Rolling bearing cross-working condition compound fault diagnosis method based on mdmaan and mk-mmd

Abstract

The application discloses a rolling bearing cross-working condition conforming fault diagnosis method based on MDMAAN and MK-MMD, the method adopts a method of multiple source domain multi-head attention mechanism, utilizes the multi-head attention mechanism to combine a convolutional neural network to perform feature extraction on multiple input source domains and a target domain, and performs weight distribution on the extracted features. A multi-core maximum mean difference is selected as a measurement function of each pair of source domains and the target domain, and a suitable parameter configuration is selected, so that an MDMAAN model and a fault diagnosis process are constructed on the basis, and the compound fault diagnosis of the rolling bearing under the cross-working condition is completed. The rolling bearing public experimental data is used, and experimental verification shows that, compared with other deep transfer learning fault diagnosis models, the diagnosis accuracy of the model is the best, which verifies the effectiveness and superiority of the method in the rolling bearing cross-working condition compound fault diagnosis, and the proposed method can well solve the problem of the rolling bearing cross-working condition compound fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of rotating machinery fault diagnosis technology, and relates to a rolling bearing fault diagnosis method, specifically a rolling bearing cross-operating condition composite fault diagnosis method based on MDMAAN and MK-MMD. Background Technology

[0002] Rotating machinery is an indispensable part of the machinery manufacturing industry, and its normal operation highly depends on the health of its bearings. Rolling bearings are the joints of machinery, widely used in almost all types of rotating machinery, such as gearboxes, rolling mills, wind turbines, high-speed trains, and gas turbines. Compared with other mechanical components, rolling bearings are prone to failure due to prolonged exposure to harsh working conditions such as high temperature, high speed, and high load. If faults are not detected in time, equipment often shuts down, causing huge economic losses and even safety accidents. To reduce maintenance costs and avoid personal injury, it is necessary to diagnose faults promptly and accurately before incalculable losses occur.

[0003] Transfer learning plays a crucial role in fault diagnosis and health management. This is primarily because transfer learning effectively addresses issues such as data scarcity, imbalance, and poor generalization ability. Especially in practical industrial applications, where available equipment fault data is often limited, transfer learning-based fault diagnosis can transfer existing pre-trained models to new tasks, effectively improving model performance and accuracy.

[0004] The application of transfer learning in fault diagnosis within PHM (Prognostics and Maintenance Management) not only improves equipment operating efficiency and reduces failure rates but also saves enterprises significant maintenance costs, thereby enhancing their economic benefits. This paper analyzes the specific application of deep transfer learning algorithms in rolling bearing fault diagnosis, improving the accuracy and generalization of fault diagnosis. This not only provides a new technical solution for intelligent operation and maintenance of rotating machinery but also promotes the advancement of intelligent manufacturing technology and enhances the intelligence level of mechanical equipment.

[0005] Domain adaptation is key to transfer learning, primarily aiming to map data from the source and target domains to the same feature space, minimizing distances within that space. This approach helps improve model performance in new domains, avoiding the domain discrepancy problem often encountered in transfer learning. For unlabeled target domains, the main strategy guides feature learning by minimizing the loss function between the feature distributions of the source and target domains. The loss function is central to feature-based deep transfer learning, with different metrics yielding varying matching effects. Common loss functions include Maximum Mean Discrepancy (MMD), Kullback-Leibler bias, and Wasserstein bias, with MMD being the most widely used. Methods using MMD for domain adaptation include Domain Adaptive Neural Networks (DaNN), Deep Domain Confusion (DDC), and Domain Adaptive Networks (DAN). Zhang et al. used a sliding frame to collect samples and applied them to a transfer learning neural network to improve fault classification performance under different operating conditions. Jiang et al. applied multi-feature fusion and a joint adaptive transfer learning network to achieve bearing fault diagnosis. Han et al. used adversarial learning to guide the feature generator to provide transferable features, thereby enabling rolling bearing fault diagnosis.

[0006] Currently, most fault diagnosis research based on transfer learning focuses on single-source domain transfer. However, in real-world scenarios, labeled data is typically collected from multiple different source domains. These data not only differ in distribution from the target domain but also in distribution among themselves, resulting in multiple source domains. Therefore, neither merging all source domains nor selecting a single source domain can achieve ideal fault diagnosis results. Multi-source domain adaptive fault diagnosis methods have emerged to address this need. Yu et al. proposed a multi-source domain adaptive method with strong generalization and robustness, making it well-suited for fault diagnosis tasks under variable conditions. Zhuang et al. proposed a novel multi-source domain adaptive method, the Deep Negative Correlation Multi-Source Domain Adaptation Network, for mechanical fault diagnosis under different operating conditions. Huang et al. designed a joint loss for aligning feature and label information distributions to mitigate the negative impact of inconsistent label spaces. Yang et al. utilized SSAE and K-means to extract better domain-invariant and discriminative features from multi-source domain data, obtaining a fault diagnosis model superior to single-source domain transfer. Xu et al. proposed a multi-source unsupervised domain adaptive network, utilizing a feature extractor to learn feature representations and constructing a domain discriminator for each source domain. Through adversarial training, domain-invariant features are learned to diagnose the fault type of the target domain. Zhang et al. organically combined multi-source domain data, proposing a dual alignment mechanism and reducing prediction inconsistencies in multi-classifiers. Their excellent cross-domain diagnostic performance was verified in multiple experimental cases. To achieve multi-domain adaptation, Wang et al. proposed a domain-specific distribution alignment module, employing intra-domain and inter-domain alignment strategies to reduce the offset between the source and target domains. Wei et al. proposed a weighted domain adaptive network, assigning corresponding weights to different source domains by measuring the MMD loss between each source and target domain, and setting a threshold for domain adaptation to prevent negative transfer. The effectiveness of multi-source domain adaptation was verified in case studies across speed and load. A literature review reveals that multi-source domain adaptive fault diagnosis methods have significant advantages when processing data from multiple different source domains; however, the problem of how to reasonably allocate weights among multiple source domains to further improve diagnostic accuracy remains to be solved.

[0007] In summary, domain adaptation has been extensively tested and applied in fault diagnosis, proving effective in this field. However, in practical applications, fault diagnosis may involve various fault types, and single-source domain adaptation often overlooks this diversity, potentially reducing the accuracy of fault diagnosis. This paper proposes a multi-source domain multi-head attention adaptation network (MDMAAN) for fault diagnosis. By combining a multi-head attention mechanism, reasonable weights are assigned to multiple source domain inputs. Multi-kernel maximum mean discrepancy (MK-MMD) is used to measure the loss between each source and target domain, thus constructing the MDMAAN network. The effectiveness and generalization ability of this model in fault diagnosis are then explored. Summary of the Invention

[0008] The purpose of this invention is to construct a multi-source domain multi-head attention adaptive network model by optimizing the source domain classification loss, domain-invariant representation, and differences between classifiers, and to combine the multi-source domain multi-head attention adaptive network with multi-kernel maximum mean difference to provide a method for diagnosing composite faults of rolling bearings across operating conditions using MDMAAN and MK-MMD.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] A method for diagnosing combined faults in rolling bearings under multiple operating conditions using MDMAAN and MK-MMD includes the following steps:

[0011] Step 1: Divide the dataset into multiple source domain data and target domain data according to different working conditions, and divide it into training set and test set; extract common features from the input data and map the extracted original features to the common feature space; after processing by the attention mechanism, concatenate them together, and obtain the output of the multi-head attention mechanism after linear transformation; assign weights to the extracted features, extract specific features from the assigned features and map them to the feature space.

[0012] Step 2: Minimize the distribution difference between the source and target domains in the feature space using multi-kernel maximum mean difference (MK-MMD) and analyze the domain adaptation loss; analyze the classification loss of the source domain using source domain labels, analyze the classification loss of the prediction results for the target domain, analyze the total loss, and update the MDMAAN model through backpropagation.

[0013] Step 3: The MDMAAN model consists of two feature extraction units and two fault diagnosis units. The fault diagnosis unit consists of multiple Softmax classifiers. The total loss consists of source domain classification loss, MK-MMD loss and classifier classification loss. The dataset is classified according to different working conditions and input into the trained MDMAAN model. The diagnostic effect is evaluated through test set data and target domain labels to complete the MDMAAN cross-working-condition composite fault diagnosis.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] 1. The MDMAAN model has high diagnostic accuracy, especially for the diagnosis of complex faults in rolling bearings. It can use laboratory data to achieve cross-condition fault diagnosis in actual working scenarios. The MDMAAN model has good comprehensive diagnostic effect and has high identification accuracy and robustness.

[0016] 2. Only a very small number of fault categories were misclassified in the diagnostic results of the MDMAAN model. The clustering effect of the model is better than that of other models. The MDMAAN model has higher fault diagnosis accuracy in multi-source domains and cross-working conditions, thus proving the stability and superiority of the MDMAAN model.

[0017] 3. Experimental verification shows that, compared with CNN, CORAL and MFSAN models, MDMAAN can effectively identify the composite fault types of rolling bearings and improve the efficiency of fault diagnosis by using the multi-kernel maximum mean difference (MK-MMD) to measure the distance between features in a specific feature space between the source and target domains. Attached Figure Description

[0018] Figure 1 The fault diagnosis process for the MDMAAN model;

[0019] Figure 2 A composite fault diagnosis model for rolling bearings across operating conditions based on MDMAAN and MK-MMD;

[0020] Figure 3 Comparison of diagnostic accuracy of the MDMAAN model with different numbers of convolutional layers;

[0021] Figure 4 Comparison of loss values ​​for different numbers of convolutional layers in the MDMAAN model;

[0022] Figure 5 It is a major component of the experimental platform;

[0023] Figure 6 For bearing failure components;

[0024] Figure 7The results of the CNN model in Experiment A are shown in the following diagrams: a) CNN feature visualization diagram; b) CNN confusion matrix diagram.

[0025] Figure 8 The CORAL model results for Experiment A are shown in the following figures: a) CORAL feature visualization; b) CORAL confusion matrix.

[0026] Figure 9 Figures showing the MFSAN model results for Experiment A: a) MFSAN feature visualization; b) MFSAN confusion matrix.

[0027] Figure 10 Figures showing the MDMAAN model results for Experiment A: a) MDMAAN feature visualization; b) MDMAAN confusion matrix.

[0028] Figure 11 Figures showing the CNN model results for Experiment J: a) CNN feature visualization; b) CNN confusion matrix.

[0029] Figure 12 The results of the CORAL model for experiment J are shown in the following figures: a) CORAL feature visualization; b) CORAL confusion matrix.

[0030] Figure 13 Figures showing the MFSAN model results for Experiment J: a) MFSAN feature visualization; b) MFSAN confusion matrix.

[0031] Figure 14 Figures showing the MDMAAN model results for experiment J: a) MDMAAN feature visualization; b) MDMAAN confusion matrix. Detailed Implementation

[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0033] This invention provides a method for diagnosing composite faults in rolling bearings across operating conditions based on MDMAAN and MK-MMD. First, an 8-head attention mechanism is proposed, using eight identical attention functions to simultaneously process the input and acquire different feature information at the same location. This design allows each attention mechanism to optimize different feature parts at each location, thus balancing the potential biases that might arise from using the same attention mechanism. Integrating the 8-head attention mechanism into feature extraction allocates more weight to valuable information. Simultaneously, multi-kernel maximum mean difference is utilized to minimize the distribution difference between the source and target domains, better addressing cross-domain fault identification. Finally, a composite fault diagnosis model for rolling bearings across operating conditions based on MDMAAN and MK-MMD is proposed. To improve fault diagnosis accuracy, two feature extraction units and two fault diagnosis units are used, as shown in the flowchart below. Figure 1 As shown, the specific implementation steps are as follows:

[0034] Step 1: Construct a dual-feature extraction network structure model:

[0035] Step 1: Divide the dataset into n source domains Ds1, Ds2, ..., Ds according to different operating conditions. n The target domain data Dt is divided into training and test sets; features are extracted from the input data, and then eight attention functions are used to process the input data simultaneously, assigning different weights to various features so that the model can focus on fault features;

[0036] Steps 1 and 2: Extract features from the model input data using a common feature extractor, and map the extracted raw features to the common feature space;

[0037] Step 13: Use a multi-head attention mechanism to assign weights to the extracted raw features, divide each set of source domain data and target domain data into multiple sample pairs, input the sample pairs into a specific feature extractor for feature extraction and mapping to a specific feature space.

[0038] The calculation formula for the multi-head attention mechanism is as follows:

[0039] Multihead(Q,K,V)=Concat(head1,…,head k ,…,head h W o

[0040] In the formula: Concat(·) is a vertical concatenation operation of the matrix, where the weight matrix is... head i The expression is:

[0041]

[0042] In the formula: X is the input, and the weight matrix is... d is the scaling factor. k =d model / h.

[0043] Step 2: Construct a multi-core maximum mean difference algorithm:

[0044] Step 2: Multi-kernel maximum mean difference. Different kernels are used to enhance the maximum mean difference, thereby selecting the optimal kernel function and analyzing the domain adaptation loss.

[0045] The formula for calculating the maximum mean difference of the multi-core kernels is as follows:

[0046]

[0047] In the formula: φ(·) represents the feature map, P and Q represent two probability distributions, and the multi-kernel maximum mean difference is... Defined as the RKHS distance between P and Q if and only if When the distributions P and Q are equal.

[0048] The formula for calculating the mean difference is as follows:

[0049]

[0050] Where: β u Represent the weights and impose constraints on them to ensure the derived multi-core... It is a feature, where m represents m different weighted sums.

[0051] Step 22: Analyze the classification loss of the source domain through source domain label analysis, analyze the classification loss of the prediction result of the target domain, analyze the total loss, and update the MDMAAN model through backpropagation.

[0052] Step 3: Construct a composite fault diagnosis model for rolling bearings across operating conditions based on MDMAAN and MK-MMD:

[0053] Step 3: The MDMAAN model consists of two feature extraction units and two fault diagnosis units. The fault diagnosis unit consists of multiple Softmax classifiers. The total loss consists of source domain classification loss, MK-MMD loss and classifier classification loss.

[0054] Step 3.2, Combined Fault Diagnosis of Rolling Bearings under Multiple Operating Conditions Using MDMAAN and MK-MMD:

[0055] The dataset from Step 1 is classified according to different working conditions and input into the trained MDMAAN model. The diagnostic effect is evaluated by using validation set data and target domain labels to complete the MDMAAN cross-working-condition composite fault diagnosis.

[0056] Step 3: Evaluation of Cross-Operating Condition Complex Fault Diagnosis: When performing cross-operating condition complex fault diagnosis of rolling bearings, the diagnostic effect was evaluated using test set data and target domain labels. The results showed that the MDMAAN model performed well in diagnosing single inner ring faults, single outer ring faults, and combined inner and outer ring faults of rolling bearings, proving the effectiveness of the MDMAAN model in diagnosing cross-operating condition complex faults of rolling bearings.

[0057] Steps 3 and 4: Evaluation of the diagnostic performance of different models: MFSAN, CNN and CORAL models were selected as comparison models to obtain feature visualization maps and confusion matrix maps of the diagnostic model outputs. The comparison shows that the MDMAAN model has a better overall diagnostic performance and has higher recognition accuracy and robustness.

[0058] Example:

[0059] In this embodiment, the rolling bearing cross-operating condition composite fault diagnosis method based on MDMAAN and MK-MMD includes the following steps:

[0060] Step 1: Construct a dual-feature extraction network structure model;

[0061] Step 2: Construct a multi-core maximum mean difference algorithm;

[0062] Step 3: Construct a composite fault diagnosis model for rolling bearings under various operating conditions based on MDMAAN and MK-MMD.

[0063] The rolling bearing cross-condition composite fault diagnosis model based on MDMAAN and MK-MMD proposed in this embodiment is as follows: Figure 2 As shown.

[0064] The bearing failure PU dataset, publicly available from the University of Paderborn, was used. The Paderborn test bench consisted of a drive motor, torque measurement shaft, rolling bearing test module, flywheel, and load motor. The tested bearing was a 6203 bearing, and the failure data included actual damage caused by human error, EDM, EDM, drilling, and fatigue, with a sampling frequency of 64 kHz. The PU dataset contained data on 6 bearings in normal condition, 12 bearings damaged by human error, and 14 bearings damaged due to accelerated life testing. Due to the complexity of the PU dataset and the variety of bearing failure types, one normal bearing dataset and some accelerated life test bearing datasets (K005, KA15, KA16, KB23, KB24, KB27, KI14, and KI21) were selected for the experiments in this section. The failed bearings included inner ring failures, outer ring failures, and combined inner and outer ring failures, as shown in Table 1. The rotational speed and load of the rolling bearings during the PU data acquisition experiments are shown in Table 2.

[0065] Table 1. Description of Some Faulty Bearings in PU

[0066]

[0067] Table 2. Operating parameters of the PU data acquisition experiment

[0068]

[0069] As shown in Table 2, the PU dataset contains four different speeds, load torques, and radial forces. Based on these four different operating conditions, nine cross-condition multi-source domain migration fault diagnosis tasks can be set up in the experiment, as shown in Table 3. The data is divided into 5000 samples, each with a length of 1024.

[0070] Table 3 Experimental setup for multi-source domain fault diagnosis

[0071]

[0072]

[0073] 80% of each dataset was randomly selected as the training set and 20% as the test set. A set of bearing fault data was randomly selected as the test set. The number of iterations was set to 30, the initial learning rate was 0.01, and SGD was used as the optimization algorithm, with a momentum of 0.9. Each model was trained for 30 epochs for fault diagnosis. A learning rate decay strategy was also used to adjust the learning rate, and the weight decay value was initialized to 0.0005. The loss function was cross-entropy. In fault diagnosis, the overall accuracy was used as the evaluation metric, defined as the ratio of correctly classified samples to the total number of samples. The initial model parameter settings are shown in Table 4. The effects of different convolutional layers, different training batches, and different numbers of source domains on the fault diagnosis of the MDMAAN model were investigated.

[0074] Table 4 Experimental settings for MDMAAN model parameters

[0075]

[0076]

[0077] The MDMAAN model was compared with different numbers of convolutional layers using the experimental conditions set in Table 3. This paper compared models from two to seven convolutional layers, selecting the most suitable number of layers based on the model's diagnostic accuracy and loss curve. The diagnostic accuracy of the MDMAAN model is compared as follows: Figure 3 As shown, the loss value pairs are, for example... Figure 4 As shown.

[0078] By comparing different numbers of convolutional layers, it was found that as the number of layers increases, the diagnostic accuracy of the MDMAAN model increases the fastest and the highest, the loss value converges the fastest, and the iteration speed is the fastest when five convolutional layers are used. This paper selects five convolutional layers to conduct multi-source domain cross-condition fault diagnosis experiments.

[0079] Using our proprietary Bearing Failure Test Bench (BPS), the BPS test bench consists of a digital force display, a motor speed controller, a motor, a radial hydraulic loading system, an axial hydraulic loading system, a spindle, two support bearings (6210 and 18720), a 1-inch bearing under test, a lever beam adjustment device, and a hydraulic device for disassembling the bearing. Figure 5 As shown in the figure. Data was collected using a triaxial accelerometer with a sampling frequency of 25.6 kHz. Experiments were conducted on single and combined faults under different damage levels of the bearing. In the experiment, a radial accelerometer was used to acquire vibration acceleration information for one rolling bearing under normal conditions and 12 bearing fault conditions. The 12 bearing fault conditions included 9 single-point faults and 3 combined faults. The combined faults were a combination of inner ring, outer ring, and rolling element faults, as detailed below. Figure 6 As shown in Table 5, taking a rotational speed of 900 r / min as an example, the basic bearing parameters and fault characteristic frequencies are shown in Table 6, the experimental parameters for different fault states are shown in Table 7, and the experimental operating condition parameters are shown in Table 7.

[0080] Table 5. Main parameters and fault characteristic frequencies of ER-16K bearings

[0081]

[0082]

[0083] Table 6 Experimental parameter settings for different fault states

[0084]

[0085] Table 7 Experimental Operating Parameter Settings

[0086]

[0087] (1) Comparative experimental verification of rolling bearing fault diagnosis under different diagnostic models across working conditions

[0088] Using PU experimental data as an example, the method proposed in this chapter is validated. First, the diagnostic accuracy of the MDMAAN model across operating conditions demonstrates the superiority of this method, as shown in Table 5. Second, several deep transfer learning fault diagnosis models are compared. The feature visualization and confusion matrix of the label classification of the MDMAAN rolling bearing deep transfer learning fault diagnosis model further illustrate that the proposed model has strong fault diagnosis capabilities.

[0089] Table 5. Diagnostic results of the MDMAAN model for different operating conditions.

[0090]

[0091] Table 5 shows that the MDMAAN model achieved an average recognition accuracy of 98.87% in the composite diagnosis of rolling bearing faults across operating conditions, demonstrating its effectiveness. Comparisons under different operating conditions reveal that the MDMAAN model performs slightly better in diagnosing faults across loads than across speeds. This indicates that the influence of speed is greater than that of load when diagnosing rolling bearing faults across operating conditions.

[0092] A random experiment A was selected from Table 5. The source domain 1 operating conditions were: speed 1500 r / min, load torque 0.7 Nm, radial force 1000 N; the source domain 2 operating conditions were: speed 900 r / min, load torque 0.7 Nm, radial force 1000 N; and the target domain operating conditions were: speed 1500 r / min, load torque 0.1 Nm, radial force 1000 N. A comparison of deep transfer learning fault diagnosis was conducted using MFSAN, CNN, and CORAL models. The diagnostic results of different models are shown in Table 6. The feature visualization and confusion matrix output by the diagnostic models are shown in the figure below. Figure 7 , Figure 8 , Figure 9 and Figure 10 As shown.

[0093] Table 6. Accuracy of different diagnostic models in Experiment A

[0094]

[0095] A comparison of the accuracy of the MDMAAN, MFSAN, CNN, and CORAL models in Table 6 reveals that the MDMAAN model significantly outperforms the other three models in diagnosing rolling bearing faults under various operating conditions. The MDMAAN model's diagnostic accuracy is 4.75% higher than the MFSAN model, 12.87% higher than the CNN model, and 9.87% higher than the CORAL model. Figure 7 a) and Figure 7 b) It is not difficult to see that the CNN model performs poorly in diagnosing K005, KB23, and KB24, especially diagnosing 30% of KB24 fault categories as KB23 and 24% of KB24 fault categories as KI14; through Figure 8 a) and Figure 8 b) It can be seen that the CORAL model also misdiagnoses the KB24 fault category, diagnosing 37% of the fault categories as KB23. This is achieved through... Figure 9a) and Figure 9 b) It can be seen that the MFSAN model has a decent diagnostic effect for each fault category, with a small number of fault labels being misclassified into other fault categories, but its diagnostic effect is significantly worse than that of the MDMAAN model; through Figure 10 a) and Figure 10 (b) It can be seen that the MDMAAN model performs well in fault diagnosis for each category of rolling bearings. The results show that the target domain is clustered into eight clusters, corresponding to eight fault states in the target domain, including single inner ring fault, single outer ring fault, and combined inner and outer ring fault. Only a very small number of combined inner and outer ring faults are classified into other fault categories. In single fault diagnosis, the MDMAAN model is also superior to the other three models. The MDMAAN model has the highest diagnostic accuracy, especially for the diagnosis of combined faults in rolling bearings. It can use laboratory data to achieve cross-condition fault diagnosis in actual working scenarios, and the MDMAAN model has good overall diagnostic performance, with high identification accuracy and robustness.

[0096] (2) Experimental verification of cross-condition composite fault diagnosis of rolling bearings based on MDMAAN

[0097] Based on the 13 bearing categories in Table 8 and the 9 operating conditions in Table 7, the multi-source domain cross-operating condition fault diagnosis experimental tasks can be set up as follows: 10 types. To reduce the number of cross-operating condition experimental tasks with the same error, two different sets of data are used for fault diagnosis. Experiments A and B have the same experimental task; Experiments C and D have the same experimental task; Experiments E and F have the same experimental task; Experiments G and H have the same experimental task; and Experiments I and J have the same experimental task. The specific source domain and target domain operating conditions are shown in Table 8.

[0098] Table 7 Experimental Operating Parameter Settings

[0099]

[0100] Table 8. Experimental Tasks for Cross-Condition Fault Diagnosis Using the MDMAAN Model

[0101]

[0102] The fault diagnosis results of the cross-operating condition fault diagnosis experiment task of the MDMAAN model deep transfer learning in Table 8 are shown in Table 9.

[0103] Table 9 Results of the MDMAAN model cross-condition fault diagnosis experiment

[0104]

[0105] For Experiment J in Table 9, experiments were conducted comparing CNN, CORAL, MFASN, and MDMAAN models. The operating conditions for source domain 1 were 900 r / min rotation speed and 200 kg radial load; for source domain 2, the operating conditions were 1500 r / min rotation speed and 0 kg radial load; and for the target domain, the operating conditions were 2400 r / min rotation speed and 100 kg radial load. The diagnostic results of different models are shown in Table 10, and the feature visualization and confusion matrix output by the diagnostic models are shown below. Figure 11 , Figure 12 , Figure 13 and Figure 14 As shown.

[0106] Table 10 Accuracy of different diagnostic models in Experiment A

[0107]

[0108] A comparison of the accuracy of the MDMAAN, MFSAN, CNN, and CORAL models in Table 10 reveals that the MDMAAN model significantly outperforms the other three models in diagnosing rolling bearing faults under cross-condition working conditions. The diagnostic accuracy of the MDMAAN model is 7.23% higher than that of the MFSAN model, 12.49% higher than that of the CNN model, and 13.05% higher than that of the CORAL model, demonstrating the effectiveness and superiority of this model in multi-source domain cross-condition fault diagnosis. Figure 11 a) and Figure 11 b) It can be observed that 54% of the moderate faults in the B10 rolling element and 50% of the moderate faults in the O03 outer ring were misclassified in the diagnostic results of the CNN model. Through Figure 12 a) and Figure 12 b) It can be observed that 50% of the B15 rolling weight defects were misclassified in the CORAL model's diagnostic results, and a small number of other defects were also misclassified. Figure 13 a) and Figure 13 b) It can be seen that 49% of the B15 rolling weight faults were misclassified in the diagnostic results of the MFSAN model, and a small number of other faults were also misclassified. Figure 14 a) and Figure 14 b) It can be seen that only a very small number of fault categories were misclassified in the diagnostic results of the MDMAAN model. The results show that the target domain was clustered into thirteen clusters, corresponding to thirteen fault states in the target domain. Moreover, the clustering effect of the model is better than that of other models, indicating that the MDMAAN model has higher fault diagnosis accuracy in multi-source domains and cross-working conditions. Therefore, it proves the stability and superiority of the MDMAAN model.

Claims

1. A method for diagnosing composite faults in rolling bearings across operating conditions based on MDMAAN and MK-MMD, characterized in that... The method includes the following steps: Step 1: Divide the dataset into multiple source domain data and target domain data according to different working conditions, and divide it into training set and test set; Step 2: Construct the MDMAAN model. The model includes two feature extraction units and two fault diagnosis units. The feature extraction units include convolutional layers, pooling layers, activation layers, fully connected layers, and multi-head attention mechanisms. The two fault diagnosis units are composed of multiple classifiers. The total loss consists of source domain classification loss, MK-MMD loss, and classifier classification loss. Step 3: Input the training sets of source domain data Ds1, Ds2, ..., Dsn and target domain data Dt under multiple different working conditions into the MDMAAN model, extract common features from the input data, and map the extracted original features to the common feature space. After being processed by the attention mechanism, the data is concatenated together and then linearly transformed to obtain the output of the multi-head attention mechanism. The extracted features are weighted and then specific features are extracted and mapped to a specific feature space. Step 4: Minimize the distribution difference between the source domain and the target domain using the multi-core maximum mean difference (MK-MMD) and calculate the domain adaptation loss; calculate the classification loss of the source domain using the source domain label, calculate the classification loss of the prediction result for the target domain, calculate the total loss and update the MDMAAN model through backpropagation; evaluate the diagnostic effect using test set data and target domain label to complete the MDMAAN cross-condition composite fault diagnosis.

2. The method for diagnosing composite faults of rolling bearings across operating conditions based on the MDMAAN domain MK-MMD as described in claim 1, characterized in that... The calculation formula for the multi-head attention mechanism is as follows: In the formula: This is a vertical concatenation operation of matrices, with the weight matrix... , The expression is: In the formula: For input, weight matrix , , , , , , Scaling factor .

3. The method for diagnosing composite faults of rolling bearings under multiple operating conditions based on MDMAAN and MK-MMD as described in claim 1, characterized in that... Step four also includes the following: Step 41: Cross-condition composite fault diagnosis and evaluation: When performing cross-condition composite fault diagnosis of rolling bearings, the diagnostic effect is evaluated using test set data and target domain labels. Step 4.2 Evaluation of the diagnostic performance of different models: Select MFSAN, CNN and CORAL models as comparison models, and obtain the feature visualization map and confusion matrix map output by the diagnostic model.

4. The method for diagnosing composite faults of rolling bearings under multiple operating conditions based on MDMAAN and MK-MMD as described in claim 3, characterized in that... The formula for calculating the maximum mean difference of the multi-core kernels is as follows: In the formula: Represents feature mapping, and Represents two probability distributions, with the maximum mean difference across multiple kernels. Defined as and The RKHS distance between them is if and only if At time, distribution and equal; The formula for calculating the mean difference is as follows: In the formula: Represent the weights and impose constraints on them to ensure the derived multi-core... It is a feature, where m represents m different weighted sums.

Images