Elevator rolling bearing fault intelligent diagnosis method and related system thereof

Abstract

The invention relates to the field of mechanical fault diagnosis, in particular to an intelligent fault diagnosis method, system and equipment for a rolling bearing of an elevator and a storage medium. Three-axis vibration signals of a motor bearing are collected and subjected to normalization processing, frequency domain and time domain features are extracted through fast Fourier transform and variational mode decomposition, and multi-dimensional feature vectors are constructed. The method comprises the following steps: extracting spatial features through a convolutional neural network, splicing the spatial features with time features to form comprehensive feature representation, capturing a sequential relationship by using Transform, enhancing the features through a time subspace attention module, and finally outputting a bearing fault type through a full connection layer. According to the method, multiple advanced model structures are fused, the method adapts to non-stationary vibration signals under complex working conditions, and the diagnosis robustness and accuracy are remarkably improved. Meanwhile, a complete bearing fault intelligent diagnosis system is constructed, full-process automatic diagnosis from original data acquisition to fault type output is realized, and the system can be deployed in an industrial field and has good expandability and engineering practicability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mechanical fault diagnosis, in particular to a hoist rolling bearing fault intelligent diagnosis method, system, device and storage medium. BACKGROUND

[0002] As a key vertical transportation equipment in dense storage systems, the running efficiency and reliability of the hoist directly affect the overall operation capacity and safety of the storage system. The motor bearing, the core component of the hoist, is one of the most prone to failure parts. Its performance state is prone to deterioration, wear or failure under high-frequency start-stop, heavy-load operation and complex working conditions. If the fault is not identified in time, it may cause abnormal stop of the equipment, seriously affecting the continuity and stability of the storage operation. Traditional bearing fault diagnosis methods are mostly based on artificial experience judgment or use feature engineering and traditional machine learning models for classification and identification. Although certain effects have been achieved in some static environments, when facing the actual scenes of frequent start-stop, complex running state and non-stationary vibration signals in dense storage systems, there are problems such as strong feature extraction dependence, poor model generalization ability and low diagnosis accuracy, making it difficult to achieve rapid and accurate identification of early faults.

[0003] In recent years, with the development of deep learning technology, more and more researches have applied models such as convolutional neural networks to the field of fault diagnosis. This kind of method has certain advantages in feature extraction, but still has deficiencies in processing time series data and capturing global information, making it difficult to meet the high-precision requirements of fault identification under complex actual working conditions. SUMMARY

[0004] The present application provides a hoist rolling bearing fault diagnosis method and related system to solve the technical problems of poor adaptability to non-stationary signals, limited feature extraction capability and low diagnosis accuracy of traditional bearing fault diagnosis methods in the prior art.

[0005] The present application is realized by the following technical solutions: A hoist rolling bearing fault diagnosis method, comprising the following steps: Obtain the original vibration signal, use Fourier transform and variational modal decomposition method to process the original vibration signal, and construct a multi-dimensional feature vector; Input the multi-dimensional feature vector into a convolutional neural network, extract spatial features through multiple layers of convolution and pooling; Concatenate the multi-dimensional feature vector and the spatial features to form a comprehensive feature representation; Based on the comprehensive feature representation, use a Transformer model to capture the time sequence relationship and output spatio-temporal fusion features; The time subspace attention module is used for feature enhancement of the spatio-temporal fusion feature, the enhanced spatio-temporal fusion feature is input into a full connection layer for feature mapping and classification, and finally the corresponding bearing fault type is output, so that bearing fault diagnosis is realized.

[0006] The fast Fourier transform performs frequency domain analysis on the collected original vibration signal, converts the time domain signal to the frequency domain, and then extracts the main frequency component in the signal.

[0007] The variational mode decomposition method is used together with the fast Fourier transform method to perform time-frequency joint analysis on the vibration signal, and the vibration signal is decomposed into a plurality of intrinsic mode functions with specific bandwidth and center frequency. Each intrinsic mode function component retains the time domain evolution characteristics of the original signal, thereby enhancing the processing capability of non-stationary signals.

[0008] In the variational mode decomposition method, the Pearson correlation coefficient is used to measure the linear correlation between the mode components in the variational mode decomposition, and the specific method is as follows:

[0009] In the formula, the Pearson correlation coefficient between the modes and , and are the values of the modes and at the time , and are the mean values of the modes and , is the sample length.

[0010] The average correlation coefficient score method is used to evaluate the correlation between the mode components, and the variational mode decomposition method is optimized, and the specific method is as follows:

[0011] In the formula, K represents the number of mode decompositions, the average correlation coefficient score, if the score value is smaller, it indicates that the correlation between the mode components is lower, that is, the decomposition effect is better, and at this time the decomposition mode number K value is optimal.

[0012] The convolutional neural network includes three convolutional layers, the convolution kernel size of each convolutional layer is 3, the stride is 1, and the padding is 1. After each layer of convolution, a ReLU activation function is applied for nonlinear transformation, and a maximum pooling operation is further used to reduce the dimension of the feature map.

[0013] The temporal subspace attention module is used to enhance the spatiotemporal fusion features. Specifically, the input features of the temporal subspace attention module are divided into subspaces, each subspace has a length of G, and each subspace is processed independently. Assume that the input feature matrix is ,in is the batch size, and is the dimension of the feature map. Split by channel dimension to obtain multiple subspaces , each subspace first convolves each channel of the input feature through a depth-wise separable convolution operation, and then uses a point-wise convolution layer to linearly combine the results of the depth-wise convolution to obtain a single-channel feature representation.

[0014] A hoist rolling bearing fault diagnosis system, comprising a data processing module, a feature extraction module, a multi-dimensional feature splicing module, a spatiotemporal feature fusion module, and a fault diagnosis module; A data processing module is used to process the original vibration signal using Fourier transform and variational mode decomposition methods to construct a multi-dimensional feature vector; The feature extraction module inputs the multi-dimensional feature vector into the convolutional neural network and extracts spatial features through multi-layer convolution and pooling; Multi-dimensional feature concatenation module, which concatenates multi-dimensional feature vectors with spatial features to form a comprehensive feature representation; The spatiotemporal feature fusion module uses the Transformer model based on comprehensive feature representation to capture temporal relationships and output spatiotemporal fusion features; The fault diagnosis module enhances the spatiotemporal fusion features through the time subspace attention module, inputs the enhanced spatiotemporal fusion features into the fully connected layer for feature mapping and classification, and finally outputs the corresponding bearing fault type to realize bearing fault diagnosis.

[0015] A computer device comprises a memory, a processor and a computer program stored in the memory and running on the processor. When the processor executes the computer program, steps of a method for diagnosing rolling bearing faults of a hoist are implemented.

[0016] A computer storage medium is provided. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the computer program implements the steps of a method for diagnosing a rolling bearing fault of a hoist.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The application is a hoist rolling bearing fault intelligent diagnosis method, which utilizes fast Fourier transform method and variational mode decomposition method to pretreat the vibration signal, extracts frequency domain features and time domain features, constructs multi-dimensional feature input through feature splicing, extracts spatial features through convolutional neural network, introduces Transformer model to model time sequence features, and finally enhances the perception ability of the model to key space-time information through time subspace attention module, so as to realize accurate diagnosis of the running state of the motor bearing. The method fuses various advanced model structures, can adapt to the characteristics of non-stationary vibration signals under complex working conditions, and improves the robustness and accuracy of diagnosis.

[0018] Further, the time subspace attention module adopted by the application effectively enhances the modeling ability of key time sequence features through mechanisms such as subspace division, depth separable convolution, attention weighting and skip connection, and introduces Dropout operation to improve the generalization performance of the model. Compared with a single model using only CNN or Transformer, the application has stronger feature extraction and fault recognition capabilities.

[0019] In addition, the application also constructs a complete bearing fault intelligent diagnosis system, including a data acquisition module, a data processing module, a model training and testing module, a diagnosis model establishment and processing module, which can realize automatic diagnosis from original data acquisition to final fault type output, improve fault recognition efficiency and reduce the degree of manual participation.

[0020] Further, the terminal equipment and computer program matched with the system can be deployed in an industrial site to realize edge computing and real-time diagnosis, and have good scalability and engineering practicability. Compared with traditional diagnosis methods relying on manual experience or shallow learning models, the application method has significant improvement in accuracy, stability and intelligence, and provides an effective solution for improving the intelligent operation and maintenance of hoist rolling bearings. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The application is a hoist rolling bearing fault diagnosis method algorithm framework; Figure 2 Time subspace attention module framework; Figure 3 Hoist rolling bearing fault diagnosis system module diagram; Figure 4 The application is a hoist rolling bearing fault diagnosis method algorithm framework; Figure 5 The application is a hoist rolling bearing fault diagnosis method algorithm framework; Figure 6 : Test set t-SNE plot of the bearing fault diagnosis method verified using the public bearing dataset of Case Western Reserve University, USA; Figure 7 : Training set and validation set accuracy plot of the bearing fault diagnosis method verified using the actual working condition dataset of the elevator; Figure 8 : Test set confusion matrix plot of the bearing fault diagnosis method verified using the actual working condition dataset; Figure 9 : Test set t-SNE plot of the bearing fault diagnosis method verified using the actual working condition dataset; Figure 10 : Confusion matrix diagram of the test set in the embodiment of the application; Figure 11 : t-SNE visualization results of the test set before and after processing by the model in the embodiment of the application. DETAILED DESCRIPTION

[0022] The application will be further described in detail below with specific embodiments, which are an explanation of the application rather than a limitation.

[0023] Embodiment 1 As shown in Figure 1 , the application discloses a kind of intelligent diagnosis method of elevator rolling bearing fault, and the specific implementation method includes the following steps: Vibration signals of motor bearing running are collected by vibration sensor, including X, Y, Z three direction dimensions vibration signals.Original vibration signal is normalized to eliminate the influence of different dimensions or amplitude difference on model training. After randomizing the normalized data, it is divided into training set, validation set and test set according to the ratio of 7:2:1. Data preprocessing is carried out on the normalized data set by using fast Fourier transform and variational mode decomposition, the frequency domain features and time domain features of vibration signal are extracted by fast Fourier transform and variational mode decomposition, the obtained frequency domain features and time domain features are spliced, and multi-dimensional feature vector is constructed. Among them, the normalized original vibration signal is analyzed in frequency domain by using fast Fourier transform, the time domain signal is converted to frequency domain, and then the main frequency component in the signal is extracted. Fast Fourier transform can effectively identify the frequency characteristics of fault signal, which is a typical global feature extraction method, and is convenient for analyzing whether there is a certain frequency corresponding to the fault mode of bearing under certain working condition.

[0024] However, the Fourier transform method cannot reflect the local dynamic characteristics of the signal changing with time. Therefore, the present application further introduces a variational mode decomposition method to perform time-frequency joint analysis on the signal. The variational mode decomposition method can decompose the vibration signal into a plurality of intrinsic mode functions (IMF) with specific bandwidth and center frequency. Each intrinsic mode function (IMF) component retains the time-domain evolution characteristics of the original vibration signal, i.e., the time characteristics of the original vibration signal, thereby enhancing the processing capability of non-stationary signals.

[0025] When using the variational mode decomposition method, the selection of the number K of decomposition modes is crucial. If the value of K is not properly set, it may lead to information redundancy between modes, uneven feature distribution, and thus affect the subsequent classification accuracy. Therefore, the present application uses the Pearson correlation coefficient (PCC) to measure the linear correlation between all mode components to optimize the selection of the number K of modes. The specific calculation formula of the Pearson correlation coefficient is as follows:

[0026] In the formula, denotes the Pearson correlation coefficient between modes and , respectively, and are the values of modes and at time , respectively, and are the mean values of modes and , respectively, is the sample length.

[0027] To evaluate the correlation between modes, the present application also introduces an average correlation coefficient score, which is the average value of the absolute values of the Pearson correlation coefficients between all mode components. The calculation formula of the average correlation coefficient score is as follows:

[0028] In the formula, denotes the number of mode decomposition. denotes the average correlation coefficient score. The smaller the score value, the lower the correlation between the mode components, i.e., the better the decomposition effect, and the number K of decomposition modes is optimal at this time. The average correlation coefficient score is calculated for the normal bearing data decomposed by the variational mode decomposition method, and the score values under different mode numbers are as follows: Figure 4As shown. It can be seen that when K=4, the average correlation coefficient score is the lowest, so the number of decomposed modes is 4, and the intrinsic mode function IMF after decomposition is as follows Figure 5 shown.

[0029] A convolutional neural network model was constructed, and the concatenated multi-dimensional feature vectors were input into the convolutional neural network. The spatial features, mostly feature vectors, were extracted through multiple layers of convolution and pooling operations within the convolutional neural network. The convolutional neural network model consisted of three convolutional layers, each with a kernel size of 3, a stride of 1, and a padding of 1. After each convolution layer, the ReLU activation function was applied to the feature vectors for nonlinear transformation, and the feature map was further reduced in dimensionality through a max pooling operation.

[0030] The spatial features output by the convolutional neural network model and the temporal features of the vibration signal after data preprocessing using the variational mode decomposition method are spliced ​​in the channel dimension to form a comprehensive feature representation. Based on the comprehensive feature representation, the Transformer model is used to capture the temporal relationship and output the spatiotemporal fusion features. The temporal subspace attention module is used to enhance the spatiotemporal fusion features. The enhanced spatiotemporal fusion features are input into the fully connected layer for feature mapping and classification, thereby improving the expressiveness of spatiotemporal features. The specific implementation is as follows: like Figure 2 As shown, the spatiotemporal fusion features are input into the time subspace attention module and divided into multiple subspaces. The time subspace attention module first divides the input features into subspaces, each subspace has a length of G, and each subspace is processed independently. Assume that the input feature matrix is ,in is the batch size, and is the dimension of the feature map. Split by channel dimension to obtain multiple subspaces , the number of channels in each subspace is , the size of each subspace is Each subspace first convolves each channel of the input feature through a depthwise separable convolution operation (convolution kernel size is 1, stride is 1), and then uses a pointwise convolution layer (convolution kernel size is 1) to linearly combine the results of the depthwise convolution to obtain a single-channel feature representation. The specific steps are as follows: Step 1: Divide the input features into multiple subspaces, and process each subspace independently; Step 2: Apply depth-wise separable convolution to each subspace to extract local features; Step 3: Batch normalization is performed on the convolved features, and a SiLU activation function is used for nonlinear transformation. Step 4: Dimensionality reduction is performed using the max pooling operation. Step 5: Point convolution is applied to compress the channel number to 1, further reducing the dimension. Step 6: Batch normalization is performed, and a SiLU activation function is used for nonlinear transformation. Step 7: The Softmax function is applied to calculate the attention weight of each position, and the input features are weighted. Step 8: The enhanced features are added to the original input features through a skip connection, preserving the input information and avoiding information loss. Step 9: Dropout is applied to the output features to prevent overfitting and enhance the model's generalization ability. Step 10: All subspace-processed features are concatenated to form the final spatio-temporal fusion feature after feature enhancement.

[0031] The time subspace attention module is used to enhance the spatio-temporal fusion feature, which is then input into the fully connected layer for feature mapping and classification, finally outputting the corresponding bearing fault type, achieving bearing fault diagnosis.

[0032] Embodiment 2 As shown in Figure 3 , this embodiment proposes a hoist rolling bearing fault diagnosis system, including a data processing module, a feature extraction module, a multi-dimensional feature concatenation module, a spatio-temporal feature fusion module, and a fault diagnosis module; to realize the process of the hoist rolling bearing fault diagnosis method proposed in the embodiment, the implementation method is as follows: The data processing module is used to process the original vibration signal using the Fourier transform and variational mode decomposition method, and construct a multi-dimensional feature vector. The feature extraction module inputs the multi-dimensional feature vector into the convolutional neural network, extracts spatial features through multiple layers of convolution and pooling. The multi-dimensional feature concatenation module concatenates the multi-dimensional feature vector and the spatial feature to form a comprehensive feature representation. The spatio-temporal feature fusion module uses the Transformer model based on the comprehensive feature representation to capture the temporal relationship and output the spatio-temporal fusion feature. The fault diagnosis module uses the time subspace attention module to enhance the spatio-temporal fusion feature, which is then input into the fully connected layer for feature mapping and classification, finally outputting the corresponding bearing fault type, achieving bearing fault diagnosis.

[0033] Example 3 To verify the effectiveness of the proposed algorithm, the algorithm was verified using the public bearing data set of Case Western Reserve University in the United States. The bearing vibration data of the driving end was selected for the experiment, the sampling frequency was 12 kHz, and the load was 0 HP. The data set contains normal bearing operation data, and inner ring fault, outer ring fault and rolling element fault data under three fault diameters of 0.007 inches, 0.014 inches and 0.021 inches, a total of 10 categories of data. The original data was sampled by sliding window, the time step was set to 1024, the overlap rate was 0.5, and the data samples were divided according to the ratio of 7:2:1. The specific experimental data information is shown in Table 1.

[0034] Table 1

[0035] The model was built using PyTorch, the optimizer used Adam to adaptively adjust the learning rate, the loss function was cross-entropy loss function, the batch size was set to 32, the initial learning rate was 0.001, and the network training times was 100. The related parameters of the CNN-Transformer-TSAM model are shown in Table 2.

[0036] Table 2

[0037] To avoid accidents, 10 repeated experiments were conducted. The experimental results show that the highest accuracy of the model reaches 99.93%, and the lowest accuracy is 99.54%. In the experiment with the lowest accuracy, the accuracy curve of the model is as shown in Figure 6 It can be seen from the figure that the accuracy of the model improves rapidly after a few rounds of training at the beginning, the accuracy exceeds 90% at the 15th round, and tends to be stable around the 27th round, and there is no significant fluctuation. This shows that the model has strong learning ability and good convergence, and can achieve high-precision classification effect in a limited number of training rounds.

[0038] The model was evaluated using the test set, and the confusion matrix of the model on the test set is shown in Figure 7 The horizontal axis represents the predicted label, and the vertical axis represents the true label. From the confusion matrix, it can be seen that all samples of all categories are accurately classified except type 3. The classification accuracy of type 3 is 94%, and the average classification accuracy of the model in the ten classification task is 99.48%. The results show that the model performs well in the ten classification task.

[0039] The t-SNE method was used to project the high-dimensional feature space of the test set samples to a two-dimensional visualization plane for analysis, and the results are shown in Figure 8It can be seen that after t-SNE dimension reduction, the sample points of the ten categories are clearly distributed in the two-dimensional space, only a few categories are misclassified, and the overall effect is good, fully indicating the accuracy of the model in the classification task.

[0040] (2) Actual hoist working condition data set example In order to reflect the superiority and adaptability of the algorithm of the present application to actual working condition problems, the actual working condition data set of the hoist is used for verification. The bearing type is 6204 deep groove ball bearing, the sampling frequency is 12kHz, the data contains four kinds of fault types, which are normal, inner ring fault, outer ring fault and rolling element fault, and the vibration signals of X, Y and Z axes are collected for each fault type. The original data is sampled by sliding window, the time step is set to 1024, the overlap rate is 0.5, and the data samples are divided into training set, validation set and test set according to the ratio of 7:2:1. The actual experimental data information is shown in Table 3.

[0041] Table 3

[0042] The parameter settings are the same as above, after training the data set under actual working condition, the accuracy curve of the model proposed in this paper in the training process is as shown in Figure 9 It can be seen from the figure that the accuracy of the training set and the validation set shows a relatively rapid upward trend in the initial stage, and after 25 rounds of training, the accuracy of the training set and the validation set is more than 90%, and after 30 rounds of training, the accuracy tends to be stable. Through training, the final validation set accuracy reaches 98.92%, and the training effect is good.

[0043] Save the model with the best performance in the training process, and use the test set to evaluate it, the confusion matrix of the test set is as shown in Figure 10 It can be seen that the model shows high classification accuracy on most categories, and the classification accuracy of 7 categories reaches 100%. But there are still some misjudgments. For example, the misjudgment rate of type 2, type 5, type 11 and type 12 is 4%. In summary, the average classification accuracy of the model in this experiment is 98.61%, indicating that the model has high classification accuracy under actual working conditions and shows good practical application potential.

[0044] The output features of the model on the test set are reduced and visualized by using t-SNE method. Figure 11The t-SNE visualization results of the test set before and after the model processing are shown. It can be found that most of the class samples form obvious clusters in the two-dimensional space, and there is almost no overlap between different classes, and only individual misjudgments exist in a few classes. This result further shows that the proposed model can still effectively distinguish different types of fault signals under actual working conditions, verifying the effectiveness of the model in practical applications.

[0045] Example 4 To further evaluate the overall performance of the model, the accuracy, precision, recall and F1 value of the test set are used to compare and analyze the four models, and the results are shown in Table 4.

[0046] Table 4

[0047] It can be seen that the four indicators of the CNN-Transformer-TSAM model are the best, followed by the CNN-TSAM model, the CNN-Transformer model ranks third, and the four indicators of the Transformer-TSAM model are the lowest. In terms of test set accuracy, the CNN-Transformer-TSAM model improves by 0.69% compared with the CNN-TSAM model, the CNN-TSAM model improves by 0.7% compared with the CNN-Transformer model, and the CNN-Transformer model improves by 1.22% compared with the Transformer-TSAM model. Further shows that the proposed CNN-Transformer-TSAM model has high precision and excellent performance in actual working conditions.

[0048] Example 5 To verify the adaptability of the proposed model under different operating load conditions, variable load working condition experiments are carried out. As shown in Table 5, the experiment collects four kinds of load working condition data: 0kg, 10kg, 20kg and 30kg, which are marked as data sets A, B, C and D respectively. In the table, training set A and test set B indicate that the model is trained on data set A and tested on data set B, and the others are the same. The final test result takes the average value of the test set accuracy of 10 experiments. After the experiment, the lowest accuracy of the model reaches 90.16%, and the highest accuracy reaches 98.61%.

[0049] Table 5

[0050] Example 6 The application further discloses an electronic device, which comprises a memory, a processor and a computer program stored in the memory and running on the processor, and the processor implements the flow steps of the hoist rolling bearing fault diagnosis method proposed in Embodiment 1 when executing the computer program. The specific implementation method is as follows: In actual work, the electronic device collects vibration signals in X, Y and Z directions from the vibration sensor at the position of the hoist motor bearing in the field. These signals are stored in the memory, i.e. the memory of the electronic device. Then, the processor calls the data processing module to perform normalization processing and other preprocessing operations on the original vibration signals according to the computer program instructions stored in the memory, and performs fast Fourier transform and variational modal decomposition operations, extracts frequency domain and time domain features and constructs a multi-dimensional feature vector. Next, the spliced features are input into the built CNN-Transformer-TSAM model for training and testing. During the model training process, the processor continuously adjusts the model parameters so that the model can accurately learn the bearing fault features and output the corresponding fault type. When the server is used as a diagnosis device, the processor executes the diagnosis program in the memory, processes and classifies the newly collected bearing vibration signals using the trained model, and finally outputs whether the bearing has a fault and the specific fault type, thereby realizing intelligent fault diagnosis of the hoist rolling bearing.

[0051] Embodiment 7 The application further discloses a computer readable storage medium, which stores a computer program, and the computer program realizes the flow steps of the hoist rolling bearing fault diagnosis method proposed in Embodiment 1 when executed by a processor. The specific implementation method is as follows: When the intelligent bearing fault diagnosis method is deployed and used on a computer device or a portable fault diagnosis device in an industrial field, the USB flash drive is inserted into the USB interface of the computer device. After the operating system of the computer device recognizes and mounts the USB drive, the user can install and run the diagnosis program stored in the USB drive on the device. During the installation process, the computer device will copy the diagnosis program and related files to the storage of the device itself (such as a hard disk or a solid state disk). Thereafter, the processor of the computer device can execute the computer program stored in the local storage to process the collected motor bearing vibration signals according to the method steps, and realize the intelligent diagnosis function of bearing faults. This way makes the intelligent bearing fault diagnosis method can be easily spread and deployed between different computer devices, improves the scalability and practicality of the method, and meets the diversified needs of bearing fault diagnosis in industrial fields.

[0052] The above merely describes the preferred embodiments of the present application, and is not intended to limit the technical solutions of the present application in any way. Those skilled in the art should understand that, without departing from the spirit and principle of the present application, the technical solutions can also be modified and replaced in several simple ways, and these modifications and replacements also all belong to the protection scope covered by the claims.

Claims

1. A method for diagnosing rolling bearing faults of a hoist, characterized in that: The following steps are involved: Obtain the original vibration signal, use Fourier transform and variational mode decomposition method to process the original vibration signal, and construct a multi-dimensional feature vector; The multi-dimensional feature vector is input into the convolutional neural network, and spatial features are extracted through multi-layer convolution and pooling; Concatenate the multi-dimensional feature vectors with the spatial features to form a comprehensive feature representation; Based on comprehensive feature representation, the Transformer model is used to capture temporal relationships and output spatiotemporal fusion features; The spatiotemporal fusion features are enhanced through the time subspace attention module, and the enhanced spatiotemporal fusion features are input into the fully connected layer for feature mapping and classification, and finally the corresponding bearing fault type is output to realize bearing fault diagnosis.

2. A method for diagnosing rolling bearing faults of an elevator according to claim 1, characterized in that: The fast Fourier transform performs frequency domain analysis on the collected original vibration signal, converts the time domain signal into the frequency domain, and then extracts the main frequency components in the signal.

3. A method for diagnosing rolling bearing faults of an elevator according to claim 1, characterized in that: The variational mode decomposition method is used together with the fast Fourier transform method to perform time-frequency joint analysis on the vibration signal, decomposing the vibration signal into several intrinsic mode functions with specific bandwidth and center frequency. Each intrinsic mode function component retains the time domain evolution characteristics of the original signal, thereby enhancing the processing capability of non-stationary signals.

4. A method for diagnosing rolling bearing faults of an elevator according to claim 3, characterized in that: In the variational mode decomposition method, the Pearson correlation coefficient is used to measure the linear correlation between the modal components in the variational mode decomposition. The specific method is as follows: Where, Indicates modality and The Pearson correlation coefficient between and Mode and exist The value of the moment, and is modal and The mean of is the sample length.

5. A method for diagnosing rolling bearing faults of an elevator according to claim 3, characterized in that: The average correlation coefficient score method is used to evaluate the correlation between modal components and optimize the variational modal decomposition method. The specific method is as follows: Where, represents the number of modal decompositions, It represents the average correlation coefficient score. If the score value is smaller, it indicates that the correlation between the modal components is lower, that is, the decomposition effect is better, and the decomposition mode number K value is optimal at this time.

6. A method for diagnosing rolling bearing faults of an elevator according to claim 1, characterized in that: The convolutional neural network contains three convolutional layers, each of which has a convolution kernel size of 3, a stride of 1, and a padding of 1. After each convolution layer, a ReLU activation function is applied for nonlinear transformation, and the dimension of the feature map is further reduced through the maximum pooling operation.

7. A method for diagnosing rolling bearing faults of an elevator according to claim 1, characterized in that: The temporal subspace attention module is used to enhance the spatiotemporal fusion features. Specifically, the input features of the temporal subspace attention module are divided into subspaces, each subspace has a length of G, and each subspace is processed independently; assuming that the input feature matrix is ,in is the batch size, and is the dimension of the feature map; by inputting features Split by channel dimension to obtain multiple subspaces , each subspace first convolves each channel of the input feature through a depth-wise separable convolution operation, and then uses a point-wise convolution layer to linearly combine the results of the depth-wise convolution to obtain a single-channel feature representation.

8. A hoist rolling bearing fault diagnosis system, based on a hoist rolling bearing fault diagnosis method according to any one of claims 1 to 7, characterized in that: It includes data processing module, feature extraction module, multi-dimensional feature splicing module, spatiotemporal feature fusion module and fault diagnosis module; A data processing module is used to process the original vibration signal using Fourier transform and variational mode decomposition methods to construct a multi-dimensional feature vector; The feature extraction module inputs the multi-dimensional feature vector into the convolutional neural network and extracts spatial features through multi-layer convolution and pooling; Multi-dimensional feature concatenation module, which concatenates multi-dimensional feature vectors with spatial features to form a comprehensive feature representation; The spatiotemporal feature fusion module uses the Transformer model based on comprehensive feature representation to capture temporal relationships and output spatiotemporal fusion features; The fault diagnosis module enhances the spatiotemporal fusion features through the time subspace attention module, inputs the enhanced spatiotemporal fusion features into the fully connected layer for feature mapping and classification, and finally outputs the corresponding bearing fault type to realize bearing fault diagnosis.

9. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the steps of a method for diagnosing a rolling bearing fault of a hoist according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the hoist rolling bearing fault diagnosis method according to any one of claims 1 to 7 are implemented.

Images