Rotating machinery cooperative cross-domain fault diagnosis system and method based on dual-domain signals

Abstract

The application discloses a rotating machinery cooperative cross-domain fault diagnosis system and method based on a dual-domain signal, relates to the technical field of rotating machinery vibration signal fault diagnosis, and comprises the following steps: based on a dual-domain signal feature extractor, time-domain and frequency-domain signal features are extracted; the signal features collected by each layer of time domain and frequency domain are combined, and dual-domain fusion feature extraction is performed based on a dual-domain fusion feature extractor; a Wasserstein distance measurer is used to measure the similarity of features between source domains and target domains; dual-domain features are fused and extracted layer by layer, and then classified by using a fault classifier and a domain classifier respectively. The application utilizes the complementary characteristics of time-domain and frequency-domain signals, has strong domain alignment capability, adopts a Swin-Transformer parallel fusion network to simultaneously extract and fuse dual-domain features, realizes domain self-adaptive technology combining adversarial training and Wasserstein distance, eliminates the influence of speed fluctuation on feature extraction, and realizes cooperative cross-domain fault diagnosis of bearings under rotating machinery speed fluctuation conditions.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for vibration signals of rotating machinery, and in particular to a collaborative cross-domain fault diagnosis system and method for rotating machinery based on dual-domain signals. Background Technology

[0002] The health condition of bearings directly affects the stability and reliability of equipment operation, and fault diagnosis methods that utilize signal processing techniques to extract fault features have been widely applied in this field. Many researchers have investigated different fault diagnosis methods to improve the computational efficiency, noise adaptability, and robustness of the diagnostic process. However, the application of these intelligent algorithms is limited by the independence and identical distribution of training and test samples. Therefore, unsupervised adaptive methods have emerged as a valuable approach to overcome this limitation.

[0003] Unsupervised domain adaptation, by reducing the distance between labeled source domain samples and unlabeled target domain samples, addresses the dependency problem of labeled data in traditional methods and has become an important research area in fault diagnosis. However, in engineering practice, the operating speed of mechanical equipment often fluctuates, leading to a complex mapping relationship between signal features and fault modes. This problem poses a significant challenge to cross-domain fault diagnosis in machinery. To address this issue, many researchers have explored the application of unsupervised domain adaptation in fault diagnosis under fluctuating speed conditions. Various methods have been proposed to improve the performance of unsupervised domain adaptive fault diagnosis. While these methods help improve the accuracy of fault diagnosis under fluctuating conditions, most are limited to enabling the network to learn and recognize frequency domain signal features or time domain signal features.

[0004] Therefore, there is still much room for improvement in unsupervised adaptive fault diagnosis under speed fluctuation conditions. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention discloses a collaborative cross-domain fault diagnosis system and method for rotating machinery based on dual-domain signals.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The first aspect of the present invention provides a rotating machinery collaborative cross-domain fault diagnosis system based on dual-domain signals.

[0008] In one embodiment, a rotating machinery collaborative cross-domain fault diagnosis system based on dual-domain signals includes:

[0009] A dual-domain signal feature extractor includes a time-domain signal feature extraction module, a frequency-domain signal feature extraction module, and a central network feature extraction module to achieve dual-domain signal feature extraction;

[0010] The dual-domain fusion feature extractor is used to fuse signal features and input them into the central network, so that the information captured by the shallow network is gradually transmitted to the deep network.

[0011] The Wasserstein distance metric is used to measure the similarity of features between the source and target domains, enhancing the robustness of domain-invariant features.

[0012] A fault classifier is used to classify different fault types.

[0013] A domain classifier is used to identify the source domain and the target domain.

[0014] A second aspect of the present invention provides a diagnostic method based on the above-described system.

[0015] In one embodiment, a rotating machinery collaborative cross-domain fault diagnosis method based on dual-domain signals includes the following steps:

[0016] S1. Based on a dual-domain signal feature extractor, extract signal features in the time domain and frequency domain;

[0017] S2. Merge the signal features acquired in each time domain and frequency domain in step S1, and perform dual-domain fusion feature extraction based on the dual-domain fusion feature extractor.

[0018] S3. Use the Wasserstein distance metric to measure the similarity of features between the source and target domains;

[0019] S4. Perform layer-by-layer fusion and extraction of the dual-domain features, and then classify them using a fault classifier and a domain classifier respectively.

[0020] Before step S1, different bearing health states are set, and the motor speed is randomly fluctuated within a certain range by manually adjusting the frequency converter, thereby obtaining the speed fluctuation condition, and collecting the vibration signal of different faulty bearings under the speed fluctuation condition.

[0021] Optionally, step S1, which involves extracting time-domain and frequency-domain signal features based on a dual-domain signal feature extractor, specifically includes:

[0022] S11. Divide the time-domain and frequency-domain signals into multiple patches and embed them based on a linear mapping, as follows:

[0023] ;

[0024] In the formula, x represents the patch after signal segmentation. For embedded patches;

[0025] S12. Input the embedded patches into the Swin-Transformer-based feature extractor to perform dual-domain signal feature extraction.

[0026] Optionally, in step S12, the dual-domain signal feature extraction employs independent channels to extract dual-domain signal features. Time-domain signal features are extracted through a time-domain signal feature extraction module, and frequency-domain signal features are extracted through a frequency-domain signal feature extraction module. These are then connected to the central network via nodes. The dual-domain signal feature extractor comprises three distinct parts: a time-domain signal feature extraction module, a frequency-domain signal feature extraction module, and a central network feature extraction module. At each stage, time-domain and frequency-domain signal feature extraction modules are established and connected to the central network via nodes for convenient information transmission. Furthermore, features from each stage of the central network are also input to nodes for information exchange. Considering that interference between different types of signals cannot be ignored, the intersection of information should not be limited to isolated early or late-stage fusion in deep structures; multi-level feature fusion is employed.

[0027] After feature extraction in step S1, the signal features are fused using average pooling and then input into the central network, thereby gradually transferring the information captured by the shallow network to the deep network. It is worth noting that, except for step S1, the input to each central network stage includes the fusion of features from three different parts of the dual-domain signal feature extractor, which allows for effective and comprehensive processing of signal information.

[0028] Optionally, in step S2, during the dual-domain fusion feature extraction process, multiple consecutive Swin-Transformer modules are selected as feature extraction networks, and a linear mapping method using convolutional neural networks as model input is adopted. Each Swin-Transformer module consists of multi-head self-attention (W-MSA) based on conventional window partitioning, multi-head self-attention (SW-MSA) based on shifted window partitioning, multiple layer normalized networks (LN), and multilayer perceptrons (MLP). The signal features are windowed before entering W-MSA. W-MSA performs multi-head self-attention calculation on a single window, while SW-MSA performs multi-head self-attention calculation on a sliding back window. Residual connections are applied after each Swin-Transformer module to achieve patch merging.

[0029] Optionally, a relative positional bias is introduced into the multi-head self-attention (W-MSA) based on conventional window partitioning and the multi-head self-attention (SW-MSA) based on shifted window partitioning to distinguish them from traditional multi-head self-attention mechanisms, as shown in the following equation:

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] In the formula, l represents the output of the Lth layer, and l+1 represents the (L+1)th layer;

[0035] ;

[0036] In the formula, B is the relative position code, Q, K, and V represent query, key, and value, respectively, and d represents a constant.

[0037] Optionally, in step S3, the Wasserstein distance metric is used to measure the similarity of features between the source and target domains, as shown in the following formula:

[0038] ;

[0039] In the formula, It is all joint distributions The set, The marginal probability is and .

[0040] Optionally, in step S4, during the classification process, the output of the central network is input into the domain classifier and the fault classifier. A fault classifier is constructed using a fully connected layer to obtain the mapping relationship between collaborative features and fault classes. Softmax cross-entropy loss is selected to evaluate the cross-entropy between the predicted probability and the actual sample data, defined as follows:

[0041] ;

[0042] In the formula, It is an indicator function. It is the Kth value of the predicted distribution, and the number of label types is defined as K;

[0043] Adversarial training is performed using a domain classifier to improve the model's ability to acquire domain-invariant features. A gradient inversion layer is placed before the domain classifier. During adversarial training, the domain classifier gradually becomes confused by the dual-domain fusion feature extractor. Specifically, when the dual-domain fusion feature extractor effectively extracts domain-invariant features, the domain classifier struggles to distinguish samples from different domains. The domain adversarial loss formula is as follows:

[0044] ;

[0045] The output features of the central network are quantified using the Wasserstein distance metric module to enhance the robustness of domain-invariant features. The distance loss is defined as follows:

[0046] ;

[0047] In the formula, It is the output of the feature extractor.

[0048] The beneficial effects of this invention are that it utilizes the complementary characteristics of time-domain and frequency-domain signals, exhibiting strong domain alignment capabilities. The model employs a Swin-Transformer parallel fusion network to simultaneously extract and fuse dual-domain features, and implements a domain adaptation technique combining adversarial training and Wasserstein distance. This eliminates the influence of speed fluctuations on feature extraction, accurately and intelligently achieving collaborative cross-domain fault diagnosis of bearings under rotating machinery speed fluctuation conditions. Attached Figure Description

[0049] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1

[0052] A collaborative cross-domain fault diagnosis system for rotating machinery based on dual-domain signals includes:

[0053] A dual-domain signal feature extractor includes a time-domain signal feature extraction module, a frequency-domain signal feature extraction module, and a central network feature extraction module to achieve dual-domain signal feature extraction;

[0054] The dual-domain fusion feature extractor is used to fuse signal features and input them into the central network, so that the information captured by the shallow network is gradually transmitted to the deep network.

[0055] The Wasserstein distance metric is used to measure the similarity of features between the source and target domains, enhancing the robustness of domain-invariant features.

[0056] A fault classifier is used to classify different fault types.

[0057] A domain classifier is used to identify the source domain and the target domain.

[0058] Example 2

[0059] A collaborative cross-domain fault diagnosis method for rotating machinery based on dual-domain signals includes the following steps:

[0060] S0. Set different bearing health states, and control the motor speed to fluctuate randomly within a certain range by manually adjusting the frequency converter, thereby obtaining the speed fluctuation condition, and collecting the vibration signal of different faulty bearings under the speed fluctuation condition.

[0061] S1. Based on a dual-domain signal feature extractor, extract signal features in the time domain and frequency domain;

[0062] Specifically, it includes:

[0063] S11. Divide the time-domain and frequency-domain signals into multiple patches and embed them based on a linear mapping, as follows:

[0064] ;

[0065] In the formula, x represents the patch after signal segmentation. For embedded patches;

[0066] S12. Input the embedded patches into the Swin-Transformer-based feature extractor to perform dual-domain signal feature extraction.

[0067] The dual-domain signal feature extraction employs independent channels to extract dual-domain signal features. Time-domain signal features are extracted through a time-domain signal feature extraction module, and frequency-domain signal features are extracted through a frequency-domain signal feature extraction module. These are then connected to the central network via nodes. The dual-domain signal feature extractor comprises three distinct parts: a time-domain signal feature extraction module, a frequency-domain signal feature extraction module, and a central network feature extraction module. At each stage, time-domain and frequency-domain signal feature extraction modules are established and connected to the central network via nodes for information transmission. Furthermore, features from each stage of the central network are also input to nodes for information exchange. Considering the significant interference between different types of signals, the intersection of information should not be limited to isolated early or late-stage fusion within deep structures; therefore, multi-level feature fusion is employed.

[0068] After feature extraction in step S1, the signal features are fused using average pooling and then input into the central network, thereby gradually transferring the information captured by the shallow network to the deep network. It is worth noting that, except for step S1, the input to each central network stage includes the fusion of features from three different parts of the dual-domain signal feature extractor, which allows for effective and comprehensive processing of signal information.

[0069] S2. Merge the signal features acquired in the time domain and frequency domain of each layer in step S1, and perform dual-domain fusion feature extraction.

[0070] In the dual-domain fusion feature extraction process, multiple consecutive Swin-Transformer modules are selected as the feature extraction network. A linear mapping method using convolutional neural networks as model input is adopted. Each Swin-Transformer module consists of multi-head self-attention (W-MSA) based on conventional window partitioning, multi-head self-attention (SW-MSA) based on shifted window partitioning, a multi-layer normalized network (LN), and a multilayer perceptron (MLP). The signal features are windowed before entering W-MSA. W-MSA performs multi-head self-attention calculation on a single window, while SW-MSA performs multi-head self-attention calculation on a sliding back window. Residual connections are applied after each Swin-Transformer module to achieve patch merging. Relative positional bias is introduced in W-MSA based on conventional window partitioning and SW-MSA based on shifted window partitioning to distinguish them from traditional multi-head self-attention mechanisms, as shown in the following equation:

[0071] ;

[0072] ;

[0073] ;

[0074] ;

[0075] In the formula, l represents the output of the Lth layer, and l+1 represents the (L+1)th layer;

[0076] ;

[0077] In the formula, B is the relative position code, Q, K, and V represent query, key, and value, respectively, and d represents a constant.

[0078] S3. Use Wasserstein distance to measure the similarity of features between the source and target domains, as shown in the following formula:

[0079] ;

[0080] In the formula, It is all joint distributions The set, The marginal probability is and ;

[0081] S4. Perform layer-by-layer fusion and extraction of the dual-domain features, and then classify them using a fault classifier and a domain classifier respectively.

[0082] During the classification process, the output of the central network is input into the domain classifier and the fault classifier. A fully connected layer is used to construct the fault classifier to obtain the mapping relationship between collaborative features and fault classes. Softmax cross-entropy loss is chosen to evaluate the cross-entropy between the predicted probability and the actual sample data, defined as follows:

[0083] ;

[0084] In the formula, It is an indicator function. It is the Kth value of the predicted distribution, and the number of label types is defined as K;

[0085] Adversarial training is performed using a domain classifier to improve the model's ability to acquire domain-invariant features. A gradient inversion layer is placed before the domain classifier. During adversarial training, the domain classifier gradually becomes confused by the dual-domain fusion feature extractor. Specifically, when the dual-domain fusion feature extractor effectively extracts domain-invariant features, the domain classifier struggles to distinguish samples from different domains. The domain adversarial loss formula is as follows:

[0086] ;

[0087] The output features of the central network are quantified using the Wasserstein distance metric module to enhance the robustness of domain-invariant features. The distance loss is defined as follows:

[0088] ;

[0089] In the formula, It is the output of the feature extractor.

[0090] Application examples

[0091] The invention is further described by acquiring vibration signals and performing intelligent diagnosis on a specially designed bearing fault test bench under speed fluctuations.

[0092] Five bearing health states were considered: normal bearing, inner ring fault, outer ring fault, rolling element fault, and combined outer ring and rolling element fault. The sampling frequency was set to 25.6 kHz. Bearing acceleration signals were collected as datasets under four stable operating conditions (1000 r / min, 1500 r / min, 1800 r / min, 2000 r / min, 1000–1500 r / min, 1500–1800 r / min, 1800–2000 r / min, and 500–3000 r / min) and fluctuating operating conditions. Each health state had 200 frequency domain samples and 200 time domain samples, with each sample containing 1024 sample points. 50% of the dataset was used as the training set, and the remainder as the test set.

[0093] To evaluate the effectiveness of this method, three state-of-the-art cross-domain fault diagnosis methods were compared: Domain Adversarial Neural Network (DANN), Hybrid Distance-Guided Adversarial Network (HDAN), and Improved Deep Subdomain Adaptive Network (MDSAN). The diagnostic accuracy of the four methods was compared across six transfer tasks: 1000 rpm → 1000 rpm ~ 1500 rpm (Task 1), 1500 rpm → 1000 rpm ~ 1500 rpm (Task 2), 1500 rpm → 1500 rpm ~ 1800 rpm (Task 3), 1800 rpm → 1800 rpm ~ 2000 rpm (Task 4), 1800 rpm → 1800 rpm ~ 2000 rpm (Task 5), and 2000 rpm → 1800 rpm ~ 2000 rpm (Task 6). The left side of the arrow represents the source domain, and the right side represents the target domain. The comparison results of the diagnostic accuracy of these four methods are shown in Table 1 below.

[0094] Table 1. Comparison of diagnostic accuracy of the four methods

[0095]

[0096] It can be seen that DANN's diagnostic accuracy is significantly lower than other methods, falling below 90% in most diagnostic tasks. In contrast, HDAN's diagnostic accuracy is above 90% in all tasks, while MDSAN's diagnostic accuracy in task 4 is even higher than HDAN's. Specifically, MDSAN performs well, achieving diagnostic accuracy above 95% in tasks 1, 5, and 6. However, the method of this invention outperforms the three comparative methods in all tasks, with an average diagnostic accuracy exceeding 98%. This demonstrates that the method of this invention is significantly effective in diagnosing migrating faults under conditions of fluctuating bearing speeds.

[0097] This invention employs a multi-level fusion feature extraction module based on the Swin-Transformer and dual-domain signals. This module extracts representative features from the dual-domain signals, mitigating the impact of rotational speed fluctuations on fault diagnosis in rotating machinery. Then, an adversarial training strategy and Wasserstein distance are used to promote domain alignment, thereby extracting domain-invariant features of mechanical equipment faults. Experimental results in bearing fault diagnosis demonstrate that this invention achieves higher diagnostic accuracy and better domain-invariant feature extraction capabilities compared to existing methods.

[0098] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A diagnostic method for a rotating machinery collaborative cross-domain fault diagnosis system based on dual-domain signals, characterized in that, The diagnostic system includes: A dual-domain signal feature extractor includes a time-domain signal feature extraction module, a frequency-domain signal feature extraction module, and a central network to achieve dual-domain signal feature extraction. The dual-domain fusion feature extractor is used to fuse signal features and input them into the central network, so that the information captured by the shallow network is gradually transmitted to the deep network. The Wasserstein distance metric is used to measure the similarity of features between the source and target domains, enhancing the robustness of domain-invariant features. A fault classifier is used to classify different fault types. Domain classifiers are used to identify source and target domains. The diagnostic method based on the above diagnostic system includes the following steps: S1. Based on a dual-domain signal feature extractor, extract signal features in the time domain and frequency domain; S2. Merge the signal features acquired in each time domain and frequency domain in step S1, and perform dual-domain fusion feature extraction based on the dual-domain fusion feature extractor. S3. Use the Wasserstein distance metric to measure the similarity of features between the source and target domains; S4. Perform layer-by-layer fusion and extraction of the dual-domain features, and then classify them using a fault classifier and a domain classifier respectively; Step S1, which involves extracting time-domain and frequency-domain signal features based on a dual-domain signal feature extractor, specifically includes: S11. Divide the time-domain and frequency-domain signals into multiple patches and embed them based on a linear mapping, as follows: ； In the formula, x represents the patch after signal segmentation. For embedded patches; S12. Input the embedded patches into the Swing-Transformer-based feature extractor to perform dual-domain signal feature extraction. In step S2, during the dual-domain fusion feature extraction process, multiple consecutive Swin-Transformer modules are selected as feature extraction networks. A linear mapping method using convolutional neural networks as model input is adopted. Each Swin-Transformer module consists of multi-head self-attention (W-MSA) based on conventional window partitioning, multi-head self-attention (SW-MSA) based on shifted window partitioning, multiple layer normalized networks (LN), and multilayer perceptrons (MLP). The signal features are windowed before entering W-MSA. W-MSA performs multi-head self-attention calculation on a single window, while SW-MSA performs multi-head self-attention calculation on a sliding back window. Residual connections are applied after each Swin-Transformer module to achieve patch merging.

2. The diagnostic method of the rotating machinery collaborative cross-domain fault diagnosis system based on dual-domain signals as described in claim 1, characterized in that, In step S12, the dual-domain signal feature extraction uses an independent channel. The time-domain signal features are extracted by the time-domain signal feature extraction module, and the frequency-domain signal features are extracted by the frequency-domain signal feature extraction module. The time-domain signal feature extraction module and the frequency-domain signal feature extraction module are connected to the central network through nodes.

3. The diagnostic method of the rotating machinery collaborative cross-domain fault diagnosis system based on dual-domain signals as described in claim 1, characterized in that, The multi-head self-attention (W-MSA) based on conventional window partitioning and the multi-head self-attention (SW-MSA) based on shifted window partitioning introduce a relative positional bias, as shown in the following formula: ； ； ； ； In the formula, l represents the output of the Lth layer, and l+1 represents the (L+1)th layer; ； In the formula, B is the relative position offset matrix, Q, K, and V represent query, key, and value, respectively, and d represents a constant.

4. The diagnostic method of the rotating machinery collaborative cross-domain fault diagnosis system based on dual-domain signals as described in claim 1, characterized in that, Step S3 involves using the Wasserstein distance metric to measure the similarity of features between the source and target domains. The formula is as follows: ； In the formula, It is all joint distributions The set, The marginal probability is and .

5. The diagnostic method of the rotating machinery collaborative cross-domain fault diagnosis system based on dual-domain signals as described in claim 1, characterized in that, Step S4, which involves classifying data using both a fault classifier and a domain classifier, specifically includes: S41. During the classification process, the output of the central network is input into the domain classifier and the fault classifier. The fault classifier is constructed using a fully connected layer to obtain the mapping relationship between collaborative features and fault classes. S42. Adversarial training using a domain classifier is employed to improve the model's ability to acquire domain-invariant features. The domain adversarial loss formula is as follows: ； S43. The output features of the central network are quantified using the Wasserstein distance metric.

6. The diagnostic method of the rotating machinery collaborative cross-domain fault diagnosis system based on dual-domain signals as described in claim 5, characterized in that, Step S41 also includes selecting the Softmax cross-entropy loss to evaluate the cross-entropy between the predicted probability and the actual sample data. The Softmax cross-entropy loss is defined as follows: ； In the formula, It is an indicator function. It is the Kth value of the predicted distribution, and the number of label types is defined as K.

7. The diagnostic method of the rotating machinery collaborative cross-domain fault diagnosis system based on dual-domain signals as described in claim 5, characterized in that, Step S43 also includes distance loss, defined as follows: ； In the formula, It is the output of the feature extractor.

Images