A method, device, medium and equipment for generalizing fault diagnosis of rolling bearings

Abstract

The specification discloses a rolling bearing domain generalization fault diagnosis method, relates to the technical field of domain generalization fault diagnosis, and comprises the following steps: obtaining a time domain signal of a rolling bearing vibration signal, performing time-frequency conversion on the time domain signal to obtain a frequency domain signal; a fault diagnosis model is constructed; the fault diagnosis model comprises a mixed feature network and a fault classifier; the mixed feature network comprises an IBN network and a Mixstyle network; the time domain signal and the frequency domain signal are jointly input into the mixed feature network, discriminative features in the frequency domain signal are extracted through the IBN network; generalization features in the time domain signal are extracted through the Mixstyle network; the discriminative features extracted through the IBN network and the generalization features extracted through the MixStyle network are spliced to form a high-dimensional feature vector; and the spliced feature signal is input into the fault classifier for fault diagnosis. The method can effectively extract rolling bearing discriminative features and generalization features, and improve the diagnosis accuracy.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis technology, and in particular to a generalized fault diagnosis method for rolling bearings. Background Technology

[0002] In the fault diagnosis of rolling bearings, traditional deep learning models rely on training and validation data following similar distributions. However, in real-world industrial scenarios, the rotational speed and load conditions of rolling bearings dynamically change, leading to significant differences in data distribution and limiting the model's generalization ability. Researchers are exploring more flexible and efficient domain generalization (DG) methods. Currently, researchers mainly achieve domain generalization through data processing, network architecture optimization, and design optimization strategies to cope with dynamic and complex operating conditions and ensure reliable diagnostic capabilities on unknown data. Therefore, this invention needs to fully utilize vibration data and further investigate how to effectively extract generalization features while also considering discriminative features in the network architecture.

[0003] For existing DG diagnostic methods, the domain generalization method cannot fully integrate the time-domain and frequency-domain information of the vibration signal in the processing of the obtained vibration signal, resulting in inaccurate fault diagnosis. Summary of the Invention

[0004] This specification provides an embodiment of a rolling bearing domain generalization fault diagnosis method, which can effectively extract the distinguishing features and generalization features of rolling bearings, thereby improving the accuracy of diagnosis. The method includes the following steps:

[0005] The time-domain signal of the rolling bearing vibration signal is obtained, and the time-domain signal is converted to frequency to obtain the frequency-domain signal; A fault diagnosis model is constructed; the fault diagnosis model includes a hybrid feature network and a fault classifier; the hybrid feature network includes an IBN network and a Mixstyle network. The time-domain signal and the frequency-domain signal are jointly input into the hybrid feature network. The discriminative features in the frequency-domain signal are extracted by the IBN network, and the generalization features in the time-domain signal are extracted by the MixStyle network. The discriminative features extracted by the IBN network and the generalization features extracted by the MixStyle network are concatenated to form a high-dimensional feature vector. The spliced ​​feature signals are input into the fault classifier for fault diagnosis.

[0006] Preferably, the method also includes optimizing the feature extractor parameters, comprising the following steps: Confirm according to different fault types individual working conditions Each working condition Experience risk is ; Construct the IRM expression; adjust the weights of the invariant constraints in the IRM expression using cross-entropy loss, with the initial weights as follows: ; The adjusted weights are: ; In the formula, Update rate; By giving more invariant constraints to the worse-performing operating conditions, we can obtain improved invariant constraints. The improved IRM formula is as follows: ; In the formula, This represents the balance parameter, used to adjust the weight between empirical risk and invariant constraints. Indicates when When, the loss function is related to gradient, The second norm of the gradient is represented. This represents the weight of the invariance constraint. Representative feature extractor, Represents a fault classifier. For all source domain operating conditions; In practical small-batch optimization, unbiased estimates are used instead of invariant constraints, as shown in the formula: ; In the formula, and It comes from the working condition Two random mini-batch samples, size ; It is a loss function. For improved invariance constraints; The empirical risk is set as the cross-entropy loss of the classifier, as shown in the formula: ; The formula for the joint optimization objective is: ; The optimization strategy is as follows: ； ; In the formula, For learning rate, These are the current feature extractor parameters. For the updated feature extractor parameters, These are the current classifier parameters. These are the updated classifier parameters.

[0007] Preferably, the time-frequency conversion of the time-domain signal includes the following steps: The phase frequency characteristic is obtained through Fourier transform, and the expression for the phase value is: ; ; In the formula, They are The real and imaginary parts are given, and the Fourier transform is calculated using the Fast Fourier Transform algorithm.

[0008] Preferably, the construction of the hybrid feature network includes the following steps: Combining IN and BN networks, IN networks are added to the shallow layers of the CNN network while BN networks are retained in the deeper layers. Instance normalization and batch normalization are then applied together within the CNN network; for a given input feature... The processes of instance normalization and batch normalization are as follows: ; ; in, and These are the mean and standard deviation calculated during instance normalization, respectively. and This is a batch-normalized statistic; Through weighted combination, the final output feature is: ; in, It is in The hyperparameters of the interval are used to adjust the balance between instance normalization and batch normalization; The MixStyle network is applied to the shallow layer of the CNN network. In the shallow layer, MixStyle generates sample features under new working conditions by mixing the working condition features of existing samples. For a given input sample and , computational feature representation and And extract the mean and standard deviation from them, that is: ； ; By combining these statistical information, new mean and standard deviation are generated: ; in, It is in arrive The hyperparameters between these parameters are used to adjust the mixing ratio of the samples; Style transfer of features is performed using a mixture of mean and standard deviation to obtain new feature representations. : ； In the hybrid feature extraction network, the discriminative features in the phase frequency signal are extracted through the IBN network, and the generalization features in the time domain signal are extracted through the Mixstyle network. The two are then concatenated and input into the fault classifier.

[0009] The present invention also provides a rolling bearing domain generalized fault diagnosis system, the system comprising: processor; A memory on which computer programs that can run on the processor are stored; The computer program, when executed by the processor, implements the steps of the rolling bearing domain generalized fault diagnosis method.

[0010] The present invention also provides a computer-readable storage medium storing a data processing program, which, when executed by a processor, implements the steps of the rolling bearing domain generalized fault diagnosis method.

[0011] The above-described at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: The rolling bearing fault diagnosis method proposed in this invention makes full use of the characteristics of vibration signals to provide multi-dimensional support for the diagnosis task from the signal level; a novel hybrid feature network is designed, using an IBN network to extract discriminative features from phase frequency information to improve the accuracy of the diagnosis task, and using a MixStyle network to extract generalization features from time domain signals to improve the generalization ability of the model.

[0012] Secondly, considering the differences in characteristics among different source domains, this paper proposes an improved IRM optimization method. By introducing classification cross-entropy loss, differentiated weights are assigned to the IRM of different source domains, focusing on the poorly performing source domains, thus ensuring that even the weakest performing source domain can achieve optimized performance. This method significantly improves the diagnostic generalization ability of the model in multi-source heterogeneous environments, providing technical support for practical applications under complex conditions. Attached Figure Description

[0013] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0014] Figure 1 This invention provides a schematic flowchart of a rolling bearing domain generalized fault diagnosis method. Figure 2 This invention provides a model for a rolling bearing domain generalized fault diagnosis method. Figure 3 The present invention provides a signal processing module that extracts amplitude-frequency and phase-frequency characteristics; wherein, Figure 3 (a) is the amplitude spectrum; Figure 3 (b) is the phase spectrum; Figure 4 A hybrid feature extractor provided by the present invention; Figure 5 The confusion matrix for a portion of the generalization experiments in the experimental setup provided by this invention; wherein, Figure 5 (a), (b), (c), (d), (e), and (f) are the confusion matrices of the generalization experiment; Figure 6 The T-SNE visualization results provided in the experimental setup of this invention; wherein, Figure 6 (a), (b), (c), and (d) are the T-SNE visualization results; Figure 7 Comparative experimental results on the bearing vibration test bench data provided in the experimental setup of this invention; Figure 8 The confusion matrix and T-SNE visualization results of the bearing vibration test bench data in the experimental setup provided by this invention; wherein, Figure 8 (a), (b), (c), and (d) are confusion matrices. Figure 8 (e), (f), (g), and (h) are the T-SNE visualization results. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0016] In the fault diagnosis of rolling bearings, the collection of fault signals is crucial. Vibration signals must be comprehensively, rationally, and fully utilized to provide multi-dimensional feature support for the diagnostic task at the signal level. However, commonly used feature network extractors often cannot adequately identify feature signals, leading to poor performance of deep models when faced with complex diagnostic tasks.

[0017] Existing DG diagnostic methods mainly achieve domain generalization through data processing, network architecture optimization, and design optimization strategies to cope with dynamic and complex operating conditions. However, they only enhance the diversity of data without solving the problem of distribution offset at the feature level. Furthermore, the optimization strategies proposed in network optimization can lead to a decrease in the performance of the diagnostic model.

[0018] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart illustrating a rolling bearing domain generalized fault diagnosis method according to the present invention. Figure 2 The model diagram includes the following steps: S101: The phase-frequency characteristics obtained from the Fourier transform contain rich high-level semantics and have low sensitivity to changes in operating conditions. For example... Figure 3 As shown in the visualization of the original signal, it can be seen that the phase frequency response contains far more information than the amplitude frequency response. Therefore, the phase frequency response contains rich invariant information, and invariant features can be extracted from it. The phase frequency response can be obtained using the following formula:

[0020] ; ； in It is an index. This is the length of the signal. The Fourier transform can be efficiently calculated using the Fast Fourier Transform algorithm. The expression for the phase value is as shown in the above equation, where... and They are The real and imaginary parts.

[0021] Time-domain signals directly reflect the dynamic changes and time characteristics of a system, which is crucial for fault diagnosis. Compared to frequency-domain features, time-domain signals retain more original data details and better describe the changes in rolling bearings under different operating conditions. Therefore, time-domain signals contain richer generalization information, and generalization features can be extracted from them.

[0022] S102: As Figure 4As shown, the hybrid feature extractor consists of two modules: an IBN network and a MixStyle network. The IBN network extracts invariant features from the phase frequency signal, while the MixStyle network extracts generalized features from the time domain signal, thereby achieving the fusion of discriminative and generalized features.

[0023] IBN: While shallow layers of CNN networks extract features primarily related to operational conditions, deeper layers extract features containing higher-level semantic information. IBN filters out environment-related information while retaining discriminative features, improving the accuracy of fault detection. It eliminates the impact of operational condition changes in shallow networks. The IBN network structure is as follows: Figure 4 As shown.

[0024] The core idea of ​​the IBN network is to perform double normalization on the feature map to achieve flexible adaptation to features from different domains. Specifically, the IBN network applies instance normalization and batch normalization after each convolutional layer, enabling the deep learning network to better capture discriminative features in the task, thereby improving generalization ability. The parameters of the IBN network are shown in Table 1. This process can be described by the following formula:

[0025] For a given input feature map The processes of instance normalization and batch normalization are as follows: ； ; in, and These are the mean and standard deviation calculated during instance normalization, respectively. and This is a batch-normalized statistic.

[0026] Through weighted combination, the final output feature is represented as: ; in, It is in The hyperparameters of the interval are used to adjust the balance between instance normalization and batch normalization.

[0027] Table 1 Detailed parameters of the IBN network Cov (1，4，8，1，1) IBN Cov (4，16，8，1，1) IBN Cov (16，32，8，1，1) BN Cov (32，64，8，1，1) BN Cov (64，128，8，1，1) BN AdaptiveMaxPool 4 / The MixStyle shallow neural network extracts features containing more operational condition changes. Therefore, the MixStyle network is applied to the shallow layers of a CNN network. The structure of the MixStyle network is designed as follows: Figure 4As shown in Table 2, the shallow MixStyle network generates new sample features by mixing the existing sample features under different working conditions. This is beneficial for deep models to extract richer generalization features. The detailed parameters of the MixStyle network are shown in Table 2. The method is described through the following steps:

[0028] First, for a given input sample and MixStyle computes their feature representations. and And extract the mean and standard deviation from them, that is: ; ; Then, by combining these statistical information, new mean and standard deviation are generated: ； ; in, It is in arrive The hyperparameters between these parameters are used to adjust the mixing ratio of the samples. Then, style transfer is performed on the features using the mean and standard deviation of the mixture to obtain a new feature representation. :

[0029] ; After feature extraction of the phase frequency signal through the IBN network, a 1*2560 dimension vector is generated. After feature extraction of the time domain signal through the MixStyle network, a 1*2560 dimension vector is also generated. This invention concatenates the two to form a 1*5120 dimension vector, which is then input into the fault classifier for fault diagnosis.

[0030] Table 2 Detailed parameters of the MixStyle network Cov (1，4，8，1，1) BN MixStyle / / Cov (4，16，8，1，1) BN MixStyle / / Cov (16，32，8，1，1) BN Cov (32，64，8，1，1) BN Cov (64，128，8，1，1) BN AdaptiveMaxPool 4 / S103: In IRM, assuming the data comes from multiple operating conditions, and each operating condition has a different distribution, IRM requires the model to maintain the same optimal predictive performance across all operating conditions by optimizing a constrained problem. The optimization objective is as follows: ； in: Indicates working condition The empirical risk in the formula is subject to the following constraints: ; Since the above equation is a two-level optimization problem, it can be simplified into a single-variable optimization problem. ; in, To become the entire invariant predictor, It is a scalar and fixed "virtual" classifier, and the gradient norm penalty is used to measure the optimality of the virtual classifier in each environment. It is a regularization parameter used to balance predictive power (the empirical risk minimization term) and the predictor. The invariance of.

[0031] However, the above formula has a problem: in rolling bearing fault diagnosis, due to the different data distribution deviations under various working conditions, the formula's approach of treating all working condition data equally often fails to achieve optimal results. Therefore, this invention uses cross-entropy loss to adjust the weights of invariant constraints, giving more invariant constraints to the worse-performing working conditions, thereby ensuring that the deep model performs well under all working conditions and improving its generalization ability. The specific improvements are as follows. Assume there are n working conditions, each working condition... The empirical risk in the current mini-batch is In the initial state, the domain weights Set as:

[0032] ; For each working condition sampled in the current mini-batch Its weight update formula is: ; in: Update rate, used to control the magnitude of updates; After obtaining the updated weights, multiply the weights by the invariance constraints to obtain the improved IRM: ; in: Indicates working conditions Experience risk It is a balancing parameter used to adjust the weight between empirical risk and invariant constraints. Indicates when When, the loss function is related to gradient, The second norm of the gradient is represented. This represents the weight of the invariance constraint. Representative feature extractor, Represents a fault classifier. This applies to all source domain conditions. However, in practical mini-batch optimization (such as stochastic gradient descent), it is difficult to directly calculate invariant constraints. Therefore, this invention uses The unbiased estimate is replaced by the following formula:

[0033] ; in, and It comes from the working condition Two random mini-batch samples, size ,and It is a loss function. For improved invariance constraints.

[0034] This invention will incorporate empirical risk. Cross-entropy loss set as the classifier: ; in, and These represent the marginal probability distribution and the conditional probability distribution, respectively.

[0035] The joint optimization objective of the proposed IBNMixNet is expressed by the following equation: ; in, For improved invariance constraints, To adjust and The hyperparameters relating these relationships. The optimization strategy is as follows:

[0036] ; ; in, For learning rate, These are the current feature extractor parameters. For the updated feature extractor parameters, These are the current classifier parameters. These are the updated classifier parameters.

[0037] The rolling bearing fault diagnosis method based on this invention fully utilizes the characteristics of vibration signals to provide multi-dimensional support for the diagnostic task at the signal level. A novel hybrid feature extraction network is designed, using an IBN network to extract discriminative features from phase frequency information to improve the accuracy of the diagnostic task, and a MixStyle network to extract generalization features from time-domain signals to enhance the model's generalization ability. Addressing the differences in characteristics among different source domains, an improved IRM optimization method is proposed: by introducing a classification cross-entropy loss, differentiated weights are assigned to the IRM of different source domains, focusing on the poorly performing source domains, thus ensuring that even the weakest performing source domain can achieve optimized performance. This method significantly improves the model's diagnostic generalization ability in multi-source heterogeneous environments, providing technical support for practical applications under complex working conditions.

[0038] The above describes a rolling bearing fault diagnosis method provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides an optimization approach for feature extractor parameters, including the following steps: Based on different fault types, multiple operating conditions are identified; Constructing IRM expressions: ； The invariant constraints in the IRM expression are weighted using cross-entropy loss, with the initial weights as follows: ; The adjusted weights are: ； in: Update rate, used to control the magnitude of updates; By applying more invariant constraints to the worse-performing working conditions, we obtain improved invariant constraints; and based on the improved invariant constraints, we obtain an unbiased estimate of the improved indeformable constraints. The improved IRM formula is as follows: ； in: Indicates working conditions Experience risk ， It is a balancing parameter used to adjust the weight between empirical risk and invariant constraints. Indicates when When, the loss function is related to gradient, The second norm of the gradient is represented. This represents the weight of the invariance constraint. Representative feature extractor, Represents a fault classifier. This applies to all source domain operating conditions.

[0039] In practical small-batch optimization, unbiased estimates are used instead of invariant constraints, as shown in the formula: ; in, and It comes from the working condition Two random mini-batch samples, size and It is a loss function. For improved invariance constraints; The empirical risk is set as the cross-entropy loss of the classifier, as shown in the formula: ; The joint optimization objective is: ; The specific optimization strategy is as follows: ; ; in, For learning rate, These are the current feature extractor parameters. For the updated feature extractor parameters, These are the current classifier parameters. These are the updated classifier parameters.

[0040] This embodiment provides specific experimental verification, including: 1: Introduction to the dataset A: CWRU Dataset: The CWRU dataset is one of the commonly used bearing fault datasets. The basic information of the CWRU dataset is shown in Table 3. This dataset includes acceleration signals collected from the drive end and fan end. At each collection point, four different operating conditions are defined based on load and speed. Each operating condition includes one normal data point and three fault data points (inner ring, outer ring, and rolling element), with damage diameters of 0.007, 0.014, 0.021, and 0.028 inches, respectively. Generalization experiments for rolling bearing fault diagnosis were conducted using eight domains and seven fault categories.

[0041] Table 3. Details of the CWRU dataset A 1797 0HP driver end 0 1 2 3 4 5 6 B 1772 1HP driver end 0 1 2 3 4 5 6 C 1750 2HP driver end 0 1 2 3 4 5 6 D 1730 3HP driver end 0 1 2 3 4 5 6 E 1797 0HP Fan end 0 1 2 3 4 5 6 F 1772 1HP Fan end 0 1 2 3 4 5 6 G 1750 2HP Fan end 0 1 2 3 4 5 6 H 1730 3HP Fan end 0 1 2 3 4 5 6 B: HIT Aero-engine Bearing Fault Dataset: The HIT dataset is a typical representative of aero-engine bearing fault datasets. The basic information of the HIT dataset is shown in Table 5. The HIT dataset is divided into 28 operating conditions based on the different high and low pressure speeds of the aero-engine. Each operating condition includes one normal condition, one inner race fault, and two outer race faults. Detailed fault conditions are shown in Table 4. To increase the experimental difficulty and test the generalization ability of the IBNMixNet model, four operating conditions with significant differences in high and low pressure speeds were selected for generalization experiments.

[0042] C: Bearing Vibration Test Bench Data: The third set of experimental data comes from the bearing vibration test bench at the experimental center, model TYS1-8. Five fault types and one normal type were designed in the experiment. The fault types are: normal, inner ring fault, outer ring fault, ball fault, outer and inner ball fault, and outer and inner ring fault. Nine working conditions were designed according to different rotational speeds and axial loads, as shown in Table 6. To increase the experimental difficulty, the designed axial load varied significantly, resulting in a more severe deviation in the vibration signal distribution, thus better verifying the superior performance of the proposed model.

[0043] Table 4 Fault Types in HIT Aircraft Engine Bearing Dataset 1 Outer ring 0.5 0.5 2 Inner circle 0.5 0.5 3 Inner circle 0.5 1.0 Table 5 shows the 28 operating conditions included in the HIT aero-engine bearing dataset. Table 6 lists the nine operating conditions included in the bearing vibration test bench data. 0 1000 0 1 1000 500 2 1000 1000 3 1500 0 4 1500 500 5 1500 1000 6 2000 0 7 2000 500 8 2000 1000 Table 7 Ablation Experimental Design MixStyle module and cross-entropy loss MixStyle module and IRM loss MixStyle and Improved IRM Loss MixStyle module and IBN module IBN module and cross-entropy loss IBNMixNet Table 8. Range of key parameters in the comparative experiment 2: Experimental Setup To ensure the fairness of the experimental results, the same main structure was used in all experiments. The experiments were conducted on a computer equipped with an NVIDIA GeForce RTX 4060Ti GPU, a 13th Gen Intel® Core™ i7 CPU, and 32GB of RAM. To verify the effectiveness of the IBNMixNet method, extensive experiments were designed, consisting of ablation experiments and control experiments. The ablation experiment design is shown in Table 7, and the parameter ranges for the control experiments are shown in Table 8. The control methods are described below:

[0044] 1): ERM: The core idea is to minimize the empirical risk of all source domains, thereby achieving domain generalization.

[0045] 2): DCORAL: Aligns the covariance between the source domain data of each group, thereby reducing distribution differences and achieving domain generalization.

[0046] 3): DCC: Aligns MMD metrics between different groups of source domain data, aligns features of different source domains, and achieves domain generalization.

[0047] 4): CNN-C: Introducing center loss into CNN improves the generalization ability of deep diagnostic models.

[0048] 5): DRO: This method combines DRO with regularization to improve the generalization ability of the overparameterized model on the worst population while maintaining average accuracy.

[0049] 6): VREx: Improves generalization ability by reducing risk differences during source domain training.

[0050] 7): MixStyle: Generates new style samples by mixing features, improving generalization ability.

[0051] 3: Experimental results on the CWRU dataset A: Ablation Experiment Results: The experimental results of various ablation models on the CWRU dataset are shown in Table 9. The confusion matrix of some generalization experiments is shown below. Figure 5 As shown in the table, the M6 ​​method achieved the highest average accuracy and also demonstrated optimal performance in the vast majority of generalization tasks. Comparing M1 and M2, it's clear that simply combining IRM with MixStyle does not improve model performance because treating all source domains equally fails to guarantee improved generalization ability of the deep diagnostic model. Combining with M3, it's evident that the improved IRM, by fully considering the worse-performing source domains, exhibits superior performance, achieving the highest accuracy in both generalization tasks, but its overall performance is still inferior to M6. Comparing M4 and M1, M5 shows that using a hybrid feature extractor yields better results than a single feature extractor, demonstrating the importance of considering both generalization features and discriminative feature extraction. Comparing M6 and M4 reveals that optimizing the deep diagnostic model with the improved IRM yields better results.

[0052] Table 9 Ablation experimental results on the CWRU dataset M1 88.28 86.72 100 97.66 83.59 90.62 95.31 88.28 91.31 M2 89.06 86.72 100 92.97 88.28 87.5 94.35 91.41 91.29 M3 93.75 80.47 100 98.44 85.16 94.53 99.22 91.41 92.87 M4 92.91 89.06 100 92.97 85.94 89.06 99.22 89.84 92.38 M5 77.34 85.94 100 94.53 81.25 81.25 87.5 77.34 85.64 M6 95.24 90.62 100 100 90.48 94.05 100 91.67 95.26 B: Comparative Experiment Results: M6 was compared with other well-known domain generalization methods on the CWRU dataset. The experimental results are shown in Table 10, and the confusion matrix is ​​as follows. Figure 5As shown in the table, the M6 ​​method achieved the highest average accuracy, exceeding other methods by 3% to 25%. In six out of eight generalization tasks, the M6 ​​method achieved the highest accuracy. The T-SNE plot demonstrates that the M6 ​​method achieved the best clustering performance compared to other methods, proving its powerful domain generalization diagnostic capability.

[0053] Table 10 Comparative experimental results on the CWRU dataset ERM 81.25 72.66 92.19 83.59 62.5 67.19 74.22 68.75 75.29 DCORAL 89.84 82.03 98.44 92.19 83.59 92.97 100 91.41 91.31 DCC 76.56 76.56 81.25 76.56 85.16 92.97 98.44 94.53 85.25 CNN-C 83.59 71.09 89.84 85.16 67.19 74.22 75.78 65.62 75.56 DRO 71.09 78.12 82.03 78.12 57.03 63.28 70.31 63.28 70.41 VREx 97.66 86.72 100 88.28 82.03 93.75 78.91 77.34 88.09 MixStyle 88.28 86.72 100 97.66 83.59 90.62 95.31 88.28 91.31 M6 95.24 90.62 100 100 90.48 94.05 100 91.67 95.26 C: Hyperparameter Analysis: This invention analyzes hyperparameters on the CWRU dataset. The analysis was conducted, and the experimental results are shown in Table 11. From the table, we can deduce that when... The accuracy on various generalized tasks shows a downward trend. The performance also shows a downward trend across various generalized tasks. The time-depth diagnostic model achieved the best results, so this experiment selected [the appropriate model]. .

[0054] Table 11 Hyperparameter Analysis = 0.5 83.59 77.34 100 98.44 85.94 92.19 100 88.28 90.72 = 1.0 91.41 90.62 100 100 90.62 93.75 100 91.41 94.73 = 2.0 89.06 85.16 100 98.44 88.28 90.62 99.22 89.06 92.48 = 4.0 89.06 82.03 100 99.22 85.16 90.62 100 87.5 91.70 = 7.0 87.50 82.03 98.44 99.22 83.59 90.62 100 87.5 91.11 4: Experimental Results of the HIT Aero-engine Bearing Failure Dataset A: Ablation Experiment Results: Experimental results for various ablation models on the HIT dataset are shown in Figure 12. Similarly, the table shows that the M6 ​​method achieved the highest average accuracy and also performed best in most generalization tasks, with an average accuracy 1%–6% higher than other deep diagnostic models. Comparison of M1, M2, and M3 reveals that the improved IRM achieved the highest accuracy, further demonstrating the superior performance of the improved IRM. Comparison of M4, M5, and M1 shows that hybrid feature extraction outperforms single feature extraction, further demonstrating the importance of balancing generalization and discriminative features.

[0055] Table 12 Ablation experimental results on the HIT dataset M1 89.06 99.22 93.36 71.88 88.38 M2 92.19 97.27 96.48 70.7 89.16 M3 95.7 100 96.09 68.75 90.14 M4 91.8 99.22 89.45 71.48 87.99 M5 87.5 100 82.42 71.48 85.35 M6 94.95 100 98.44 71.48 91.22 B: Comparison of Experimental Results: Two sets of generalization experiments were conducted on the HIT dataset. The first set of experiments generalized three domains to an invisible target domain, as shown in Table 13. The second set of experiments generalized two domains to two invisible target domains, as shown in Table 14. The experimental results are shown in the tables, and the T-SNE visualization is shown below. Figure 6As shown in the figure, in the first set of generalization experiments, method M6 achieved the highest average accuracy, exceeding other methods by 1%–6%. M6 achieved the highest accuracy in three out of the four generalization tasks. In the second set of generalization experiments, it again achieved the highest average accuracy, exceeding other methods by 2%–5%. The T-SNE plot reveals that by fusing discriminative features and generalization feature extraction, the M6 ​​model reduces the intra-class distance between various faults and increases the inter-class distance, resulting in the best clustering performance. The results of these two sets of generalization experiments further demonstrate the strong domain generalization capability of the proposed method.

[0056] Table 13 Comparative experimental results on the HIT dataset (1) ERM 89.54 98.05 91.02 71.88 87.62 DCORAL 80.47 99.22 86.33 71.48 84.38 DDC 84.38 96.48 95.31 70.7 86.72 CNN-C 90.62 100 96.88 71.88 89.85 DRO 74.61 88.28 85.53 71.48 79.98 VREx 72.27 83.59 80.47 71.09 76.86 MixStyle 89.06 99.22 93.36 71.88 88.38 M6 94.53 100 98.44 71.48 91.11 Table 14 Comparative experimental results on the HIT dataset (2) ERM 75.36 70.12 79.11 75.99 75.14 DCORAL 74.81 71.88 71.68 76.56 73.73 DDC 80.06 67.78 72.47 79.11 74.85 CNN-C 80.45 68.95 72.66 75.39 74.36 DRO 74.61 68.95 71.88 76.17 72.90 VREx 69.14 70.51 71.88 75.20 71.68 MixStyle 73.83 69.73 77.93 75.78 74.32 M6 80.86 73.05 81.45 76.18 77.88 5: Experimental results of bearing vibration test bench A: Ablation Experiment Results: The ablation experiment results on the bearing vibration test bench at the experimental center are shown in Table 15. The experimental results show that the proposed method achieved the highest average accuracy, exceeding other methods by 2%–5%. Comparison of M1, M2, and M3 reveals that the improved IRM performs best in accuracy, further validating its superior performance. Further comparison of M4, M5, and M1 shows that hybrid feature extraction significantly outperforms single feature extraction, highlighting the importance of simultaneously focusing on generalization and discriminative features.

[0057] B: Comparison of experimental results: The ablation test results on the bearing vibration test bench at the experimental center are as follows: Figure 7 As shown, the confusion matrix and TSNE visualization of some comparative experiments are as follows: Figure 8 As shown in the figures, these experimental results clearly demonstrate that the proposed method maintains superior performance in generalization diagnosis, improving upon other methods by 5%–8%. Analysis of the confusion matrix further confirms the high classification accuracy of the proposed method, significantly reducing the misdiagnosis rate. Visualization of the T-SNE plot reveals that the proposed method, by considering both generalization and discriminative features, more effectively clusters fault samples, achieving larger inter-class margins and smaller intra-class margins, thus obtaining superior diagnostic performance.

[0058] Table 15 Ablation test results on the bearing vibration test bench dataset M1 76.31 90.39 76.19 94.4 97.92 94.05 81.9 92.02 81.48 87.18 M2 73.69 91.55 90.97 85.36 97.8 86.43 81.07 97.86 86.69 87.94 M3 73.93 96.06 96.18 95.60 99.88 92.5 79.76 99.17 81.48 90.51 M4 77.4 92.57 92.01 81.43 94.56 84.64 77.4 94.29 84.03 86.48 M5 77.02 95.16 84.49 95.71 99.31 92.50 79.88 91.9 72.22 87.58 M6 77.4 95.27 97.34 97.5 99.88 96.19 82.38 99.64 87.86 92.61 The above is one embodiment of the rolling bearing domain generalized fault diagnosis method provided in this embodiment. Based on the same idea, this embodiment also provides a corresponding rolling bearing domain generalized fault diagnosis system. Specific limitations of the rolling bearing domain generalized fault diagnosis system can be found in the limitations of the rolling bearing domain generalized fault diagnosis method described above, and will not be repeated here. Each module in the above-described rolling bearing domain generalized fault diagnosis system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0059] This embodiment also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A generalized fault diagnosis method for rolling bearing domains is provided.

[0060] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0061] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A generalized fault diagnosis method for rolling bearings, characterized in that, include: The time-domain signal of the rolling bearing vibration signal is obtained, and the time-domain signal is converted to frequency to obtain the frequency-domain signal; A fault diagnosis model is constructed; the fault diagnosis model includes a hybrid feature network and a fault classifier; the hybrid feature network includes an IBN network and a Mixstyle network. The time-domain signal and the frequency-domain signal are jointly input into the hybrid feature network. The discriminative features in the frequency-domain signal are extracted by the IBN network, and the generalization features in the time-domain signal are extracted by the MixStyle network. The discriminative features extracted by the IBN network and the generalization features extracted by the MixStyle network are concatenated to form a high-dimensional feature vector. The spliced ​​feature signals are input into the fault classifier for fault diagnosis. It also includes the optimization of feature extractor parameters, including the following steps: Confirm according to different fault types individual working conditions Each working condition Experience risk is ; Construct the IRM expression; adjust the weights of the invariant constraints in the IRM expression using cross-entropy loss, with the initial weights as follows: ； The adjusted weights are: ； In the formula, For update rate; By giving more invariant constraints to the worse-performing operating conditions, we can obtain improved invariant constraints. The improved IRM formula is as follows: ； In the formula, This represents the balance parameter, used to adjust the weight between empirical risk and invariant constraints. Indicates when When, the loss function is related to gradient, The second norm of the gradient is represented. This represents the weight of the invariance constraint. Representative feature extractor, Represents a fault classifier. For all source domain operating conditions; In practical small-batch optimization, unbiased estimates are used instead of invariant constraints, as shown in the formula: ； In the formula, and It comes from the working condition Two random mini-batch samples, size ; It is a loss function. For improved invariance constraints; The empirical risk is set as the cross-entropy loss of the classifier, as shown in the formula: ； The formula for the joint optimization objective is: ； The optimization strategy is as follows: ； ； In the formula, For learning rate, These are the current feature extractor parameters. For the updated feature extractor parameters, These are the current classifier parameters. The updated classifier parameters; The construction of the hybrid feature network includes the following steps: Combining IN and BN networks, IN networks are added to the shallow layers of the CNN network while BN networks are retained in the deeper layers. Instance normalization and batch normalization are then applied together within the CNN network; for a given input feature... The processes of instance normalization and batch normalization are as follows: ； ； in, and These are the mean and standard deviation calculated during instance normalization, respectively. and This is a batch-normalized statistic; Through weighted combination, the final output feature is: ； in, It is in The hyperparameters of the interval are used to adjust the balance between instance normalization and batch normalization; The MixStyle network is applied to the shallow layer of the CNN network. In the shallow layer, MixStyle generates sample features under new working conditions by mixing the working condition features of existing samples. For a given input sample and , computational feature representation and And extract the mean and standard deviation from them, that is: ； ； By combining these statistical information, new mean and standard deviation are generated: ； in, It is in arrive The hyperparameters between these parameters are used to adjust the mixing ratio of the samples; Style transfer of features is performed using a mixture of mean and standard deviation to obtain new feature representations. : ； In the hybrid feature extraction network, the discriminative features in the phase frequency signal are extracted through the IBN network, and the generalization features in the time domain signal are extracted through the Mixstyle network. The two are then concatenated and input into the fault classifier.

2. The rolling bearing domain generalized fault diagnosis method as described in claim 1, characterized in that, The time-frequency conversion of the time-domain signal includes the following steps: The phase frequency characteristic is obtained through Fourier transform, and the expression for the phase value is: ； ； In the formula, They are The real and imaginary parts are given, and the Fourier transform is calculated using the Fast Fourier Transform algorithm.

3. A rolling bearing domain generalized fault diagnosis system, characterized in that, The system includes: processor; A memory on which computer programs that can run on the processor are stored; When the computer program is executed by the processor, it implements the steps of the rolling bearing domain generalized fault diagnosis method as described in any one of claims 1 to 2.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a data processing program, which, when executed by a processor, implements the steps of the rolling bearing domain generalized fault diagnosis method as described in any one of claims 1 to 2.

Images