Fault diagnosis method and system based on improved residual error and GRF feature fusion

Abstract

The invention discloses a fault diagnosis method and system based on improved residual error and GRF feature fusion. The method comprises the following steps: integrating signals of a multi-source sensor; reducing dimensions by using kernel principal component analysis; extracting features in parallel by adopting an improved residual network, an LSTM (Long Short Term Memory) and a Transform; after normalization and projection transformation are carried out on output of each branch, feature weighted fusion is carried out by adopting a dynamic gating weight mechanism, and timing sequence modeling capability is enhanced by utilizing a gating recursion unit; and finally, stage mark embedding and dynamic zooming processing are introduced, and diagnosis results are output through a full connection layer and Softmax classification. According to the method, the global and local information of the time-frequency domain of multiple sensors is fused, and the diagnosis precision of the complex fault of the rolling bearing is remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mechanical equipment signal processing, and particularly relates to a fault diagnosis method and system based on improved residual and GRF feature fusion. BACKGROUND

[0002] As the core transmission component of mechanical systems, the health status of rolling bearings directly determines the operation efficiency and service life of the equipment. According to statistics, about 30% of mechanical failures in industrial equipment are caused by bearing failure, and sudden failures not only cause unplanned downtime, but also may cause serious safety accidents. Therefore, state monitoring and intelligent fault diagnosis of rolling bearings have become an important research direction in the field of industrial operation and maintenance, and have significant engineering application value and economic benefits.

[0003] With the rapid development of artificial intelligence and machine learning technology, intelligent diagnosis methods based on deep learning have been widely studied and applied. This kind of method can automatically extract fault features from massive sensor data, reducing the dependence on expert prior knowledge and manual feature engineering. In the prior art, some methods use convolutional neural networks (CNN) to fuse spatial features of multiple vibration sensor signals; some other methods combine CNN and long short-term memory network (LSTM) to capture the spatial and temporal dependence of signals; some researches try to introduce the combination of Transformer structure and LSTM to enhance the modeling ability of long sequence signals, so as to realize the fusion diagnosis of multi-sensor information in the time domain or frequency domain.

[0004] However, the existing fusion diagnosis methods still have a significant defect: most of the methods only extract and fuse features of multi-sensor signals from a single domain (such as time domain or frequency domain), lack the ability to cooperatively represent global and local information in time-frequency domain, and have limited recognition ability for complex fault patterns, making it difficult to achieve comprehensive and robust fault diagnosis. SUMMARY

[0005] The purpose of the present application is to provide a fault diagnosis method based on improved residual and GRF feature fusion, which can fuse global and local information in time-frequency domain of multi-sensor and improve the accuracy of complex fault diagnosis of rolling bearings. On the other hand, a fault diagnosis system based on improved residual and GRF feature fusion is provided.

[0006] Technical solution: The fault diagnosis method provided by the present application comprises the following steps: Step S1, integrating and standardizing the original signals from multiple sensors to form a multi-channel data set; using kernel principal component analysis method to extract and reduce the dimension of nonlinear features of the multi-channel data set, and obtaining a low-dimensional fusion feature matrix; Step S2, input the low-dimensional fusion feature matrix into three parallel feature extraction branches respectively: the first branch adopts a residual network containing a fast Fourier convolution module for feature extraction; the second branch adopts a long short-term memory network for feature extraction; and the third branch adopts a Transformer encoder for feature extraction; Step S3, normalize and project the features output by the three branches; calculate the dynamic gating weight between the features of each branch, and fuse the features by weighting to generate a fused feature vector; Step S4, recursively process the fused feature vector using a gated recurrent unit; introduce stage label embedding and dynamically scale the feature vector; Step S5, input the processed feature vector into a fully connected layer, classify by a Softmax function, and output the fault diagnosis result.

[0007] The fault diagnosis method described in the application effectively improves the data consistency and comparability by integrating and standardizing the original signals of multiple sensors; significantly enhances the feature expression ability and reduces the computational complexity by using kernel principal component analysis for nonlinear feature extraction and dimensionality reduction; efficiently extracts spatial local features, time series dependent features and global context features by using three branches of residual network (containing fast Fourier convolution), long short-term memory network and Transformer encoder in parallel; the complementary advantages of multi-source features are fully retained based on the feature normalization and weighted fusion mechanism of dynamic gating weight; the representation ability of the feature sequence is further optimized by combining the recursive processing of the gated recurrent unit and the dynamic scaling of the stage label embedding; finally, the accuracy and robustness of fault diagnosis are comprehensively improved by full connection layer and Softmax classification output.

[0008] Preferably, the integration and standardization processing of the original signals from multiple sensors in step S1 includes: Align and integrate the original signals from multiple heterogeneous sensors by time point to construct a multi-channel original data matrix:

[0009] In the formula wherein each row corresponds to multi-source sensor data of a time point, n represents the number of sampling points, m represents the number of channels; Standardize the multi-channel original data matrix:

[0010] In the formula and are the mean and standard deviation of the first j channel, respectively.

[0011] By time point alignment and integration of raw signals from multiple heterogeneous sensors, a unified multi-channel data matrix is effectively constructed, eliminating the time delay differences and format barriers between multi-source data. Further, by using standardization processing based on the mean and standard deviation of each channel, the dimensional differences and distribution deviations between sensors are significantly eliminated, laying a high-quality data foundation for subsequent feature extraction and fusion.

[0012] Preferably, the step S1 of using the kernel principal component analysis method to perform nonlinear feature extraction and dimensionality reduction on the multi-channel data set comprises: For the standardized multi-channel data set, the original data space is mapped to a high-dimensional feature space through a nonlinear kernel function:

[0013] In the formula, σ is a kernel parameter, which controls the scale of nonlinear mapping; Calculate the kernel value of all sample pairs to obtain a symmetric kernel matrix, where the element is:

[0014] Center the kernel matrix:

[0015] In the formula, L N is an element of 1 / N × N matrix; N Solve the characteristic equation for the centered kernel matrix:

[0016] Obtain the eigenvalue and the eigenvector , and select the first k principal components according to the set cumulative contribution rate threshold, satisfying:

[0017] Map the original data to the first k principal component direction to obtain a low-dimensional fusion feature matrix:

[0018] Where is the m th component of the j th eigenvector, and the principal component of the fault is obtained .

[0019] ​The KPCA method is used to map the original data to a high-dimensional feature space by using a nonlinear kernel function, so as to effectively capture the complex nonlinear relationship in the multi-channel data; the principal components with the maximum variance contribution are extracted by constructing, centering and decomposing the kernel matrix, so that the representation capability of the features is significantly enhanced; the principal components are adaptively selected according to the cumulative contribution rate, so that the data dimensionality reduction is realized while the key fault information is retained, the redundant noise is eliminated, and a low-dimensional and high-discriminative fusion feature matrix is provided for the subsequent diagnosis process.

[0020] Preferably, the first branch in step S2 adopts a residual network comprising a fast Fourier convolution module to perform feature extraction, including: Based on the input feature map , the spatial domain convolution is converted into a frequency domain operation by using the fast Fourier convolution:

[0021] The intermediate feature is subjected to batch normalization processing and ReLU activation operation: ;

[0022] In the formula, is the mean value of the batch data, and are a scale factor and a translation factor, respectively, is a small positive number not equal to 0; The processed feature is subjected to fast Fourier convolution and batch normalization operation again:

[0023]

[0024] The channel dimension of the feature in the skip connection path is adjusted by 1x1 convolution, so as to match the feature dimension output by the main path:

[0025] The main path output feature and the skip connection path feature are added element by element, and the final output is obtained through an activation function: ;

[0026] .

[0027] By adopting the residual network containing a fast Fourier convolution module, the first branch converts the spatial domain convolution into a frequency domain operation, significantly improves the calculation efficiency and enhances the capture ability of the frequency domain features; combined with batch normalization processing and ReLU activation function, effectively speeds up the training convergence process and improves the gradient flow; using 1x1 convolution to adjust the channel dimension of the jump connection path, ensures the dimensional matching of feature fusion; finally through the element-wise addition of residual connection, while preserving the original information, highlights the detailed features, significantly improves the model's extraction ability of spatial local patterns and complex fault features.

[0028] Preferably, the normalization and projection transformation of the features output by the three branches in step S3 comprises: performing layer normalization processing on the original output features of the first branch, the second branch and the third branch respectively; transforming the features of each branch after layer normalization processing through a learnable linear projection matrix, and mapping to a unified dimensional feature space, as follows: ; ; ;

[0029] wherein F f , F l , F t respectively represent the original output features of the three branches; is a learnable projection matrix; LayerNorm is a layer normalization operation, and the calculation formula is as follows:

[0030] wherein and are the mean and standard deviation of the features, and are learnable scaling and offset parameters.

[0031] By performing layer normalization processing on the features output by the three branches respectively, the distribution of the features is effectively stabilized and the internal covariate shift is reduced, and the stability of the training process is enhanced; then, the features of each branch are mapped to a unified dimensional feature space by using a learnable linear projection matrix, which not only eliminates the scale and structural differences between different feature sources, but also creates a compatible and comparable representation basis for subsequent feature fusion, significantly improving the coordination and effectiveness of multi-source heterogeneous feature fusion.

[0032] Preferably, the calculation of the dynamic gating weight between the features of each branch in step S3 comprises: calculating the cosine similarity between each two branch output feature vectors i and jThe average cosine similarity at different time steps is:

[0033] wherein and ; T denotes the total number of time steps; t is an index ; and denotes two vectors at time t ; The average cosine similarity is input into a Softmax function to generate a dynamic fusion weight corresponding to each branch:

[0034] wherein the sum of all weights is 1.

[0035] The average cosine similarity at different time steps is calculated for each two branch output feature vectors, effectively quantifying the degree of dynamic correlation between different feature sources; then the average cosine similarity is input into a Softmax function to generate a dynamic fusion weight, realizing the normalization and adaptive adjustment of weight distribution; this mechanism can dynamically adjust the contribution proportion of each branch according to the real-time correlation between features, significantly enhancing the flexibility and discriminability of multi-source feature fusion, and ultimately improving the representation quality and robustness of the fused feature vector.

[0036] Preferably, the step S3 of weighting and fusing each feature comprises: At each time step, the three normalized branch features are linearly weighted according to the dynamically generated gating weight:

[0037] wherein t denotes the time step, C t denotes the fused feature vector; the weighted calculation result is taken as the fused feature vector at the time step.

[0038] At each time step, the three normalized branch features are linearly weighted according to the dynamically generated gating weight, realizing adaptive fusion of multi-source features; this mechanism can dynamically adjust the contribution proportion of each branch according to the real-time correlation between features, which not only fully retains the complementary advantages of spatial local features, temporal dependence features and global context features, but also significantly enhances the discriminability and information integrity of the fused feature vector, providing high-quality feature representation for subsequent temporal modeling and classification.

[0039] Preferably, the step S4 of using a gated recurrent unit to recursively process the fused feature vector comprises: The fused feature vector is processed using an LSTM unit, which calculates an input gate, a forget gate, an output gate, and a candidate state according to the input of the current time step and the state of the previous time step, and the gate signals are calculated as follows:

[0040] wherein i t 、 f t 、 o t、 g t represents the input gate, the forget gate, the output gate, and the candidate state in the gating mechanism; represents a sigmoid activation function; tanh represents a hyperbolic tangent activation function; represents the state vector of the previous time; b represents a bias vector; The state of the current time step is updated according to the gate signals, and an output state is calculated as follows: The state update formula is as follows:

[0041] The output is calculated as follows: .

[0042] The dynamic gating weight is used to linearly weight and fuse the features of each branch, realizing adaptive optimization allocation of the contribution degrees of multiple source features, and significantly enhancing the discriminability and information integrity of the fused feature vector; the recursive processing of the gated recurrent unit (GRU) is used to effectively capture the long-term and short-term dependencies in the feature sequence through the cooperative mechanism of the update gate and the reset gate, and to selectively retain important historical information, thereby further improving the representation ability and diagnostic robustness for the time-series dynamic fault patterns.

[0043] Preferably, the introducing phase label embedding and dynamically scaling the feature vector in step S4 include: A corresponding phase label embedding vector is generated according to the fault severity, and the phase label embedding is calculated as follows:

[0044] wherein is a learnable phase embedding vector; The phase label embedding vector is combined with the recursively processed features, and the combined features are dynamically scaled:

[0045] wherein αis a scalar hyper-parameter, is a stage weight matrix; after the above processing, the output of the GRFU at the time step t is and passes the state to the next time step.

[0046] By introducing a learnable stage label embedding vector, prior knowledge of fault severity is encoded into the feature representation, effectively enhancing the model's ability to distinguish fault evolution stages; by combining stage labels with recurrent features and performing dynamic scaling, adaptive adjustment of the feature representation is achieved, highlighting key information highly relevant to the current fault stage while suppressing irrelevant feature interference, significantly improving the model's representation accuracy and diagnostic robustness for complex fault states.

[0047] In a second aspect, the fault diagnosis system of the present application comprises: A data layer fusion module is used to integrate and standardize the input multi-sensor raw signals, and a kernel principal component analysis method is used for nonlinear feature extraction and dimensionality reduction, outputting a low-dimensional fusion feature matrix; A feature extraction module includes three parallel branches: The first branch is an improved residual network, which includes a fast Fourier convolution module for extracting joint features of the input features in the frequency domain and spatial domain; The second branch is a long short-term memory network for extracting time series dependent features of the input features; The third branch is a Transformer encoder for extracting long-range dependent features of the input features; A feature fusion module is used to normalize and project the output features of the three branches, calculate the dynamic gating weights between the features of each branch, and weight and fuse each feature to generate a fused feature vector; it also includes a gated recurrent unit for recursive processing and state updating of the fused features; A fault classification module is used to receive the features output by the feature fusion module, calculate through a fully connected layer and a Softmax function, and output a fault classification result.

[0048] Advantages: Compared with the prior art, the present application has the following significant advantages: 1. innovatively fuses multi-sensor global and local information in the time-frequency domain, significantly improving the diagnosis accuracy of rolling bearing complex faults; 2. automatically calculates dynamic weights according to feature importance, achieving adaptive optimal fusion of each branch feature; 3. uses a gated recurrent unit to capture long-range dependencies in fault signals, enhancing the model's time series state memory ability; 4. introduces a fault stage label, enabling the model to perceive fault severity and dynamically scale features for more precise diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 Schematic diagram of the method flow of the present invention; Figure 2 Schematic diagram of the FFC-ResNet model of the present invention; Figure 3 It is the GRF structure diagram of the present invention; Figure 4 Schematic diagram of the FRLT bearing fault diagnosis model framework of the present invention; Figure 5 This is a schematic diagram of the fusion model diagnosis process of the present invention; Figure 6 Schematic diagram of the accuracy and loss curves of the training set and validation set of the present invention; Figure 7 Schematic diagram of the test set confusion matrix of the present invention; Figure 8 This is the t-SNE visualization diagram of the present invention; Figure 9 Schematic diagram of confusion matrix of each comparison model of the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0051] The present invention provides a fault diagnosis method based on improved residual and GRF feature fusion, such as Figure 1 As shown, the following steps are included: Step S1: Data layer fusion based on kernel principal component analysis The basic idea of ​​kernel principal component analysis (KPCA) is to transform the original data space into a high-dimensional feature space using a nonlinear kernel function. Since the relationships between variables in the high-dimensional feature space are linear, linear principal component analysis can be used. Given the diverse information collected by multiple sensors and the complexity of faults, KPCA is used for data-level fusion to obtain kernel principal components.

[0052] First, the heterogeneous sensor signals are integrated into a unified multi-channel data set, preserving the temporal and spatial correlation of the original data and providing structured input for subsequent fusion. The multi-channel original data matrix is ​​represented as:

[0053] In the formula , where each row corresponds to multi-source sensor data at a time point, n Indicates the number of sampling points, m Indicates the number of channels.

[0054] Secondly, in order to ensure that the contribution of each channel in the kernel space is balanced due to the difference in numerical scale, the data is standardized to eliminate the dimensional difference. The formula is as follows:

[0055] In the formula and Respectively j The mean and standard deviation of each channel.

[0056] The complex relationship between multi-sensor signals is implicitly captured by the nonlinear kernel function, and the Gaussian kernel function is used to map the data into a high-dimensional feature space. The formula is as follows:

[0057] In the formula σ is the kernel parameter, which controls the scale of the nonlinear mapping. Calculate the kernel value of all sample pairs and obtain a symmetric kernel matrix, where the elements are:

[0058] Then, in order to eliminate the mean shift in high-dimensional space, make the principal component analysis conform to the zero mean assumption, and ensure the validity of the covariance matrix calculation, the kernel matrix is ​​centered:

[0059] In the formula L N All elements are 1 / N of N × N In order to extract the direction with the largest variance in high-dimensional space, retain the main characteristic patterns of multi-sensor data, and suppress noise and redundant information, the characteristic equation is solved:

[0060] Get the eigenvalue and eigenvectors According to the cumulative contribution rate threshold (such as θ =95%), before selecting k principal components, satisfying:

[0061] Finally, the multi-channel nonlinear correlation information is compressed into a low-dimensional representation to achieve data layer fusion, minimize information loss, and provide high-discrimination features for subsequent tasks. k The principal component directions are obtained to obtain the fused low-dimensional feature matrix:

[0062] in is the first component of the first feature vector, and the principal component of the fault can be obtained by using kernel principal component analysis: m j

[0063] Step S2, improving the residual network model The core of the Fast Fourier Convolution (FFC) module is to convert the traditional spatial domain convolution to the frequency domain, which is realized by using the Fast Fourier Transform (FFT). This conversion brings significant improvement in computational efficiency and model accuracy. In addition, the module also introduces the Local Fourier Unit (LFU), which is a learnable filter that performs in the frequency domain, which can further optimize the feature representation and enhance the performance of the model. FFC not only has a non-local receptive field, but also performs cross-scale information fusion within the convolution.

[0064] FFC consists of two interconnected paths: a spatial (local) path that performs ordinary convolution on part of the input feature channels, and a spectral (global) path that works in the spectral domain. Each path can capture complementary information with different receptive fields.

[0065] Therefore, in order to make up for the shortcomings of traditional residual local convolution and enhance the modeling of long-distance dependence, the conventional convolution in ResNet is replaced by Fast Fourier Convolution, which simultaneously processes local details and global context. The main path extracts features through joint convolution in the frequency domain and the spatial domain, and the 1x1 convolution path adjusts the dimension to realize the combination of residual learning and global features. Compared with traditional residual blocks, FFC-ResNet not only retains the advantages of residual (relief of gradient disappearance, support for deep training), but also enhances the capture of global information and improves the feature expression. The FFC-ResNet feature extraction model is shown in Figure 2 .

[0066] Input feature map , the FFC-ResNet model calculation process is as follows: first, the spatial domain convolution is converted to frequency domain operation by using Fast Fourier Convolution, which efficiently extracts global features, and the formula is as follows:

[0067] After FFC, batch normalization (BN) is introduced to normalize the intermediate features, accelerate the convergence speed, reduce the gradient disappearance, and improve the stability of the model. At the same time, the ReLU activation function is introduced to introduce nonlinearity, enhance the expression ability of the model, and avoid the degradation problem of linear layer stacking. The formula is as follows: ;

[0068] ​​In the formula, is the mean of the batch data, and are the scale factor and the translation factor, respectively, is a small positive number not equal to 0. To further extract the frequency domain features, secondary FFC is performed, and the formula is as follows:

[0069] Then, to ensure the stability of the feature distribution, batch normalization is performed again:

[0070] To ensure the feasibility of residual addition, the channel number is adjusted by 1x1 convolution to realize dimension matching, and the calculation of the skip connection path is as follows:

[0071] Finally, residual addition and activation are performed, and the residual learning mechanism is used to add elements, y is the final output of the FFC-ResNet model: ; .

[0072] Step S3, feature layer fusion based on gated recurrent fusion (GRF) (1) Input feature preprocessing The outputs of the three feature extraction branches (FFC-ResNet frequency domain features, LSTM time series features, and Transformer long-range features) are first processed by layer normalization to eliminate the distribution differences of different branch features. Then, through linear projection transformation, the features are mapped to a unified feature space to ensure dimension consistency and comparability, and the formula is as follows: ; ; ;

[0073] In the formula F f , F l、 F t represent the original output features of the three branches, respectively; is a learnable projection matrix; LayerNorm is a layer normalization operation, and the calculation formula is as follows:

[0074] where and are the mean and standard deviation of the features, and are learnable scaling and offset parameters.

[0075] (2) Dynamic gating weight generation The cross-modal mutual information between three features is calculated, and the complementarity strength between different feature modalities is quantified by cosine similarity. Specifically, the cosine similarity of two feature vectors at each time step is calculated, and the average value along the time dimension is taken. Finally, the dynamic fusion weight is generated through the Softmax function, ensuring that the weight sum is 1. This mechanism can automatically reduce the weight of highly correlated features, avoid information redundancy, and at the same time, improve the weight of complementary features.

[0076] Calculate the feature i The average cosine similarity of the feature j :

[0077] where and ; T denotes the total number of time steps; t is the index ; and denotes the two vectors at time t .

[0078] Generate dynamic fusion weight:

[0079] (3) Weighted feature fusion According to the dynamically generated gating weight, the three normalized features are linearly weighted and fused. This step is performed independently at each time step to generate a combined feature vector containing multi-modal information. This weighting method can adaptively adjust the contribution of each branch according to the strength of the complementarity between features. The calculation formula is as follows:

[0080] where t denotes the time step, C t denotes the fused feature vector.

[0081] (4) Recurrent state update The fused features are processed using LSTM units to maintain the time evolution state of the fault features. LSTM contains four core components: input gate, forget gate, output gate, and candidate state.

[0082] The gating signal is calculated as follows:

[0083] where i t , ft 、 o t、 g t denotes the input gate, the forget gate, the output gate and the candidate state in the gating mechanism; denotes the sigmoid activation function; tanh denotes the hyperbolic tangent activation function; denotes the state vector at the previous time step; b denotes the bias vector.

[0084] The state update formula is as follows:

[0085] The output calculation is as follows:

[0086] (5) Fault stage perception enhancement Based on the fault severity (mild / moderate / severe), a stage label embedding is added, and the feature representation is dynamically scaled. The fault stage is automatically divided according to the energy intensity of the frequency domain features: when the low-frequency energy dominates, it is a mild fault, and when the high-frequency impact is significant, it is a severe fault. This mechanism enables the model to adaptively adjust the feature representation intensity for different fault stages.

[0087] The stage label embedding calculation is as follows:

[0088] wherein is a learnable stage embedding vector.

[0089] Feature dynamic scaling:

[0090] wherein α is a scalar hyperparameter (may be used to control the scaling strength of the weight), is the stage weight matrix.

[0091] After the above processing, the output of the GRFU at time step t is , and the state is passed to the next time step. The structure diagram is shown in Figure 3 .

[0092] Based on similar inventive concepts, the embodiments of the present application also provide a fault diagnosis system corresponding to the fault diagnosis method, which includes three modules of data layer fusion, feature layer fusion and fault classification, and the model framework is shown in Figure 4 .

[0093] Firstly, the data layer fusion will collect the multi-sensor signals after data preprocessing and sample division, and construct a multi-channel data set corresponding to the fault type. Then, in the feature layer fusion stage, the multi-channel data set fused by the data layer is input into three feature extraction branches for feature extraction. Branch one adopts fast Fourier convolution residual network (FFC-ResNet) to realize global receptive field coverage through frequency domain convolution kernel, and extracts multi-scale frequency domain features combined with residual jump connection, which effectively captures the local impact component and global frequency shift characteristics in the signal; Branch two captures the global time sequence dependence of the signal through the LSTM network and extracts the time sequence features of the signal; Branch three captures the long-distance features of the time sequence signal through the Transformer encoder, and uses the multi-head self-attention mechanism to establish the feature correlation across time steps, which significantly improves the model's ability to capture long-range correlation of fault features. The three features are fused through the GRF feature layer fusion method, and finally sent to the fully connected layer (FC) through the Softmax function (SM) for classification, realizing the fault diagnosis of multi-sensor data.

[0094] As shown in Figure 5 , the fault diagnosis process corresponding to the fusion model includes: Step S1, data acquisition and preprocessing. The fault signal of the bearing equipment is collected through the data acquisition device, and appropriate preprocessing is performed; Step S2, data set construction. The preprocessed signal data is divided into training set, validation set and test set; Step S3, establish FRLT model. Based on the FFC-ResNet-LSTM-Transformer three-branch feature extraction method, a fault diagnosis model is established; Step S4, FRLT model training and verification. The training set and validation set obtained in step 2 are used to train and verify the model constructed in step 3; Step S5, FRLT model test. Through step 4 to save the optimal parameters of model training, the test set generated by the same preprocessing in step 2 is input into the model, and the corresponding results are output for analysis.

[0095] In order to verify the effect of FRLT model in bearing fault diagnosis, based on python3.9 and pytorch2.1 framework, experimental verification is carried out. The network model can output accuracy graph and loss value graph, confusion matrix graph and scatter plot. Through the above results, the fault diagnosis accuracy and classification effect are analyzed, and the performance of the proposed network model is evaluated.

[0096] Step S1, data set division Deep learning often needs a large amount of data for intelligent diagnosis. The bearing data set of Southeast University is used to retain the correlation of adjacent signals and expand the data sample set by using the method of segmented overlap interception. In order to ensure that each sample and the overlapping part contains at least one period, 1024 data points are selected as a sample, and the sample is intercepted according to the overlap rate of 50%. According to the above method, 10000 samples are generated, including 2000 samples of each fault type. Next, the training set, validation set and test set are divided according to the ratio of 7:2:1.

[0097] Step S2, network training and result analysis The network training sets the iteration number epoch to 50 times, sets the number of samples processed by each batch batch size to 64, sets the network learning rate to 0.0003, and uses the Adam optimization algorithm to update the network parameters. The performance of the deep learning network needs a series of indicators to judge, and the accuracy and loss are introduced as the standard of judgment. The accuracy is the ratio of the correctly classified samples to the total number of samples, and the loss is used to measure the difference between the predicted label and the true label in the training process.

[0098] The accuracy curve and loss curve of the model on the training set and the validation set are shown in Figure 6 . The maximum number of iterations of the model is set to 50 times, and after 10 iterations of training, the model network tends to be stable, and the final model training accuracy reaches 100%, and the accuracy in the validation set also reaches 100%, and the loss value also tends to 0, indicating that the model training is completed, and the accuracy is high. Apply the model to the test set, and the final result is shown in Figure 7 The confusion matrix. The recognition accuracy of different types of faults is 100%, and there is no overfitting phenomenon, which indicates that the model training effect is good.

[0099] Step S3, network test visual analysis In order to more intuitively show the changes of inter-class distance and intra-class distance after network classification, t-distributed stochastic neighbor embedding (t-SNE) is used. t-SNE is a data dimension reduction visualization technology that reduces high-dimensional data to two or three dimensions by preserving the similarity relationship between data.

[0100] By visualizing the test set original data through t-SNE, a scatter plot is drawn, as shown in Figure 8 (a), the features of the input layer are intertwined with each other, and there is no obvious boundary between different fault types. After feature extraction by the FRLT model, the output of the fully connected layer is reduced to two dimensions, and different types of samples are distinguished using different color points. Figure 8(b) in the figure shows that for the output of FRLT, data of the same category are clustered and have a clear distance from other categories. There are no overlapping sample points, and the diagnostic accuracy reaches 100%.

[0101] Step S4: Model comparison and analysis To further verify the effectiveness of the FRLT model, the proposed FRLT model is compared and analyzed with the classic deep learning models ResNet model, CNN-LSTM model, LSTM-Transformer model and ResNet-LSTM-Transformer (RLT) model using the same dataset.

[0102] In order to visually observe the fault classification effect of each model, the output confusion matrix of each model is drawn to compare and analyze the classification results. The confusion matrices of the four comparison models are as follows: Figure 9 As shown. Take the RLT model confusion matrix as an example, Figure 9 As shown in (d) in the figure, the other models are analyzed in the same way. This model incorrectly predicts 11 samples of the bearing rolling element fault C2 as outer race fault C3. After the RLT model is improved and fast Fourier convolution is introduced, the classification effect reaches 100%, as shown in Figure 7 As shown in Figure 2, the FRLT model has higher recognition accuracy than the other four classic models, which verifies the advantages of the FRLT model in multi-sensor signal bearing fault state recognition.

Claims

1. A fault diagnosis method based on improved residual and GRF feature fusion, characterized in that: The following steps are involved: Step S1: Integrate and standardize the original signals from multiple sensors to form a multi-channel data set; use the kernel principal component analysis method to perform nonlinear feature extraction and dimensionality reduction on the multi-channel data set to obtain a low-dimensional fusion feature matrix; Step S2: input the low-dimensional fusion feature matrix into three parallel feature extraction branches respectively: the first branch uses a residual network including a fast Fourier convolution module for feature extraction; the second branch uses a long short-term memory network for feature extraction; and the third branch uses a Transformer encoder for feature extraction; Step S3: normalize and project the features output by the three branches; calculate the dynamic gating weights between the features of each branch, and perform weighted fusion on each feature to generate a fused feature vector; Step S4: recursively process the fused feature vector using a gated recursive unit; introduce stage marker embedding and dynamically scale the feature vector; Step S5: Input the processed feature vector into the fully connected layer, perform classification through the Softmax function, and output the fault diagnosis result.

2. The fault diagnosis method according to claim 1, characterized in that: The integration and standardization of the raw signals from the multiple sensors in step S1 includes: Align and integrate the raw signals from multiple heterogeneous sensors by time point to construct a multi-channel raw data matrix: Where , where each row corresponds to multi-source sensor data at a time point, n Indicates the number of sampling points, m Indicates the number of channels; The multi-channel raw data matrix is ​​normalized: Where and Respectively j The mean and standard deviation of each channel.

3. The fault diagnosis method according to claim 1, characterized in that: The step S1 of performing nonlinear feature extraction and dimensionality reduction on the multi-channel data set using the kernel principal component analysis method includes: For the standardized multi-channel data set, the original data space is mapped to a high-dimensional feature space through a nonlinear kernel function: Where σ is the kernel parameter, which controls the scale of the nonlinear mapping; Calculate the kernel values ​​of all sample pairs and obtain a symmetric kernel matrix, where the elements are: ; Centralize the kernel matrix: Where L N All elements are 1 / N of N × N matrix; Solve the characteristic equation for the centered kernel matrix: ; Get the eigenvalue With the eigenvector , and select the first k principal components according to the set cumulative contribution rate threshold, satisfying: ; Map the original data to the front k principal component directions, and obtain the low-dimensional fusion feature matrix: ;in It is m The first eigenvector j components and obtain the principal components of the fault .

4. The fault diagnosis method according to claim 1, characterized in that: The first branch described in step S2 uses a residual network including a fast Fourier convolution module to perform feature extraction, including: Based on the input feature map , use fast Fourier convolution to convert spatial convolution into frequency domain operation: ; For intermediate features Perform batch normalization and ReLU activation operations: ; Where, is the mean of the batch data, and The scale factor and the translation factor are respectively, is a small positive number not equal to 0; Perform fast Fourier convolution and batch normalization operations again on the processed features: ; ; Adjust the channel dimension of the features in the jump connection path through 1×1 convolution to match the output feature dimension of the main path: ; Add the main path output features and the skip connection path features element by element, and obtain the final output through the activation function: ; .

5. The fault diagnosis method according to claim 1, characterized in that: The normalization and projection transformation of the features output by the three branches in step S3 includes: The original output features of the first branch, the second branch, and the third branch are respectively subjected to layer normalization processing; the features of each branch after layer normalization are transformed by a learnable linear projection matrix and mapped to a feature space of uniform dimension. The formula is as follows: ; ; Where F f 、 F l 、 F t Represent the original output features of the three branches respectively; is a learnable projection matrix; LayerNorm is a layer normalization operation, and the calculation formula is as follows: ;in and are the mean and standard deviation of the feature, and are learnable scale and offset parameters.

6. The fault diagnosis method according to claim 1, characterized in that: The calculation of the dynamic gating weights between the branch features in step S3 includes: Calculate the output eigenvector of each two branches i and j Average cosine similarity at different time steps: ;in and ; T represents the total number of time steps; t For index ; and Indicates at time t Two vectors of The average cosine similarity is input into the Softmax function to generate the dynamic fusion weight corresponding to each branch: ; where the sum of all weights is 1.

7. The fault diagnosis method according to claim 1, characterized in that: The weighted fusion of the features described in step S3 includes: At each time step, a linear weighted calculation is performed on the three normalized branch features according to the dynamically generated gating weights: ;in t represents the time step, C t Represents the fused feature vector; the weighted calculation result is used as the fused feature vector of this time step.

8. The fault diagnosis method according to claim 1, characterized in that: The recursive processing of the fused feature vector by using the gated recursive unit in step S4 includes: The fused feature vector is processed using an LSTM unit. The LSTM unit calculates the input gate, forget gate, output gate, and candidate state based on the input of the current time step and the state of the previous time step. The gate control signal is calculated as follows: ;in i t 、 f t 、 o t、 g t Represents the input gate, forget gate, output gate, and candidate state in the gating mechanism; represents the sigmoid activation function; tanh represents the hyperbolic tangent activation function; Represents the state vector at the previous moment; b represents the bias vector; Update the state of the current time step according to the gating signal and calculate the output state: The state update formula is as follows: ; The output is calculated as follows: .

9. The fault diagnosis method according to claim 1, characterized in that: The introduction phase of step S4, marking embedding and dynamically scaling the feature vector, includes: Generate the corresponding stage marker embedding vector according to the fault severity. The stage marker embedding is calculated as follows: ;in is the learnable stage embedding vector; The stage marker embedding vector is combined with the recursively processed features, and the combined features are dynamically scaled: ;in α is a scalar hyperparameter, is the stage weight matrix; after the above processing, GRFU is in time step t The output is , and pass the status to the next time step.

10. A fault diagnosis system based on improved residual and GRF feature fusion, characterized in that: include: The data layer fusion module is used to integrate and standardize the input multi-sensor raw signals, and uses the kernel principal component analysis method to perform nonlinear feature extraction and dimensionality reduction, and output a low-dimensional fusion feature matrix; The feature extraction module consists of three parallel branches: The first branch is an improved residual network, which includes a fast Fourier convolution module to extract the joint frequency and spatial domain features of the input features; The second branch is the long short-term memory network, which is used to extract the temporal dependency features of the input features; The third branch is the Transformer encoder, which is used to extract long-range dependency features of the input features; The feature fusion module is used to normalize and project the output features of the three branches, calculate the dynamic gating weights between the features of each branch, and perform weighted fusion of each feature to generate a fused feature vector. It also includes a gated recursive unit for recursive processing and state update of the fused features. The fault classification module is used to receive the features output by the feature fusion module, perform calculations through the fully connected layer and the Softmax function, and output the fault classification results.

Images