Fault diagnosis method and device based on cross disturbance domain self-adaption and medium

Abstract

The invention relates to the technical field of fault diagnosis, and discloses a fault diagnosis method and device based on cross disturbance domain self-adaption and a medium, and the fault diagnosis method comprises the following steps: obtaining deep features of original data; inputting the deep feature into a cross Gaussian disturbance unit to obtain a first disturbance feature and a second disturbance feature; generating a fault classifier based on a deep learning network according to the first disturbance feature; and reconstructing the second disturbance characteristic signal to generate judgment data, and inputting the judgment data into the fault classifier to form a fault category probability. According to the fault diagnosis method, the diagnosis precision is improved.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis technology, and in particular to a fault diagnosis method, device and medium based on cross-disturbance domain adaptation. Background Technology

[0002] Fault diagnosis and prediction for rotating machinery such as train engines, aircraft engines, and steam turbine generators play a crucial role in their safe operation. Rolling bearings are generally considered key components of rotating machinery, making their health monitoring and fault prediction an important area of ​​research and application in sparse decomposition, manifold learning, and deep learning. However, due to data distribution biases, most existing methods trained in a single working scenario are incompatible with other working scenarios. Their dependence on the same data distribution remains a challenge in the field of intelligent diagnostics. Fortunately, to address the performance degradation problem across multiple working scenarios, transfer learning, which aims to obtain domain-invariant models between two related but different domains, offers a new approach to improving scenario generalization.

[0003] In recent years, intelligent fault diagnosis methods based on transfer learning have been extensively studied. Existing transfer learning-based diagnostic methods typically input monitoring data from both the source and target domains into a shared domain network to extract transferable features, thus transferring category knowledge from the source domain to the target domain. However, existing domain adaptation methods mainly employ domain adversarial networks as a distribution alignment mechanism, which easily leads to the neglect of local feature alignment. Furthermore, the batch normalization parameters are fixed in the source domain and cannot be dynamically adjusted with the target domain data, affecting the transfer effect. Therefore, improving diagnostic accuracy is a technical problem that needs to be solved. Summary of the Invention

[0004] This application provides a fault diagnosis method, device, and medium based on cross-disturbance domain adaptation, which improves the diagnostic accuracy.

[0005] To achieve the above objectives, the main technical solutions adopted in this application include:

[0006] In a first aspect, embodiments of this application provide a fault diagnosis method based on cross-perturbation domain adaptation, comprising the following steps: acquiring deep features of raw data; inputting the deep features into a cross-Gaussian perturbation unit to acquire a first perturbation feature and a second perturbation feature; generating a fault classifier based on a deep learning network based on the first perturbation feature; reconstructing the second perturbation feature signal to generate judgment data, and inputting the judgment data into the fault classifier to form a fault category probability.

[0007] The first aspect of this application proposes a fault diagnosis method based on cross-disturbance domain adaptation. By generating a first disturbance feature and a second disturbance feature through cross-Gaussian disturbance units, it can more comprehensively capture the multi-scale and multi-dimensional fault information contained in the original data, effectively reduce the interference of data distribution fluctuations under varying operating conditions of rotating machinery on feature representation, and enhance the robustness of features in judging fault modes. The first disturbance feature is used to construct a classifier, and the second disturbance feature signal is reconstructed to assist in the judgment. It not only relies on the deep discriminativeness of the first disturbance feature to lay an accurate classification foundation for the fault classifier, but also makes up for the information loss problem caused by adversarial learning in traditional domain adaptation methods through signal reconstruction. This makes the judgment data more complete in preserving the details of fault features, further reducing misjudgment and omission, and improving the diagnostic accuracy.

[0008] Optionally, the step of inputting the deep features into the cross-Gaussian perturbation unit to obtain the first perturbation feature and the second perturbation feature includes:

[0009] Deep features are divided equally by channel and then cross-stitched, and weighted features are generated using Softmax gating;

[0010] The perturbation weights are determined based on the weighted characteristics, and the recalibration characteristics are determined based on the perturbation weights.

[0011] In the above scheme, deep features are evenly divided into channels and then cross-stitched, which helps to integrate the local fault details contained in each channel, such as high-frequency impacts and low-frequency fluctuations in rotating machinery vibration signals, avoiding the limitations of feature representation caused by the one-sided information of a single channel. At the same time, the introduction of Softmax gating to generate weighted features can dynamically highlight key features strongly correlated with fault modes and suppress redundant interference information through adaptive weight allocation, reducing the influence of irrelevant signals caused by operating condition fluctuations, and significantly improving the fault discriminativeness of weighted features. It enhances the feature's generalization ability to data distribution fluctuations under varying operating conditions through reasonable perturbation, while avoiding excessive perturbation that destroys core fault information. The recalibrated features further strengthen the representation intensity of key fault information through the channel-by-channel effect of weights and weighted features, ultimately making the output first and second perturbation features both comprehensive and accurate, providing high-quality feature support for subsequent fault classifier construction and signal reconstruction, effectively alleviating the problems of insufficient local feature alignment and weak feature anti-interference ability in existing domain adaptation methods, and improving the fault diagnosis accuracy in cross-operating condition scenarios.

[0012] Optionally, determining the perturbation weights based on the weighted characteristics includes:

[0013] Calculate the unimodal Gaussian weights for each of the weighted features;

[0014] The single-peak Gaussian weights are normalized to obtain the perturbation weights.

[0015] In the above scheme, calculating unimodal Gaussian weights for each weighted feature accurately captures the core dimensions strongly correlated with fault modes, significantly improving the focus of the weights on key fault information. On the other hand, normalizing the unimodal Gaussian weights unifies the weights of different dimensions to the same numerical range, eliminating the imbalance in weight scale caused by differences in the numerical ranges of the various dimensions of the weighted features. It also ensures the comparability and synergy of the perturbation weights for each dimension, enabling subsequent feature recalibration based on these weights to accurately apply to each feature dimension at a balanced scale. This enhances the feature's generalization ability to fluctuating data distributions under varying operating conditions through reasonable perturbation without compromising the integrity of core fault information. This effectively solves the problems of weak perturbation weight targeting and excessive perturbation of non-fault features in existing domain adaptation methods, contributing to improved accuracy and robustness of fault diagnosis across different operating conditions.

[0016] Optionally, determining the recalibration features based on the perturbation weights includes:

[0017] The recalibrated features are obtained by multiplying the weighted features one channel at a time.

[0018] In the above scheme, each channel's weighted feature can be matched with a dedicated perturbation weight, achieving a precise correspondence between feature dimensions and weight dimensions. This reduces the problem of key channels being weakened and interference channels being over-strengthened, ensuring that the weight's regulation of features is precisely applied to the fault-related dimensions of each channel, allowing the fault information of each channel to be strengthened in accordance with its importance. On the other hand, the weighted features have already preliminarily screened out fault-related information, and the perturbation weight further focuses on the core fault dimensions. After multiplying the two, it can amplify the representation intensity of key fault information, making the signal amplitude of the fault feature frequency band more prominent, and also perform secondary suppression of non-fault interference such as residual operating condition fluctuations and signal noise in the weighted features, significantly improving the distinguishability between faults and interference in the recalibrated features. At the same time, the operation logic of multiplying channel by channel is simple and highly interpretable, avoiding the destruction of core feature information by complex calculations. This ensures that the recalibrated features have stronger resistance to variable operating condition interference while retaining the integrity of the original fault features, effectively alleviating the problems of weak feature recalibration targeting and easy masking of core fault information in existing methods.

[0019] Optionally, the step of dividing deep features equally according to channels and then cross-stitching them includes:

[0020] The deep features are input into the large-scale path and the small-scale path respectively for multi-scale feature extraction to obtain the first output feature and the second output feature. The first output feature and the second output feature are then cross-stitched together.

[0021] In the above scheme, deep features are input into both large-scale and small-scale paths, enabling the targeted capture of key information from different dimensions in rotating machinery vibration signals. The large-scale path effectively extracts low-frequency, global operating condition-related features, avoiding the omission of the impact of operating condition fluctuations on feature distribution due to the limited field of view at the small scale. The small-scale path accurately captures high-frequency, local fault impact features, solving the problem of easily blurred fault details in large-scale extraction and achieving full-dimensional feature coverage. On the other hand, cross-splitting the first and second output features breaks the channel independence of the two scale features, allowing the operating condition adaptation information of the large-scale path and the fault detail information of the small-scale path to be deeply integrated at the channel dimension. This ensures that each new channel simultaneously carries both the operating condition background and the fault signal, avoiding the problem of separation between operating condition and fault information during subsequent Softmax gating weight allocation. This setup helps reduce the perturbation bias caused by isolated scale information and effectively alleviates the problem of insufficient local feature alignment in existing domain adaptation methods.

[0022] Optionally, the acquisition of deep features from the raw data includes:

[0023] Raw vibration signals under multiple operating conditions are collected, and source domain datasets and target domain datasets are formed based on the raw vibration signals.

[0024] In the above scheme, the acquisition of original vibration signals under multiple operating conditions accurately matches the dynamic changes in operating conditions such as speed and load during the actual operation of rotating machinery. This avoids the problem of incomplete scene coverage in traditional single-condition data acquisition. By covering vibration signals under different operating conditions, the scheme can fully capture the changing patterns of fault characteristics in operating condition fluctuations, significantly improving the model's diagnostic generalization ability in actual variable operating condition scenarios. On the other hand, the source domain dataset provides benchmark data with clearly labeled fault categories, laying the foundation for subsequent fault classifier construction based on the first perturbation feature and the transfer of fault category knowledge. The target domain dataset corresponds to the unknown operating condition data to be diagnosed, forming an inter-domain comparison relationship with the source domain data. This gives subsequent domain adaptation operations such as measuring distribution differences by maximum mean difference and dynamically adjusting parameters through adaptive batch normalization a clear target, effectively solving the problem of sharp drop in diagnostic accuracy across operating conditions in traditional diagnostic methods. At the same time, the model learns fault discrimination ability and inter-domain adaptation ability simultaneously during training, ultimately improving the accuracy and reliability of fault diagnosis under variable operating conditions.

[0025] Optionally, the acquisition of deep features from the raw data further includes:

[0026] The original vibration signal is preprocessed using a sliding time window and Z-score normalization to obtain source domain samples and target domain samples.

[0027] In the above scheme, a sliding time window is used to process the original vibration signal, which can segment the continuous temporal vibration signal of rotating machinery into multiple equal-length sample segments. This not only fully preserves the local temporal features contained in the signal, such as the instantaneous impact signal and periodic vibration pattern at the time of failure, but also avoids the problem of dilution or truncation of fault details caused by direct processing of long signals. Furthermore, it generates sufficient training samples, solving the problem of model overfitting due to insufficient original temporal data sample size. At the same time, the equal-length sample format can accurately adapt to the fixed-dimensional input requirements of subsequent multi-layer one-dimensional convolutional networks, laying the data format foundation for the model to stably extract deep features. On the other hand, introducing Z-score standardization to process the segmented samples can effectively eliminate the interference of signal amplitude fluctuations caused by differences in operating conditions. This allows the standardized samples to focus on the feature changes brought about by the fault, such as signal frequency shifts and the addition of harmonic components caused by the fault, rather than irrelevant variables related to the operating conditions. This improves the accuracy and robustness of fault diagnosis across different operating conditions.

[0028] Optionally, the acquisition of deep features from the raw data further includes:

[0029] Construct a multi-layer one-dimensional convolutional network and embed an adaptive batch normalization layer after each convolutional layer, so that the mean and variance of the batch normalization are updated adaptively with the target domain data.

[0030] In the above scheme, by progressively deepening the convolution operation, noise interference can be gradually stripped from the original vibration signal, and multi-dimensional features from shallow local impacts to deep global patterns can be extracted. This avoids the limitation of a single convolutional layer in representing complex fault features. On the other hand, embedding an adaptive batch normalization layer after each convolutional layer can solve the problem of data distribution offset between the source domain and the target domain. The adaptive batch normalization allows the mean and variance to be dynamically updated with the target domain data, and can calibrate the distribution deviation in real time at each stage of feature extraction. The adaptive batch normalization can adjust the parameters in real time to ensure that the features output by each convolutional layer retain the core fault information and are aligned with the distribution characteristics of the target domain, avoiding the problem of "effective fault features being suppressed and operating condition interference being amplified" caused by distribution differences.

[0031] Secondly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor; when the processor runs the computer program, it implements the fault diagnosis method described in any of the above embodiments.

[0032] The electronic device proposed in this application improves the accuracy of fault diagnosis because it can implement the fault diagnosis method described in any of the embodiments.

[0033] Thirdly, embodiments of this application provide a storage medium storing a computer program, which, when executed by a processor, implements the fault diagnosis method of any embodiment.

[0034] The storage medium proposed in this application embodiment can implement the fault diagnosis method of any embodiment, thereby improving the accuracy of fault diagnosis. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0036] Figure 1 This is a schematic diagram of the fault diagnosis method steps in some embodiments of this application;

[0037] Figure 2 This is a schematic diagram of the fault diagnosis method steps in some other embodiments of this application;

[0038] Figure 3 This is a schematic diagram of the fault diagnosis method steps in some other embodiments of this application;

[0039] Figure 4 Here are flowcharts of fault diagnosis methods in some embodiments of this application;

[0040] Figure 5 This is a visual diagram illustrating the target domain fault features extracted by a mechanical fault diagnosis method based on cross-disturbance domain adaptation. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0042] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used in the description of this application is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms "comprising" and "having," and any variations thereof, in the description, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the description, claims, or accompanying drawings of this application are used to distinguish different objects, not to describe a specific order or hierarchy.

[0043] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0044] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "attachment" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0045] In this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, in this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0046] In this application, "multiple" refers to two or more (including two), and similarly, "multiple groups" refers to two or more (including two), and "multiple pieces" refers to two or more (including two).

[0047] Fault diagnosis and prediction for rotating machinery such as train engines, aircraft engines, and steam turbine generators play a crucial role in their safe operation. Rolling bearings are generally considered key components of rotating machinery, making their health monitoring and fault prediction an important area of ​​research and application in sparse decomposition, manifold learning, and deep learning. However, due to data distribution biases, most existing methods trained in a single working scenario are incompatible with other working scenarios. Their dependence on the same data distribution remains a challenge in the field of intelligent diagnostics. Fortunately, to address the performance degradation problem across multiple working scenarios, transfer learning, which aims to obtain domain-invariant models between two related but different domains, offers a new approach to improving scenario generalization.

[0048] In recent years, intelligent fault diagnosis methods based on transfer learning have been extensively studied. Existing transfer learning-based diagnostic methods typically input monitoring data from both the source and target domains into a shared domain network to extract transferable features, thus transferring category knowledge from the source domain to the target domain. However, existing domain adaptation methods mainly employ domain adversarial networks as a distribution alignment mechanism, which easily leads to the neglect of local feature alignment. Furthermore, the batch normalization parameter is fixed in the source domain and cannot be dynamically adjusted with the target domain data, affecting the transfer effect.

[0049] Therefore, the insufficient accuracy of fault diagnosis is a technical problem that urgently needs to be solved.

[0050] In view of this, in order to improve the diagnostic accuracy and robustness under cross-operating conditions / cross-equipment scenarios, this application provides a fault diagnosis method based on cross-perturbation domain adaptation. Deep features are input into a cross-Gaussian perturbation unit to obtain first and second perturbation features; a fault classifier based on a deep learning network is generated based on the first perturbation features; judgment data is reconstructed from the second perturbation feature signal, and the judgment data is input into the fault classifier to form fault category probabilities. This method can simultaneously constrain global distribution differences and local feature differences, and dynamically adjust normalization parameters. It can adaptively formulate different activation schemes according to different input data, improving the model's feature learning ability for different operating conditions and significantly improving the fault diagnosis accuracy of the diagnostic model under varying operating conditions.

[0051] The following is in conjunction with the appendix Figure 1 -Appendix Figure 3 This application provides a specific embodiment of a fault diagnosis method based on cross-disturbance domain adaptation.

[0052] In a first aspect, embodiments of this application provide a fault diagnosis method based on cross-disturbance domain adaptation, comprising the following steps: obtaining deep features of the original data;

[0053] The deep features are input into the cross-Gaussian perturbation unit to obtain the first perturbation feature and the second perturbation feature;

[0054] A fault classifier based on a deep learning network is generated based on the first perturbation feature;

[0055] The second disturbance feature signal is reconstructed to generate judgment data, and the judgment data is input into the fault classifier to form the fault category probability.

[0056] The first aspect of this application proposes a fault diagnosis method based on cross-disturbance domain adaptation. By generating a first disturbance feature and a second disturbance feature through cross-Gaussian disturbance units, it can more comprehensively capture the multi-scale and multi-dimensional fault information contained in the original data, effectively reduce the interference of data distribution fluctuations under varying operating conditions of rotating machinery on feature representation, and enhance the robustness of features in judging fault modes. The first disturbance feature is used to construct a classifier, and the second disturbance feature signal is reconstructed to assist in the judgment. It not only relies on the deep discriminativeness of the first disturbance feature to lay an accurate classification foundation for the fault classifier, but also makes up for the information loss problem caused by adversarial learning in traditional domain adaptation methods through signal reconstruction. This makes the judgment data more complete in preserving the details of fault features, further reducing misjudgment and omission, and improving the diagnostic accuracy.

[0057] In other embodiments, inputting deep features into a cross-Gaussian perturbation unit to obtain a first perturbation feature and a second perturbation feature includes:

[0058] Deep features are divided equally by channel and then cross-stitched, and weighted features are generated using Softmax gating;

[0059] The perturbation weights are determined based on the weighted characteristics, and the recalibration characteristics are determined based on the perturbation weights.

[0060] In the above scheme, the deep features are divided equally by channel and then cross-stitched, which helps to integrate the local fault details contained in each channel, such as high-frequency impacts and low-frequency fluctuations in the vibration signal of rotating machinery, and avoids the limitation of feature representation caused by the one-sided information of a single channel. At the same time, the introduction of Softmax gate to generate weighted features can dynamically highlight key features that are strongly correlated with the fault mode and suppress redundant interference information through adaptive weight allocation, reduce the influence of irrelevant signals caused by operating condition fluctuations, and greatly improve the fault discrimination of weighted features.

[0061] This design enhances the generalization ability of features to data distribution fluctuations under varying operating conditions through reasonable perturbation, while avoiding excessive perturbation that could damage core fault information. Meanwhile, the recalibrated features further strengthen the representation intensity of key fault information through the channel-by-channel effect of weighted features and weighted features. Ultimately, the output first and second perturbation features are both comprehensive and accurate, providing high-quality feature support for subsequent fault classifier construction and signal reconstruction. This effectively alleviates the problems of insufficient local feature alignment and weak feature anti-interference ability in existing domain adaptation methods, and improves the fault diagnosis accuracy under cross-operating condition scenarios.

[0062] In other embodiments, determining the perturbation weights based on weighting characteristics includes:

[0063] Calculate the unimodal Gaussian weights for each weighted feature;

[0064] The unimodal Gaussian weights are normalized to obtain the perturbation weights.

[0065] In the above scheme, a single-peak Gaussian weight is calculated for each weighted feature, which can accurately capture the core dimensions that are strongly correlated with the fault mode in the weighted features and greatly improve the focus of the weight on key fault information.

[0066] On the other hand, normalizing the unimodal Gaussian weights can unify the Gaussian weights of different dimensions to the same numerical range. This eliminates the imbalance in weight scale caused by differences in the numerical ranges of the weighted features, and ensures that the perturbation weights of each dimension are comparable and synergistic. This allows subsequent feature recalibration based on these weights to accurately apply to each feature dimension at a balanced scale. It enhances the feature's generalization ability to fluctuations in the distribution of data under varying operating conditions through reasonable perturbation, without compromising the integrity of core fault information. This effectively solves the problems of weak perturbation weight targeting and excessive perturbation of non-fault features in existing domain adaptation methods, and helps improve the accuracy and robustness of fault diagnosis in cross-operating condition scenarios.

[0067] In other embodiments, determining the recalibration features based on the perturbation weights includes:

[0068] The recalibrated features are obtained by multiplying the weighted features one by one channel.

[0069] In the above scheme, the weighted features of each channel can be matched with exclusive perturbation weights, which can achieve a precise correspondence between feature dimensions and weight dimensions, reduce the problem of key channels being weakened and interference channels being over-strengthened, and ensure that the weights' regulation of features is precisely applied to the fault-related dimensions of each channel, so that the fault information of each channel can be strengthened in accordance with its importance.

[0070] On the other hand, the weighted features have initially screened out fault-related information, and the perturbation weights further focus on the core fault dimensions. After the two are multiplied, the representation intensity of key fault information can be amplified, making the signal amplitude of the fault feature frequency band more prominent. At the same time, non-fault interference such as residual operating condition fluctuations and signal noise in the weighted features can be suppressed for a second time, which greatly improves the distinguishability of faults or interference in the recalibrated features.

[0071] Meanwhile, the operation logic of multiplying channel by channel is simple and highly interpretable, which can avoid the destruction of core feature information by complex calculations. It ensures that the recalibrated features have stronger resistance to interference from changing operating conditions while retaining the integrity of the original fault features, effectively alleviating the problems of weak feature recalibration and easy obscuring of core fault information in existing methods.

[0072] In other embodiments, the deep features are divided into equal channels and then cross-stitched, including:

[0073] The deep features are input into the large-scale path and the small-scale path respectively for multi-scale feature extraction to obtain the first output feature and the second output feature. The first output feature and the second output feature are then cross-stitched together.

[0074] In the above scheme, deep features are input into the large-scale path and the small-scale path respectively, which can capture key information in different dimensions of rotating machinery vibration signals. The large-scale path can effectively extract low-frequency, global working condition correlation features and avoid missing the impact of working condition fluctuations on feature distribution due to the limited field of view of the small scale. The small-scale path can accurately capture high-frequency, local fault impact features, solve the problem of easily blurred fault details when extracting large-scale features, and achieve full-dimensional feature coverage.

[0075] On the other hand, cross-splitting the first and second output features can break the channel independence of the two scale features, enabling the working condition adaptation information of the large-scale path and the fault details of the small-scale path to be deeply integrated in the channel dimension. This allows each new channel to simultaneously carry the working condition background and fault signal, avoiding the problem of the separation of working condition and fault information when the subsequent Softmax gating is assigned weights.

[0076] This setting helps reduce perturbation bias caused by isolated scale information and effectively alleviates the problem of insufficient local feature alignment in existing domain adaptation methods.

[0077] In other embodiments, obtaining deep features of the raw data includes:

[0078] Raw vibration signals under multiple operating conditions are collected, and source domain datasets and target domain datasets are formed based on the raw vibration signals.

[0079] In the above scheme, the acquisition of original vibration signals under multiple working conditions accurately matches the dynamic changes in working conditions such as speed and load during the actual operation of rotating machinery. This avoids the problem of incomplete scenario coverage in traditional single-working-condition data acquisition. By covering vibration signals under different working conditions, the scheme can fully capture the changing patterns of fault characteristics in working condition fluctuations, and significantly improve the diagnostic generalization ability of the model in actual variable working-condition scenarios.

[0080] On the other hand, the source domain dataset can provide benchmark data with clearly labeled fault categories, laying the foundation for subsequent fault classifier construction based on the first perturbation feature and the transfer of fault category knowledge; the target domain dataset corresponds to the unknown working condition data to be diagnosed, forming an inter-domain comparison relationship with the source domain data, so that subsequent domain adaptation operations such as maximum mean difference measurement of distribution difference and adaptive batch normalization dynamic adjustment of parameters have a clear target, effectively solving the problem of sharp drop in cross-working condition diagnosis accuracy of traditional diagnostic methods.

[0081] At the same time, the model learns fault discrimination ability and inter-domain adaptability simultaneously during training, ultimately improving the accuracy and reliability of fault diagnosis under varying operating conditions.

[0082] In other embodiments, obtaining deep features from the raw data further includes:

[0083] The original vibration signal was preprocessed using a sliding time window and Z-score normalization to obtain source domain samples and target domain samples.

[0084] In the above scheme, the sliding time window is used to process the original vibration signal, which can divide the continuous temporal vibration signal of rotating machinery into multiple sample segments of equal length. This not only preserves the local temporal features contained in the signal, such as the instantaneous impact signal when the fault occurs and the periodic vibration pattern, but also avoids the problem of the fault details being diluted or truncated due to the direct processing of long signals. Furthermore, it can generate a sufficient number of training samples, solving the problem that the insufficient sample size of the original time series data can easily lead to model overfitting.

[0085] Meanwhile, the equal-length sample format can accurately adapt to the fixed-dimensional input requirements of subsequent multi-layer one-dimensional convolutional networks, laying the data format foundation for the model to stably extract deep features. On the other hand, introducing Z-score standardization to process the segmented samples can effectively eliminate signal amplitude fluctuation interference caused by differences in operating conditions, allowing the standardized samples to focus on the feature changes brought about by the fault, such as signal frequency shifts and the addition of harmonic components caused by the fault, rather than irrelevant variables related to the operating conditions; thus improving the accuracy and robustness of fault diagnosis in cross-operating condition scenarios.

[0086] In other embodiments, obtaining deep features from the raw data further includes:

[0087] Construct a multi-layer one-dimensional convolutional network and embed an adaptive batch normalization layer after each convolutional layer, so that the mean and variance of the batch normalization are updated adaptively with the target domain data.

[0088] This setup helps to accurately extract deep fault features, improves the ability to adapt to differences in data distribution across operating conditions, and provides robust basic feature support for subsequent feature perturbation and classification.

[0089] Understandably, multi-layer one-dimensional convolutional networks are naturally suited to the temporal characteristics of rotating machinery vibration signals. Through progressively deeper convolutional operations, they can gradually remove noise interference from the original vibration signals and extract multi-dimensional features from shallow local impacts to deep global patterns, thus avoiding the limitations of a single convolutional layer in representing complex fault features.

[0090] Meanwhile, traditional batch normalization relies solely on fixed mean and variance of source domain data, making it difficult to adapt to changes in data distribution in the target domain (such as new operating conditions), which can easily lead to feature shifts or even distortion of fault information. Adaptive batch normalization, on the other hand, allows the mean and variance to be dynamically updated with the target domain data, enabling real-time calibration of distribution deviations at each stage of feature extraction, thereby reducing the data distribution shift between the source and target domains.

[0091] For example, when the target domain rotational speed changes from 20Hz to 40Hz, the overall statistical characteristics of the vibration signal change. Adaptive batch normalization can adjust the parameters in real time to ensure that the features output by each convolutional layer retain the core fault information and are aligned with the distribution characteristics of the target domain, thus avoiding the problem of "effective fault features being suppressed and operating condition interference being amplified" caused by distribution differences.

[0092] Meanwhile, the "layer-by-layer embedding" design enables distributed calibration to run through the entire feature extraction process. Both shallow features (such as the basic waveform of a signal) and deep features (such as the abstract pattern of a fault) can be adapted in a targeted manner, avoiding the cumulative deviation caused by insufficient calibration in a single link. The final output deep features have both "fault discrimination" and "cross-domain adaptability", providing high-quality input for the accurate perturbation of the cross-Gaussian perturbation unit and the reliable training of the fault classifier. This improves the diagnostic accuracy and stability of the model in variable operating conditions from the source of feature extraction.

[0093] In one specific embodiment, please refer to Figure 4 and Figure 5 The fault diagnosis method proposed in this application includes the following steps:

[0094] S1: Under multiple operating conditions, the original vibration signals of the internal components of rotating machinery under different health conditions are collected to form source domain datasets and target domain datasets. On a self-built three-phase asynchronous motor fault simulation experimental platform, six operating conditions are designed by adjusting the motor speed (20Hz, 30Hz, 40Hz) and torque load (1A, 2A). For details of the specific operating conditions, please refer to Table 1.

[0095] S2: The original vibration signal is preprocessed using a sliding time window and Z-score normalization to obtain source domain samples and target domain samples;

[0096] S3: Construct a multi-layer one-dimensional convolutional network to extract deep features from the source domain samples and the target domain samples respectively, and embed an adaptive batch normalization layer after each convolutional layer so that the mean and variance are updated adaptively with the target domain data.

[0097] S4: Input the deep features into the cross-Gaussian perturbation unit, output recalibrated features R1 and R2, and merge R1 and R2 in the channel dimension to obtain the perturbed features. The cross-Gaussian perturbation unit data stream includes:

[0098] S4.1: Deep features are input into large-scale and small-scale paths for multi-scale feature extraction, resulting in features U1 and U2. The large-scale path CNN-L has a kernel size of 17×1 and 16 channels; the small-scale path CNN-S has a kernel size of 3×1 and 16 channels.

[0099] S4.2: Divide the output features of CNN-L and CNN-S equally by channel and cross-concatenate them, and use Softmax gating to generate weighted features V1 and V2;

[0100] U1→[U 1a U 1b ]

[0101] U2→[U 2a U 2b ]

[0102] F1 = Concat(U 1a U 2a )

[0103] F2 = Concat(U 1b U 2b )

[0104] G = Softmax(F1⊕F2)

[0105]

[0106] Among them, U 1a U represents the first half of the channel dimension features. 1b This represents the second half of the feature in the channel dimension. Concat represents the concatenation operation in the channel dimension, and ⊕ represents the element-wise addition operation. This indicates an element-wise multiplication operation.

[0107] S4.3: Calculate the unimodal Gaussian weights Θ1(V1) and Θ2(V2) for the weighted features V1 and V2 respectively;

[0108] S4.4: Normalize the single-peak Gaussian weights to obtain the final perturbation weights;

[0109]

[0110] S4.5: Multiply the perturbation weights with the original weighted features V1 and V2 channel by channel to obtain the recalibration features R1 and R2;

[0111]

[0112] S4.6: Calculate the covariance loss of R1 and R2 to constrain the two features to maintain low correlation.

[0113] S5: Measuring source domain perturbation characteristics based on maximum mean difference Perturbation characteristics of the target domain The differences in distribution between them;

[0114] S6: Construct an upsampling-convolutional decoder to perturb the features of the target domain. Signal reconstruction is performed to mitigate information loss caused by adversarial learning;

[0115] S7: Source domain perturbation characteristics A fault classifier consisting of a fully connected layer and a Softmax output layer is used to determine the fault category.

[0116] S8: During the training phase, the source domain and target domain data are simultaneously input into the diagnostic network, and the RMSprop optimizer is used for training.

[0117] S9: In the inference phase, only the target domain data is input, and after forward propagation, the classifier outputs the probability of the fault category.

[0118] Table 1. Variable Operating Condition Task Settings

[0119]

[0120] Table 2. Fault Diagnosis Results

[0121]

[0122]

[0123] Table 2 shows the comparison results between the model proposed in this invention and the existing variable operating condition fault diagnosis model under different variable operating condition tasks in Table 1. By comparing the fault diagnosis accuracy, it can be seen that the present invention improves the accuracy and reliability of fault identification compared with the existing methods, and can effectively and accurately diagnose the fault category of rolling bearings.

[0124] Secondly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor; when the processor runs the computer program, it implements the fault diagnosis method of any of the above embodiments.

[0125] The electronic device proposed in this application improves the accuracy of fault diagnosis because it can implement the fault diagnosis method described in any of the embodiments.

[0126] Thirdly, embodiments of this application provide a storage medium storing a computer program, which, when executed by a processor, implements the fault diagnosis method of any embodiment.

[0127] The storage medium proposed in this application embodiment can implement the fault diagnosis method of any embodiment, thereby improving the accuracy of fault diagnosis.

[0128] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0129] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0130] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

[0131] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A fault diagnosis method based on cross-disturbance domain adaptation, characterized in that, Includes the following steps: Obtain deep features from raw data; The deep features are input into the cross-Gaussian perturbation unit to obtain the first perturbation feature and the second perturbation feature; A fault classifier based on a deep learning network is generated based on the first perturbation feature; The second disturbance feature signal is reconstructed to generate judgment data, and the judgment data is input into the fault classifier to form the fault category probability.

2. The fault diagnosis method according to claim 1, characterized in that, The step of inputting the deep features into the cross-Gaussian perturbation unit to obtain the first perturbation feature and the second perturbation feature includes: Deep features are divided equally by channel and then cross-stitched, and weighted features are generated using Softmax gating; The perturbation weights are determined based on the weighted characteristics, and the recalibration characteristics are determined based on the perturbation weights.

3. The fault diagnosis method according to claim 2, characterized in that, The step of determining the perturbation weights based on the weighted characteristics includes: Calculate the unimodal Gaussian weights for each of the weighted features; The single-peak Gaussian weights are normalized to obtain the perturbation weights.

4. The fault diagnosis method according to claim 3, characterized in that, The determination of recalibration features based on perturbation weights includes: The recalibrated features are obtained by multiplying the weighted features one channel at a time.

5. The fault diagnosis method according to claim 2, characterized in that, The step of dividing deep features equally according to channels and then cross-splitting them includes: The deep features are input into the large-scale path and the small-scale path respectively for multi-scale feature extraction to obtain the first output feature and the second output feature. The first output feature and the second output feature are then cross-stitched together.

6. The fault diagnosis method according to claim 1, characterized in that, The deep features obtained from the raw data include: Raw vibration signals under multiple operating conditions are collected, and source domain datasets and target domain datasets are formed based on the raw vibration signals.

7. A fault diagnosis method according to claim 6, characterized in that, The deep features for obtaining the raw data also include: The original vibration signal is preprocessed using a sliding time window and Z-score normalization to obtain source domain samples and target domain samples.

8. The fault diagnosis method according to claim 1, characterized in that, The deep features for obtaining the raw data also include: Construct a multi-layer one-dimensional convolutional network and embed an adaptive batch normalization layer after each convolutional layer, so that the mean and variance of the batch normalization are updated adaptively with the target domain data.

9. An electronic device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor; when the processor runs the computer program, it implements the fault diagnosis method as described in any one of claims 1 to 8.

10. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the fault diagnosis method as described in any one of claims 1 to 8.

Images