bearing faults are effectively characterized. The above two data preprocessing modes convert the original vibration signal from different dimensions to form complementary two-dimensional matrix features, providing multi-modal input for the subsequent model and avoiding the loss of feature information caused by single preprocessing.

[0022] The representation ability of the model is enhanced through multi-scale and deep feature extraction. A two-channel feature extraction model is constructed. In the MS-CNN channel, the large-scale convolutional layer performs preliminary global feature extraction on the SSM matrix to capture the macroscopic patterns of self-similar structures. The multi-scale convolution branch performs parallel calculations through small, medium, and large-scale convolution kernels to extract detailed features (small scale), local structures (medium scale), and global contours (large scale) respectively, realizes the fusion of features with different granularities, and avoids the insufficient representation of complex fault features by single-scale convolution. The feature fusion layer splices the multi-scale outputs to form a comprehensive feature vector containing multi-granularity information, improving the ability to capture multi-dimensional features of bearing faults. In the ResNet channel, the residual block is used to solve the problem of gradient disappearance in the deep network. The number of channels is gradually increased through 6 residual layers to extract deep abstract features and capture the non-linear complex patterns of bearing faults. Global average pooling avoids the overfitting risk of the fully connected layer and enhances the generalization ability of the model to data under different working conditions.

[0023] The two-channel feature extraction model fuses the phase correlation features after GASF preprocessing and the time structure features after SSM preprocessing through independent extraction in the two channels, realizes the complementarity of "time domain structure information + frequency domain phase information", and avoids the dependence on specific features in a single channel. The fused features contain multi-modal, multi-scale, and multi-level information, more comprehensively represent the health state of the bearing, and reduce the classification confusion caused by single features. The fully connected layer outputs the category probability distribution based on the fused features, and combines activation functions such as Softmax to achieve accurate classification of fault categories and reduce the misjudgment problem caused by insufficient features in traditional models.

[0024] The combination of data preprocessing and the dual-channel network enables the model to adapt to the vibration signal differences under different working conditions, avoid overfitting of the traditional model to a specific dataset, and improve the cross-scenario diagnosis ability. The application of the residual network and batch normalization further stabilizes the training process and reduces performance fluctuations caused by changes in data distribution. Description of the Drawings

[0025] Figure 1 : Architecture diagram of the bearing fault diagnosis system based on GASF-SSM-MSR-FusionNet; Figure 2 : Flowchart of the bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet; Figure 3 : Module diagram of the bearing fault diagnosis system based on GASF-SSM-MSR-FusionNet; Figure 4 : Operating result diagram of the bearing fault diagnosis system based on GASF-SSM-MSR-FusionNet; Figure 5 : Schematic diagram of the accuracy result of the GASF-ResNet model; Figure 6 : Schematic diagram of the accuracy result of the SSM-MSCNN model; Figure 7 : Schematic diagram of the comparison of the accuracy of the GASF-ResNet model, SSM-MSCNN model and the validation set of the system of the present invention; Figure 8 : Schematic diagram of the final loss curve and accuracy curve of the bearing fault diagnosis system based on GASF-SSM-MSR-FusionNet; Figure 9 : Schematic diagram of the operating result of the GASF-ResNet model; Figure 10 : Schematic diagram of the operating result of SSM-MS; Figure 11 : Schematic diagram of the comparison of the accuracy of the GASF-ResNet model, SSM-MSCNN model and the validation set of the system of the present invention. Detailed Embodiments

[0026] To further understand the content of the present invention, the following describes the present invention in detail with reference to the drawings and specific embodiments. It should be understood that the embodiments are only for explaining the present invention and not for limiting it.

[0027] In the following text, only some exemplary embodiments are briefly described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the present invention. Therefore, the accompanying drawings and the description are considered to be exemplary in nature and not restrictive.

[0028] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0029] Embodiment 1 A bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet of the present invention, the implementation method process is as Figure 2 , including the following steps: Step 1: Obtain the original vibration signal.

[0030] Step 2: Divide the original vibration signal into data sets, including a training set, a validation set, and a test set.

[0031] Step 3: Preprocess the data sets using GASF and SSM respectively.

[0032] Step 3.1: Use GASF to capture the integrity and periodicity characteristics of the signals in the data set, and convert the one-dimensional signal into a two-dimensional Gram sum angular field matrix.

[0033] Step 3.2: Use SSM to mine the similarity characteristics of the signals between different time periods in the data set, and convert the one-dimensional vibration data into a two-dimensional self-similarity matrix.

[0034] Step 4: Construct an MSR-FusionNet feature extraction model, and use the training set data to train the MSR-FusionNet feature extraction model.

[0035] Step 4.1: Input the preprocessed SSM data into the MS-CNN channel in the MSR-FusionNet feature extraction model, and sequentially perform global feature extraction by the large-scale convolutional layer, different-scale feature extraction by the multi-scale convolutional layer, and feature fusion and flattening by the fusion layer.

[0036] Step 4.2: Input the preprocessed GASF data into the ResNet channel in the MSR-FusionNet feature extraction model, and sequentially perform global feature extraction by the large-scale convolutional layer, deep feature extraction by the 6-layer Resnet convolutional block, and flattening operation.

[0037] Step 5: The fusion layer performs feature fusion on the features extracted from the MS-CNN channel and the ResNet channel.

[0038] Step 6: The fully connected layer classifies and evaluates the fused features.

[0039] Step 7: Input the validation set into the MSR-FusionNet feature extraction model to verify the accuracy of the model.

[0040] Step 8: Set the iteration threshold. Whether the actual number of iterations is greater than or equal to it. If satisfied, output the optimal model. If not satisfied, continue to iterate.

[0041] Step 9: Input the test set of the original vibration signal into the MSR-FusionNet feature extraction model for evaluation and output the diagnostic result.

[0042] Based on the above Figure 2 For the bearing fault diagnosis method process described above, a bearing fault diagnosis system based on GASF-SSM-MSR-FusionNet proposed in this embodiment has the following specific implementation method: Obtain the original vibration signal and divide the original vibration signal into a training set, a validation set, and a test set; Use SSM (Self-Similarity Matrix) and GASF (Gramian Angular Summation Field) to perform data preprocessing on all original vibration signals. SSM reveals the internal patterns and characteristics of the data by capturing the similarity between data at different time periods or features; the core idea of GASF is to encode time series data in the polar coordinate system to generate a two-dimensional image with a time series relationship.

[0043] Among them, the data preprocessing method of SSM for the original vibration signal is as follows: Segment the long-time time series signal, that is, divide the data using a fixed-size window. Divide the original time series signal with a length of According to the fixed window length For division, each sub-window contains k Time points, that is: 。 。

[0044] Calculate the similarity between sub-windows and select the cosine similarity As the similarity measurement method, as shown in Equation (1): (1) Among them, And Are two adjacent window vectors, Is the inner product of the two vectors, And respectively represent the norms of vectors.

[0045] Construct a self-similarity matrix according to the similarity . As shown in Equation (2): (2) The data preprocessing method of GASF for the original vibration signal is as follows: Normalize the numerical range of the data set to [-1, 1] to form time series data , ensuring the stability of the data in polar coordinate transformation.

[0046] Convert the normalized time series data into the angle in the polar coordinate system and the radius , the angle in the polar coordinate system and the radius The conversion formulas are shown in the following formulas (3) and (4): (3) (4) Among them, is the time step, is the total length of the time series taken.

[0047] Calculate the Gram matrix, generate the elements in GASF by calculating the cosine similarity of each pair of points in the time series , as shown in Equation (5): (5) Among them, and are the angles at the corresponding time steps.

[0048] Construct the Gram sum angle field matrix : (6) Construct a feature extraction model of MSR-FusionNet (Multi-Scale ResNet Fusion Network, MSR-FusionNet, multi-scale fusion residual network extracts features) with dual-channel data input, and input the training set of the original vibration signal into the MSR-FusionNet feature extraction model for model training. The structure of the constructed MSR-FusionNet feature extraction model is as Figure 1 shown, the data input adopts dual-channel input, and the preprocessed data is respectively passed through the MS-CNN channel (such as Figure 1 the Sin input data channel shown inFigure 1 Perform feature extraction on the Gin input data channel shown in the figure, and combine multi-scale feature extraction with the residual learning mechanism to improve the accuracy and robustness of fault diagnosis.

[0049] The MS-CNN channel includes three-layer structures: a large-scale convolutional layer, a multi-scale convolutional layer, and a feature fusion layer, to capture feature information of different granularities, thereby enhancing the model's ability to recognize complex fault patterns.

[0050] The large-scale convolutional layer is like Figure 1 Conv ReLU Maxpool shown in the figure. The large-scale convolutional layer is mainly divided into large-scale two-dimensional convolution, batch normalization, and pooling operations. The large-scale two-dimensional convolution is used to perform preliminary feature extraction on the self-similar matrix output after preprocessing the SSM data. Using a large convolutional kernel can capture the global information in the self-similar matrix of the input data. The output of the large-scale two-dimensional convolutional layer The output of the large-scale convolutional layer is calculated as shown in formula (7): (7) Where is the input matrix of the initial large convolutional layer, is the convolutional kernel weight, is the bias, and m, n are the positions corresponding to the weight matrix.

[0051] Use the ReLU activation function to perform batch normalization and activation on the output of the large-scale convolutional layer to stabilize the feature distribution and accelerate training convergence. The output after batch normalization and activation is as follows in formula (8): (8) Finally, perform max pooling operation on the output after batch normalization and activation to reduce the size of the feature map, lower the computational complexity, and retain key features at the same time. Define the output as follows in formula (9): (9) Where and are the width and height of the pooling window.

[0052] ​The multi-scale convolutional layer includes three types: a small-scale convolutional branch, a medium-scale convolutional branch, and a large-scale convolutional branch, which perform parallel computational branching of three scales on the output of the large-scale convolutional layer. Taking the small-scale convolutional branch as an example, its structure consists of three convolutional blocks. Each convolutional block sequentially includes a convolutional layer, a pooling layer, and an activation operation inside. The structures of the medium-scale convolutional branch and the large-scale convolutional branch are the same as that of the small-scale convolutional branch. The After being processed by the parallel computational branching of the convolutional branches of the three scales, features of different scales are output , and .

[0053] The third layer is the feature fusion layer. The feature fusion layer concatenates the features of different scales output in three channels , and to obtain a fused vector . The feature fusion layer integrates the feature information at different scales, enriches the feature representation, and improves the model's comprehensive understanding and classification performance of fault patterns.

[0054] The ResNet channel is mainly used to extract deep and high-level feature representations from the Gram sum angular field matrix after the preprocessing of GASF data. The ResNet channel is mainly divided into three layers, namely the initial convolutional layer (Conv ReLU), the residual layer (Resnet Blockn), and the global average pooling layer.

[0055] The input data, the Gram sum angular field matrix of the ResNet channel first enters the large convolutional kernel of the initial convolutional layer. Its output channel number is 64, which is used for preliminary feature extraction. Then, batch normalization and activation are performed on the output of the initial convolutional layer. Finally, it is the max pooling layer, which further reduces the size of the feature map. The structure of the initial convolutional layer is the same as that of the large-scale convolutional layer of the MS-CNN channel and will not be elaborated here.

[0056] The residual layer contains 6 residual blocks, such as Figure 1 shown as Resnet Block1~Resnet Block6. Each residual block contains two convolutional layers with a convolutional kernel size of 3×3. The ReLU activation function is selected, and the output channel numbers are 64, 128, 256, 512, 512, 512 in sequence. By gradually increasing the channel number, deep features are extracted. The residual learning mechanism of the residual layer can effectively alleviate the gradient vanishing problem in the deep network, promote the learning of deep features, and improve the deep feature mining ability and training stability of the model.

[0057] The global average pooling layer flattens the deep feature data extracted by the residual layer into It effectively retains the global feature information of each channel, ensures that important fault mode features are not lost, while significantly reducing the computational complexity of the subsequent fully connected layer and improving the computational efficiency of the model, preparing for subsequent feature splicing.

[0058] Input the multi-scale features extracted by the MS-CNN channels and the deep features extracted by the ResNet channels into the fusion layer. The fusion layer is the Fully Connected Layer as described in Figure 1 . In the fusion layer, the multi-scale features and the deep features are sequentially spliced to enhance the richness and diversity of the model's feature expression.

[0059] Input the fused features into the fully connected layer. The fully connected layer combines the Softmax function to convert the original classification scores into probability values for each category, and finally obtains the probability distribution of each input sample for each category.

[0060] Embodiment 2 As shown in Figure 3 , this embodiment proposes a bearing fault diagnosis system based on GASF-SSM-MSR-FusionNet, which mainly includes a data processing module, a feature extraction module, a feature fusion module, and a fault diagnosis module; it can implement a bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet described in Embodiment 1. The data processing module is used to divide the dataset of the original vibration signal, perform GASF data preprocessing on the dataset to generate a Gram sum angular field matrix , perform SSM data preprocessing on the dataset to generate a self-similarity matrix . The feature extraction module is used to construct an MSR-FusionNet feature extraction model. The MSR-FusionNet feature extraction model has a dual data input channel, and the data after GASF and SSM data preprocessing are respectively input into the dual data input channels of the MSR-FusionNet feature extraction model for feature extraction. The feature fusion module is used to fuse the features extracted by the dual data input channels respectively. The fault diagnosis module is used to perform a probability distribution of the fused features for each category, and output an evaluation result according to the probability distribution to form fault diagnosis data.

[0061] Embodiment 3 Training is carried out using the CWRU public dataset provided by the Department of Mechanical Engineering, Case Western Reserve University (CWRU), Ohio, USA. The CWRU dataset is a commonly used dataset widely applied in the fields of fault diagnosis and predictive maintenance. It is based on bearing vibration tests in a laboratory environment and provides rich data resources. Among the relevant components of the test bench, the bearing model used is SKF6205. The sampling frequencies include 12 kHz and 48 kHz, and the motor speeds include 1797 rpm, 1772 rpm, 1750 rpm, and 1730 rpm. The dataset covers four bearing operating states: normal state, inner race fault, outer race fault, and ball fault. For each fault type, three different diameters of fault points, namely 0.007 inches, 0.014 inches, and 0.021 inches, are considered respectively, totaling 10 fault types.

[0062] The experimental data information is shown in Table 1. First, the original data is sampled using a sliding window, with a time step set to 256 and an overlap rate of 0.5. Then, the sampled data is preprocessed separately using SSM and GASF, and the preprocessed dataset is divided into a training set, a validation set, and a test set. The training set accounts for 70%, providing sufficient data volume for the model to fully learn the complex patterns and features of bearing faults, thus constructing a robust prediction ability. The validation set accounts for 20% and is used to adjust the model hyperparameters and monitor the overfitting risk during the training process. This proportion ensures a sufficient sample size to reliably optimize the model configuration. The test set accounts for 10% and is used to finally evaluate the generalization ability of the model on unseen data, while avoiding wasting too much data for training, with the proportions being 0.7, 0.2, and 0.1 respectively. Finally, the divided dataset is used to train and test the model.

[0063] Table 1

[0064] The computer configuration used in the experiment is CPU i9 13900 with 16GB of memory. The learning rate is set to 0.0001, the batch size is 32, the optimizer used is Adam, the loss function is the cross-entropy loss function, the number of epochs is set to 100 rounds, and 1000 samples of each fault category are set, with each sample containing 512 data points. The relevant parameters of the model are shown in Table 2.

[0065] Table 2

[0066] 1. Fault diagnosis results First, verify the effectiveness of the GASF-SSM-MSR-FusionNet system. Through training, the final running results of the model are asFigure 4 As shown by Figure 4 -a) and the loss curve and accuracy curve in 4-b), it can be seen that the convergence speed of the training set of the model is higher than that of the validation set. The accuracy of the training set reached 97.46% in the 3rd round, and the accuracy of the validation set reached 98.50% in the 22nd round. The confusion matrix of the model is as Figure 4 -c) shows that the accuracy of samples of type 1 is 96.00%, the classification accuracy of samples of type 4 is 98.00%, and the classification accuracy of samples of other types is 100%, with a relatively high overall diagnostic accuracy. As Figure 4 -d), it can be seen from the visualization graph of t-distributed Stochastic Neighbor Embedding (t-SNE) that the data of ten categories can be clearly distributed in the two-dimensional space, demonstrating the high efficiency of the GASF-SSM-MSR-FusionNet model in classification.

[0067] 2. Comparative Experiments To further analyze the model performance, the GASF-ResNet model and the SSM-MSCNN model were designed for comparison with the model in this paper, and the three models were trained 10 times under the same conditions. The experiments show that the average accuracy of the GASF-SSM-MSR-FusionNet model is 99.70%, the average accuracy of the GASF-ResNet model is 99.20%, and the average accuracy of the SSM-MSCNN model is 96.30%. The lowest accuracy results in the 10 runs were compared. The accuracy running results of the GASF-ResNet model are as Figure 5 -b) shows that the model accuracy gradually increased in the first 20 rounds, then fluctuated, and began to converge around the 40th round. The final validation set accuracy of the model is 98.60%. The confusion matrix of the GASF-ResNet model Figure 5 -c) shows that the classification accuracy of 5 types of the model is 100%. Among the misclassified samples, the accuracy of types 0 and 5 is the highest, at 99.00%, and the accuracy of type 4 is the lowest, at 94.00%. From Figure 5 -d) of the t-SNE graph, it can be seen that there is a certain degree of overlap among types 0, 1, and 4 in the feature space, and the feature distributions of other categories are relatively clear, with obvious inter-class boundaries.

[0068] The accuracy running results of the SSM-MSCNN model are as Figure 6 shown: As shown by Figure 6-a) and 6-b) show that the convergence rates of the training set and validation set of the model are both slow, and the accuracy is low. The accuracy of the training set stabilizes after the 10th round, and the accuracy of the validation set stabilizes after the 31st round. The final validation set accuracy of the SSM-MSCNN model is 96.40%; From Figure 6 -c)'s test set confusion matrix, it can be seen that the model has a classification accuracy of 100% for 5 types. Among the misclassified samples, the accuracies of type 1 and type 6 are the highest, at 96.00%, and the accuracy of type 4 is the lowest, at 84.00%. From Figure 6 -d)'s t-SNE graph, it can be seen that there is a certain degree of overlap in the features of type 1 and type 4, resulting in misclassification. The feature distributions of other categories are relatively clear, and the inter-class boundaries are relatively obvious, with good classification effects.

[0069] Compare the validation set accuracies of the three models, as Figure 7 shown. In terms of the number of iteration rounds, the GASF-SSM-MSR-FusionNet model gradually converges after 20 rounds of operation, the GASF-ResNet model gradually converges after 40 rounds of operation, and the SSM-MSCNN model gradually converges after 31 rounds of operation. The GASF-SSM-MSR-FusionNet model has fewer iteration times. In terms of accuracy, the accuracy of the GASF-SSM-MSR-FusionNet model reaches 99.60%, the accuracy of the GASF-ResNet model reaches 98.60%, and the accuracy of the SSM-MSCNN model reaches 96.40%. The accuracy of the GASF-SSM-MSR-FusionNet model is 1.01% higher than that of the GASF-ResNet model and 3.32% higher than that of the SSM-MSCNN model, further demonstrating the superiority of the combined model in terms of accuracy.

[0070] To further analyze the model performance, compare the test set accuracy, precision, recall, and F1 value of the three models, as shown in Table 3. It can be seen that the four indicators of the GASF-SSM-MSR-FusionNet model are the highest, followed by the GASF-ResNet model, and the four indicators of the SSM-MSCNN model are the lowest. The test set accuracy of the GASF-SSM-MSR-FusionNet model is 0.91% higher than that of the GASF-ResNet model and 3.97% higher than that of the SSM-MSCNN model. This further shows that the feature extraction effect of the residual network is more superior to that of the ordinary convolutional network.

[0071] Table 3

[0072] 3. Comparison with Literature Methods To further verify the effectiveness of the GASF-SSM-MSR-FusionNet model, under the condition of the same dataset, the model designed in this chapter is compared with the MPGCN-Resnet model, WPT-CNN-SA model, DenseNet-121 model, and SDP-MSCNN-LSTM model proposed in the literature, as shown in Table 4. It can be seen that the model accuracies from high to low are: GASF-SSM-MSR-FusionNet model, WPT-CNN-SA model, SDP-MSCNN-LSTM model, MPGCN-Resnet model, DenseNet-121 model. It can be seen that the GASF-SSM-MSR-FusionNet model has the highest accuracy and the best diagnostic effect.

[0073] Table 4

[0074] 4. Analysis of Results of Self-built Dataset To further verify the effectiveness of the model under actual working conditions, a fault diagnosis experimental platform is used for verification. The self-built experimental platform sets the sampling frequency to 12 kHz and the rotational speed to 1700 rpm. The fault types collected for the bearing include inner race fault, outer race fault, ball fault, and normal bearing. Each fault type contains vibration data in three directions: X, Y, and Z. The experimental data information is shown in Table 5.

[0075] Table 5

[0076] First, verify the effectiveness of the GASF-SSM-MSR-FusionNet system in fault diagnosis under normal working conditions. Through training and optimization, the final loss curve and accuracy curve are as Figure 8 shown. As can be seen from Figure 8 Figure 8-a) and 8-b), the convergence speed of the training set of the model is higher than that of the validation set. The accuracy of the training set reached 93.80% in the 8th round and 98.75% in the 20th round, while the accuracy of the validation set reached 84.92% in the 7th round and 98.39% in the 23rd round, and then fluctuated slightly around 99%. The confusion matrix of the model is as Figure 8 shown in Figure 8-c). The classification accuracy of 5 types of the model is 100%. Among the remaining misclassified samples, the highest classification accuracy is 98.00% and the lowest is 96.00%. As Figure 8 shown in Figure 8-d), it can be seen from the t-SNE graph that except for a small amount of feature overlap, the data of twelve categories can be clearly distributed in the two-dimensional space, and the overall diagnostic accuracy is relatively high and the diagnostic effect is good.

[0077] 5. Comparative Experiment on Self-built Datasets To further analyze the model performance, the GASF-ResNet model and the SSM-MSCNN model were designed for comparison with the model in this paper, and the three models were trained 10 times under the same conditions. Experiments show that the average accuracy of the GASF-SSM-MSR-FusionNet model is 99.20%, the average accuracy of the GASF-ResNet model is 98.10%, and the average accuracy of the SSM-MSCNN model is 95.10%. The lowest accuracy results in the 10 runs were compared. The running results of the GASF-ResNet model are as Figure 9 shown. As can be seen from Figure 9 -a), the loss of the training set of the model decreases faster than that of the validation set. The loss of the validation set fluctuates significantly in the first 20 rounds and then becomes relatively stable, showing an overall downward trend. As shown in Figure 9 -b), the accuracy of the model gradually increases in the first 20 rounds and begins to converge around the 30th round. The final validation set accuracy of the model is 97.90%. The confusion matrix of the GASF-ResNet model Figure 9 -c) shows that the classification accuracy of 5 types of the model is 100%. Among the misclassified samples, the accuracies of type 1, type 2, and type 3 are the highest, at 98.00%, and the accuracy of type 4 is the lowest, at 90.00%. As can be seen from the t-SNE diagram in Figure 9 -d), there is a certain degree of overlap of type 4, type 5, and type 6 in the feature space. The inter-class boundaries of other classes are relatively obvious, and the classification effect is good.

[0078] The accuracy running results of the SSM-MSCNN model are as Figure 10 shown. As can be seen from Figure 10 -a), the loss of the training set of the model decreases faster than that of the validation set, and there are certain fluctuations in the loss of the validation set, showing an overall downward trend. As can be seen from Figure 10 -b), the convergence speeds of both the training set and the validation set of the model are relatively slow, and the accuracy is low. The accuracy of the training set tends to be stable after the 40th round, and the accuracy of the validation set tends to be stable after the 50th round. The final validation set accuracy of the SSM-MSCNN model is 94.70%; as can be seen from Figure 10 -c), the classification accuracy of 5 types of the model is 100%. Among the misclassified samples, the accuracies of type 4, type 10, and type 11 are the highest, at 98.00%, and the accuracy of type 7 is the lowest, at 73.00%. As can be seen from Figure 10From the t-SNE plot in (d), it can be seen that there is a certain degree of overlap among Type 5, Type 3, and Type 7 in the feature space, and the class distances are relatively close. The feature distributions of other classes are clear, and the classification effect is good.

[0079] Compare the validation set accuracies of the three models as Figure 11 shown. In terms of the number of iteration rounds, the GASF-SSM-MSR-FusionNet model gradually converges after 30 rounds of operation, the GASF-ResNet model gradually converges after 30 rounds of operation, the SSM-MSCNN model gradually converges after 40 rounds of operation, and the GASF-SSM-MSR-FusionNet model has fewer iteration times. In terms of accuracy, the accuracy of the GASF-SSM-MSR-FusionNet model reaches 98.80%, the accuracy of the GASF-ResNet model reaches 97.90%, and the accuracy of the SSM-MSCNN model reaches 94.70%. The accuracy of the GASF-SSM-MSR-FusionNet model is 0.92% higher than that of the GASF-ResNet model and 4.33% higher than that of the SSM-MSCNN model, further proving the superiority of the combined model in terms of accuracy.

[0080] To further analyze the model performance, compare the test set accuracy, precision, recall, and F1-score of the three models, as shown in Table 6. It can be seen that the four indicators of the GASF-SSM-MSR-FusionNet model are the highest, followed by the GASF-ResNet model, and the four indicators of the SSM-MSCNN model are the lowest. The test set accuracy of the GASF-SSM-MSR-FusionNet model is 1.13% higher than that of the GASF-ResNet model and 4.78% higher than that of the SSM-MSCNN model. This further shows that the feature extraction effect of the residual network is more superior to that of the ordinary convolutional network.

[0081] Table 6

[0082] In response to the problems of insufficient feature extraction and low model accuracy in fault diagnosis methods, this paper proposes a bearing fault diagnosis model based on GASF-SSM-MSR-FusionNet. First, the two data preprocessing techniques of GASF and SSM are combined to capture the periodic features in the vibration signal and the similarity between features. Secondly, the MSR-FusionNet dual-channel input structure is adopted. Through deep convolution and multi-scale feature fusion, the model can deeply explore the deep features in the signal and realize the efficient fusion of multi-scale information through the feature integration layer, which significantly improves the accuracy and robustness of fault diagnosis. Finally, experimental verification is carried out using public datasets and self-built datasets. The results show that the average accuracy of the model in this chapter on the public dataset is 99.70%, and the average accuracy on the self-built dataset is 99.20%. Compared with traditional models and models in the literature, the diagnostic accuracy of GASF-SSM-MSR-FusionNet is better than other models, which verifies the effectiveness and stability of the model in bearing fault diagnosis.

[0083] In addition, it should be understood that although this specification describes the embodiments, not every embodiment contains only one independent technical solution. This description is for clarity only. Those skilled in the art should consider the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art. The above content is only for the purpose of illustrating the technical concept of the present invention and cannot be used to limit the scope of protection of the present invention. Any changes made based on the technical solution in accordance with the technical concept proposed by the present invention fall within the scope of protection of the claims of the present invention.

Claims

1. A bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet, characterized in that It includes the following steps: Obtain the original vibration signal and divide the data set; Perform GASF and SSM data preprocessing on the data set respectively; Construct an MSR-FusionNet feature extraction model. The MSR-FusionNet feature extraction model has dual data input channels. The data after GASF and SSM data preprocessing are respectively input into the dual data input channels of the MSR-FusionNet feature extraction model for feature extraction; Perform feature fusion on the features extracted from the dual data input channels respectively; Perform probability distribution on the fused features in terms of categories, and output the evaluation result according to the probability distribution to generate fault diagnosis data.

2. The bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet according to claim 1, wherein The GASF data preprocessing is performed on the dataset respectively, specifically: normalizing the numerical range of the dataset to [-1, 1] to form time series data , and converting the time series data into the angle and radius in the polar coordinate system. According to the angle in the polar coordinate system, calculate the cosine similarity of each pair of points in the time series to generate the elements in GASF . According to , construct the Gram sum angle field matrix ; The conversion formulas for the angle and radius in the polar coordinate system are as follows: wherein, is the time step, is the total length of the selected time series; The elements in GASF are generated according to the following formula: , wherein, and are the angles corresponding to the time step.

3. A bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet according to claim 1, characterized in that, The SSM data preprocessing of the data set is specifically as follows: the data is segmented into several windows by using a window with a fixed size, the similarity between the windows is calculated by using the cosine similarity, and a self-similarity matrix is constructed according to the similarity ; The calculation method of the cosine similarity is as follows: Wherein, is the cosine similarity, and are two adjacent window vectors, is the inner product of the two vectors, and respectively represent the norms of the vectors.

4. A bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet according to claim 1, characterized in that, The MSR-FusionNet feature extraction model includes dual-channel data input, a fusion layer, and a fully connected layer; the dual-channel data input is: the MS-CNN channel and the ResNet channel. The data after data preprocessing are respectively used for feature extraction through the MS-CNN channel and the ResNet channel. The fusion layer fuses the data with features extracted from the dual channels respectively, and then the fully connected layer performs probability distribution on the fused features in terms of categories, and outputs the evaluation result according to the probability distribution.

5. A bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet according to claim 4, characterized in that, The MS-CNN channel includes three-layer structures: a large-scale convolutional layer, a multi-scale convolutional layer, and a feature fusion layer, to capture feature information of different granularities. The large-scale convolutional layer is mainly divided into large-scale two-dimensional convolution, batch normalization, and pooling operations; the multi-scale convolutional layer contains three branches: a small-scale convolution branch, a medium-scale convolution branch, and a large-scale convolution branch, which respectively perform three-scale branch parallel computations on the output of the large-scale convolutional layer; the feature fusion layer splices the features of different scales output by the three channels to obtain a fused vector , and to obtain a fused vector .

6. The bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet according to claim 5, characterized in that The large-scale two-dimensional convolution of the large-scale convolutional layer is used for the self-similarity matrix output after preprocessing the SSM data to perform preliminary feature extraction and output . The output calculation formula of the large-scale convolutional layer is as follows: In the formula, is the input matrix of the initial large convolutional layer, is the convolutional kernel weight, is the bias, where m and n are the positions corresponding to the weight matrix; the ReLU activation function is used to perform batch normalization and activation on the output of the large-scale convolutional layer. The output after batch normalization and activation is as follows: Finally, for the output after batch normalization and activation A max pooling operation is adopted, and its calculation method is as follows: In the formula, is the defined output of the pooling operation, and are the width and height of the pooling window.

7. A bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet according to claim 5, characterized in that, The multi-scale convolutional layer includes three types: a small-scale convolutional branch, a medium-scale convolutional branch, and a large-scale convolutional branch. Each of the three-scale branches consists of three convolutional blocks. Inside each convolutional block, there are sequentially a convolutional layer, a pooling layer, and an activation operation. The output of the large-scale convolutional layer After parallel computing and processing through the convolutional branches of the three scales, features of different scales are output , and .

8. A bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet according to claim 5, characterized in that, The feature fusion layer integrates feature information at different scales. The feature fusion layer splices the features at different scales output in three channels , and to obtain the fused vector .

9. A bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet according to claim 4, characterized in that, The ResNet channel is mainly divided into three layers, namely the initial convolution layer, the residual layer, and the global average pooling layer; The data after GASF preprocessing enters the large convolutional kernel of the initial convolutional layer, whose output channels are 64 for preliminary feature extraction. Then, batch normalization and activation are performed on the output of the initial convolutional layer. Finally, it is the max pooling layer to further reduce the size of the feature map. The residual layer contains 6 residual blocks, and each residual block contains two convolutional layers with a convolutional kernel size of 3×3. The ReLU activation function is selected, and the output channels are 64, 128, 256, 512, 512, and 512 in sequence. By increasing the number of channels layer by layer, deep features are extracted. The global average pooling layer flattens the deep feature data extracted by the residual layer into .

10. A bearing fault diagnosis system based on GASF-SSM-MSR-FusionNet, based on a bearing fault diagnosis method based on GASF-SSM-MSR-FusionNet according to any one of claims 1 to 9, characterized in that, It includes a data processing module, a feature extraction module, a feature fusion module, and an evaluation module; The data processing module is used to divide the original vibration signal into a data set and perform GASF data preprocessing and SSM data preprocessing on the data set respectively; The feature extraction module is used to construct an MSR-FusionNet feature extraction model. The MSR-FusionNet feature extraction model has dual data input channels. The data after GASF and SSM data preprocessing are respectively input into the dual data input channels of the MSR-FusionNet feature extraction model for feature extraction; The feature fusion module is used to perform feature fusion on the features extracted from the dual data input channels respectively; The fault diagnosis module is used to perform probability distribution on the fused features in terms of categories, and output the evaluation result according to the probability distribution to generate fault diagnosis data.

Images