Rotating machinery fault diagnosis method, device, medium and product under multiple variable conditions

Abstract

The application discloses a rotating machine fault diagnosis method, device, medium and product under multiple variable working conditions. In the diagnosis method, firstly, a double-cycle cross-domain fault diagnosis framework based on meta-learning is constructed, then a multi-source continuous working condition model based on reliability constraint is constructed, reliability screening is performed on the multi-source continuous working condition samples, and Dirichlet multi-source mixed generation is adopted to generate enhanced samples and soft labels thereof; then, feature extraction is performed on the enhanced samples to obtain enhanced sample features; the support set enhanced samples are preliminarily classified to generate a support set enhanced sample set, and fast adaptation of the model in the current domain is completed; then, a joint multi-constraint objective function is constructed to realize stable cross-domain fault diagnosis. The application is reasonable in design, greatly reduces the false alarm and missed alarm rates of fault diagnosis, and is high in fault diagnosis reliability.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment fault prediction technology, specifically to a method, device, medium, and product for diagnosing faults in rotating machinery under varying operating conditions. Background Technology

[0002] As industrial equipment develops towards high-end and intelligent directions, rotating machinery, as a core functional component of industrial production systems, is widely used in energy, power, chemical, and manufacturing fields. Its operating status directly affects the safety and production continuity of industrial systems. Analyzing the operating data of rotating machinery to achieve early fault identification and accurate diagnosis is an important part of the industrial intelligent operation and maintenance system.

[0003] In practical industrial applications, rotating machinery typically operates under multiple speeds, loads, and complex conditions for extended periods. These conditions are not independent but rather exhibit continuous changes and transitional characteristics. Changes in operating conditions significantly affect the time-domain, frequency-domain, and statistical distribution of vibration signals, causing the same fault type to show marked differences under different operating conditions, thus increasing the complexity of fault diagnosis. Furthermore, most in-service equipment operates in a healthy state, making fault sample acquisition costly and limited in quantity. The actual collected operational data inevitably contains environmental noise, sensor errors, and operating condition disturbances. Under multiple operating conditions, the quality and stability of data from different conditions vary considerably, placing higher demands on the reliable application of fault diagnosis methods in real-world industrial scenarios.

[0004] In the existing technology, although there are studies on fault diagnosis methods for rotating machinery, there are still some problems, such as: (1) the lack of reliability assessment of multi-condition data samples, which leads to weak anti-interference ability of model feature learning and high false alarm and false negative rates. When dealing with multi-condition rotating machinery fault diagnosis tasks, existing methods mainly focus on alleviating the data shortage problem through data augmentation, multi-source domain adaptation and other techniques to enhance the generalization boundary of the model. However, these strategies often ignore the quality control of source data and lack prior assessment and weighted constraint mechanisms for sample reliability. In the actual industrial environment, the collected vibration data is inevitably affected by environmental noise, transient operating conditions and sensor errors, and some samples have high uncertainty or are on the fault discrimination boundary. Existing methods often assume that all training samples have equal credibility, which leads to unreliable samples being reused and spread to the training set during the augmentation process, thereby introducing noise patterns and non-essential feature errors into the model feature space. This indiscriminate use of low-quality samples weakens the model's robustness to environmental noise and operating condition disturbances, causing the model to be overly sensitive to abnormal signals or fail to identify weak fault features in actual diagnosis, ultimately leading to an increase in false alarm and false negative rates, which seriously affects the engineering practicality of the diagnostic system; (2) It is difficult to adapt to the continuous changes in the operating conditions of rotating machinery, resulting in insufficient stability and reliability of cross-operating condition fault diagnosis. Existing cross-operating condition fault diagnosis methods usually adopt the idea of ​​discretization modeling, treating different operating conditions as several independent steady states, and performing joint training or feature forced alignment based on discrete operating condition data, aiming to extract common features that are not sensitive to changes in operating conditions. However, in actual industrial scenarios, the operating conditions of rotating machinery often exhibit continuous evolution and transition characteristics, and the equipment is often in a transitional state of atypical operating conditions. Due to the lack of explicit modeling of the continuity of operating conditions and transitional states, existing methods are difficult to adapt to the current feature distribution when the actual operating state deviates from the discrete steady state covered by training. Furthermore, some methods, in pursuing cross-domain generalization, excessively emphasize the consistency of feature representations under different operating conditions, neglecting the inherent physical differences of the same fault under different operating conditions. This leads to excessive compression of the feature space and blurred boundaries between fault classes. These shortcomings cause diagnostic results to oscillate when operating conditions fluctuate, and the same fault may lead to contradictory conclusions over consecutive time periods, thereby reducing the stability and reliability of cross-operating-condition fault diagnosis.

[0005] In other words, existing fault diagnosis methods for rotating machinery have the following problems: the lack of reliability assessment of multi-condition data samples leads to weak anti-interference ability of model feature learning and high false alarm and false negative rates; in addition, they are difficult to adapt to continuous changes in the operating conditions of rotating machinery, resulting in insufficient stability and reliability of cross-condition fault diagnosis. Summary of the Invention

[0006] This invention discloses a method, device, medium, and product for fault diagnosis of rotating machinery under varying operating conditions. It solves the technical problems of high false alarm and false negative rates in fault diagnosis of rotating machinery under multiple operating conditions, and the difficulty in characterizing continuous changes in operating conditions, leading to insufficient stability and reliability in cross-condition diagnosis. It features a reasonable design, significantly reduced false alarm and false negative rates, and high reliability in fault diagnosis. The technical solution adopted is as follows: A fault diagnosis method for rotating machinery applicable to various operating conditions, applied to rotating machinery, includes the following steps: S1. Construct a dual-loop cross-domain fault diagnosis framework based on meta-learning. The dual-loop cross-domain fault diagnosis framework is designed such that, in each training round, at least a support domain and a query domain are divided from the source working condition domain set, and a support set and a query set are constructed respectively. The dual-loop cross-domain fault diagnosis framework includes at least an inner loop and an outer loop. In the inner loop, the parameters in the dual-loop cross-domain fault diagnosis framework are updated with a finite number of steps using the support set as input to obtain fast weight parameters. In the outer loop, the cross-domain generalization performance of the model updated by the inner loop is evaluated using the query set as input, and the global parameters are meta-updated. S2. Construct a multi-source continuous operating condition model based on reliability constraints. The multi-source continuous operating condition model is designed to perform reliability screening on multi-source continuous operating condition samples and use Dirichlet multi-source hybrid generation to generate enhanced samples and their soft labels. S3. Call the augmented sample, extract features from the augmented sample, and obtain augmented sample features; based on the augmented sample features, combine the adaptive decision module of general discrimination and instance bias collaboration, perform preliminary classification of support set augmented samples, generate support set augmented sample set, and complete the rapid adaptation of the model in the current domain; S4. Construct a joint multi-constraint objective function to achieve stable cross-domain fault diagnosis. The joint multi-constraint objective function includes a joint generalization task loss function, a gradient consistency alignment loss function, and a semantic centroid convergence loss function. The joint loss value is output based on the evaluation results of the query set in the outer loop to update the global parameters of the model.

[0007] Based on the above technical solution, the support domain and the query domain come from different source operating condition domains. The source operating condition domain is an independent set of operating conditions with different operating conditions and different data distributions, in order to simulate the diagnostic scenario of generalizing from seen operating conditions to unseen operating conditions.

[0008] Based on the above technical solution, the steps of the multi-source continuous operating condition model based on reliability constraints, which involves reliability screening of multi-source continuous operating condition samples and generating enhanced samples and their soft labels using Dirichlet multi-source hybrid generation, include: S21. For any sample Feature extraction was performed to obtain The classifier outputs the number of fault categories. Corresponding prediction probability distribution ; S22. Measuring sample uncertainty through information entropy:

[0009] Calculate the sample reliability weight based on the information entropy: ; S23, When the sample reliability weight When adding samples to the candidate sample pool Complete reliability screening. This is the reliability threshold; S24, from the candidate sample pool Extracted from Samples from different source operating conditions and corresponding tags The mixing coefficient vector of the sampled Dirichlet distribution , These are the parameters of the Dirichlet distribution; S25. Correct the mixing coefficient vector by combining the sample reliability weights, where k represents the index of the kth source operating condition sample, with a value ranging from 1 to K:

[0010] Enhanced samples are generated based on the corrected mixing coefficients:

[0011] Soft labels for enhanced samples are generated based on the corrected mixing coefficients, where y represents the sample label corresponding to the source operating condition domain: .

[0012] Based on the above technical solution, the adaptive decision-making module that combines general discrimination and instance bias includes a general classifier and a bias prediction network. The general classifier has a weight matrix, which is:

[0013] For the set of real numbers, The dimension of the feature vector output by the feature extractor. The number of fault categories; the weight matrix is ​​shared among the various source operating condition domains; For input samples Feature extraction is performed to obtain the corresponding instance features. ; The bias prediction network takes instance features as input and outputs a bias term corresponding to the classifier: ; The deviation term is normalized, and a preset scaling factor is introduced. By applying constraints, adaptive classification weights for instances are constructed based on the normalized bias term and the weight matrix of the general classifier: The instance features are processed based on the instance adaptive classification weights to obtain the sample classification prediction. .

[0014] Based on the above technical solution, in the internal loop, enhanced samples are generated and quickly adapted using a multi-source continuous operating condition model with reliability constraints, and classification is completed through an adaptive decision module that combines general discrimination and instance bias. The inner loop fast adapt function is: , Internal learning rate To support the enhancement of the sample set; In the outer loop, the model is updated using meta-optimization based on the joint multi-constraint objective function. This meta-optimization update uses stochastic gradient descent to update the global parameters.

[0015] in The external learning rate is used to obtain a fault diagnosis model with stable cross-domain generalization ability under unknown operating conditions.

[0016] Based on the above technical solutions, a regularization constraint on the bias modulus is introduced by combining standard cross-entropy loss to construct a domain-based combined classification loss, avoiding the model's over-reliance on instance-specific information during cross-domain meta-optimization:

[0017] For standard cross-entropy loss, This is the regularization coefficient, used to balance classification accuracy and bias constraint strength.

[0018] Based on the above technical solution, the generalized task loss function is:

[0019] For query set; The gradient consistency alignment loss function is:

[0020] For the first The gradient corresponds to each source operating condition domain. The average gradient; The semantic centroid convergence loss function is:

[0021] For the first In the first source operating condition domain Centroid of class sample features The global centroid of this category is defined as follows: The joint multi-constraint objective function is:

[0022] This is the balance coefficient.

[0023] A fault diagnosis device for rotating machinery applicable to various operating conditions, comprising: The data acquisition module is used to collect vibration signals and source operating condition information of rotating machinery and perform preprocessing. The data storage module is used to store raw data, preprocessed data, soft tag data, and source operating condition metadata; The diagnostic module is used to diagnose the fault status of rotating machinery and output the fault category and its corresponding confidence level according to the preset diagnostic report template.

[0024] A storage medium storing a computer program thereon, which, when executed by a processor, enables the implementation of the rotating machinery fault diagnosis method applicable to varying operating conditions as described above.

[0025] A computer program product, wherein when the instructions in the computer program product are executed by a processor of an electronic device, the electronic device performs the rotating machinery fault diagnosis method applicable to varying operating conditions as described above.

[0026] Beneficial effects This invention is a reasonably designed fault diagnosis method for rotating machinery under various operating conditions. It models the cross-domain diagnosis process as a task-level meta-learning problem and achieves stable generalized diagnosis under unknown operating conditions through a multi-source continuous operating condition model with reliability constraints and a general discriminant-instance bias collaborative decision-making mechanism.

[0027] The meta-learning-based dual-loop cross-domain fault diagnosis framework employs a dual-loop optimization structure with an inner and outer loop. In the inner loop, enhanced samples are generated and rapidly adapted using a multi-source continuous operating condition model with reliability constraints. Classification is then completed through an adaptive decision module that coordinates general discriminant and instance bias, while simultaneously eliminating sample noise interference. In the outer loop, the model is meta-optimized and updated based on the joint multi-constraint objective function to obtain a fault diagnosis model with stable cross-domain generalization ability under unknown operating conditions. This overcomes, to some extent, the problems of insufficient continuity modeling of multi-source continuous operating conditions, the susceptibility of unreliable enhanced samples to negative transfer, and the degradation of discriminant ability due to strong cross-domain feature alignment in existing technologies, resulting in diagnostic instability and insufficient generalization. This framework provides a more stable and robust cross-domain diagnostic capability under unknown operating conditions. That is, this application adopts a closed-loop collaborative structure of "sample augmentation - adaptive decision-making - multi-constraint meta-optimization" to achieve high-precision cross-domain diagnosis in complex working conditions. With meta-learning dual loop as the main line, in each round of training, a real application mode simulating "generalization from seen working conditions to unseen working conditions" is constructed through support set / query set tasks. This enables the model to obtain the ability to quickly adapt to unknown working conditions and the ability to generalize stably, thereby improving the accuracy of rotating machinery fault diagnosis. It also shows significant advantages for generalization methods in typical application fields. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only one embodiment of the present invention. For those skilled in the art, other embodiments can be derived from the provided drawings without creative effort.

[0029] Figure 1 Flowchart of the rotating machinery fault diagnosis method applicable to various working conditions in this invention; Figure 2 : Working principle diagram of the rotating machinery fault diagnosis method applicable to various working conditions in this invention; Figure 3 : A schematic diagram of a multi-source continuous operating condition model for reliability constraints of the rotating machinery fault diagnosis method applicable to various operating conditions in this invention; Figure 4 : A schematic diagram of the adaptive decision-making mechanism of general discrimination and instance deviation collaboration in the fault diagnosis method for rotating machinery applicable to various working conditions in this invention; Figure 5 : A visual comparison diagram of the features of the rotating machinery fault diagnosis method applicable to various working conditions in this invention with other comparative methods; Detailed Implementation The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some embodiments may include or substitute parts and features of other embodiments. The scope of the embodiments herein encompasses the entire scope of the claims and all available equivalents thereof. Throughout this document, the terms “first,” “second,” etc., are used only to distinguish one element from another without requiring or implying any actual relationship or order between the elements. Indeed, a first element can also be referred to as a second element, and vice versa. Furthermore, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a structure, apparatus, or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a structure, apparatus, or device. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the structure, apparatus, or device that includes said element. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0030] The terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer" used in this document to indicate orientations or positional relationships are based on the orientations or positional relationships shown in the accompanying drawings. They are used solely for the convenience of describing the document and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In the description herein, unless otherwise specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two elements; they can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0031] In this document, unless otherwise stated, the term "multiple" means two or more.

[0032] In this article, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0033] In this article, the term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0034] like Figure 1 and2 The method for diagnosing faults in rotating machinery under varying operating conditions, as shown, includes the following steps: S1. Construct a dual-loop cross-domain fault diagnosis framework based on meta-learning. The dual-loop cross-domain fault diagnosis framework is designed to divide at least the support domain and the query domain from the source operating condition domain set in each training round, and construct the support set and the query set respectively. In this embodiment, the support domain and the query domain come from different source operating condition domains. The source operating condition domain is an independent set of operating conditions with different operating conditions and different data distributions, so as to simulate the diagnostic scenario of generalizing from seen operating conditions to unseen operating conditions.

[0035] Specifically, the dual-loop cross-domain fault diagnosis framework includes at least an inner loop and an outer loop. In the inner loop, in each training round, a support domain and a query domain are randomly divided from the source domain set, and support sets and query sets are constructed respectively to simulate the actual diagnosis scenario of "generalizing from seen working conditions to unseen working conditions".

[0036] In the inner loop, the parameters in the dual-loop cross-domain fault diagnosis framework are updated in a finite number of steps using the support set as input to obtain the fast weight parameters; In the outer loop, the cross-domain generalization performance of the model after being updated by the inner loop is evaluated using the query set as input, and the global parameters are meta-updated. S2. Construct a multi-source continuous operating condition model based on reliability constraints. The multi-source continuous operating condition model is designed as follows: Figure 3 As shown, reliability screening is performed on multi-source continuous operating condition samples, and Dirichlet multi-source hybrid generation is used to generate enhanced samples and their soft labels, specifically: In production practice, the vibration signals collected in the operating environment of rotating machinery are easily affected by environmental noise and non-stationary operating conditions. Some samples are distributed near the classification decision boundary, resulting in high prediction uncertainty. If low-confidence samples are directly used for interpolation or hybrid enhancement, the noise and fuzzy semantics will be amplified and introduced into the generated virtual enhanced samples, thereby interfering with the model's discrimination boundary learning and reducing cross-domain generalization performance.

[0037] The steps involved in reliability screening of multi-source continuous operating condition samples and generating enhanced samples and their soft tags using Dirichlet multi-source hybrid technology include: S21. For any sample Feature extraction was performed to obtain The classifier outputs the number of fault categories. Corresponding prediction probability distribution ; S22. Measuring sample uncertainty through information entropy:

[0038] The larger the entropy value, the more uncertain the prediction. Further, the sample reliability weight is calculated based on the information entropy. To quantify its contribution to the enhancement process; among which As the normalization factor, it makes , The closer to 1, the higher the sample confidence level.

[0039] S23, When the sample reliability weight When the sample is deemed a high-reliability sample, it is added to the candidate sample pool. Complete reliability screening. The reliability threshold is set at 1; furthermore, low-confidence samples that do not meet the condition are not included in the subsequent augmentation process, thereby suppressing noise propagation and negative migration risk. Compared with traditional data augmentation, which usually only involves linear interpolation between two samples and has the disadvantages of limited sample distribution and difficulty in covering the interior of the convex hull of a high-dimensional distribution space composed of multiple source domains, the samples in this application have good reliability, which is conducive to improving prediction accuracy.

[0040] S24. To construct a multi-source continuous operating condition feature space, this invention employs a Dirichlet distribution to introduce a multi-source mixing strategy from the candidate sample pool. Extracted from Samples from different source operating conditions and corresponding tags The mixing coefficient vector of the sampled Dirichlet distribution , Dirichlet is a distribution parameter; S25. To strengthen the dominant role of high-confidence samples in the mixing process, the mixing coefficient vector is modified in conjunction with the sample reliability weights, where k represents the index of the k-th source operating condition sample, with a value ranging from 1 to K:

[0041] Enhanced samples are generated based on the corrected mixing coefficients:

[0042] Soft labels for enhanced samples are generated based on the corrected mixing coefficients, where y is the sample label corresponding to the source operating condition domain: .

[0043] By employing the aforementioned multi-source hybrid strategy with reliability constraints, we can effectively explore the internal region of the convex hull formed by the distribution of multi-source domain samples while ensuring the semantic reliability of the enhanced samples. This enables us to achieve approximate coverage from discrete source conditions to the potential continuous condition space. At the same time, the information entropy reliability screening and reliability weighting mechanism explicitly suppress the propagation of noise from low-confidence samples, reducing the risk of negative migration caused by unreliable enhanced samples. This provides high-quality data support for subsequent meta-learning to quickly adapt and generalize across domains.

[0044] S3. Invoke the augmented samples, extract features from them, and obtain augmented sample features; based on these features, combine the adaptive decision module that integrates general discrimination and instance bias to perform preliminary classification of the support set augmented samples, generate a support set augmented sample set, and complete the rapid adaptation of the model within the current domain, such as... Figure 4 As shown, specifically: Based on the features of the augmented samples, in the meta-learning inner loop stage, the augmented samples are input into the feature extraction network to obtain discriminative feature representations. The feature extraction network itself and the method for obtaining discriminative feature representations are existing technologies and will not be described in detail here.

[0045] For any input sample x (including the original sample and the enhanced sample), the corresponding instance features are extracted by the feature extractor: ,in , For the set of real numbers, The dimension of the feature vector output by the feature extractor.

[0046] By extracting features from the enhanced samples, the model can cover a wider range of potential continuous operating conditions during the training phase and provide more representative feature representations for subsequent classification decisions.

[0047] The general discriminant structure is based on shared modeling. To characterize the shared fault discrimination semantics across operating conditions, a general classifier is constructed, and its weight matrix is ​​defined as follows:

[0048] in Number of fault categories It is shared between different source operating condition domains to provide a stable cross-domain discrimination benchmark, thereby ensuring cross-domain semantic consistency.

[0049] Instance bias prediction and controlled introduction: Considering that the same fault category may exhibit continuously changing discrimination differences under different operating conditions, a single fixed classifier is prone to causing compression of intra-class discrimination structure and degradation of decision boundary.

[0050] This invention introduces a lightweight bias prediction network. Based on the current instance feature distribution as input, the classifier bias term is dynamically output for any instance feature. Its deviation term is given by the following formula, which realizes fine-grained compensation for the general decision boundary.

[0051]

[0052] To prevent excessive fluctuations in the magnitude of the bias term with different instances from compromising the cross-domain stability of the general classifier, the bias term is normalized, and a preset scaling factor is introduced. Constraints are imposed to control the influence of bias in the overall decision-making process, thereby introducing instance adaptation capability in a controlled manner. Instance adaptive classification weights are constructed based on the normalized bias term and the weight matrix of the general classifier.

[0053] Instance features are processed based on instance adaptive classification weights to obtain sample classification predictions:

[0054] This mechanism allows samples of the same fault category under different operating conditions to be distributed within a class with a reasonable structure around a common semantic center, without being forced to be mapped to completely consistent locations. This preserves the necessary operating condition characteristics while maintaining the separability of fault categories, thereby improving the reliability of discrimination under unknown operating conditions.

[0055] The adaptive decision-making module that combines general discrimination and instance bias also includes a bias modulus regularization term. To prevent the adaptive decision-making module from overfitting due to excessive reliance on instance bias to achieve short-term performance improvements during cross-domain meta-optimization, a bias modulus regularization term is introduced. This bias modulus regularization term constructs a domain-based combined classification loss based on the cross-entropy loss.

[0056] For standard cross-entropy loss, This is the regularization coefficient, used to balance classification accuracy and bias constraint strength.

[0057] This domain-specific classification loss ensures accurate classification while limiting instance bias to a reasonable range, allowing the instance adaptation mechanism to compensate only when necessary. This enables collaborative modeling of domain invariance and domain specificity during the rapid adaptation of the inner loop of meta-learning and the cross-domain optimization of the outer loop, and improves the generalization stability across different working conditions.

[0058] S4. Construct a joint multi-constraint objective function to achieve stable cross-domain fault diagnosis. The joint multi-constraint objective function includes a joint generalization task loss function, a gradient consistency alignment loss function, and a semantic centroid convergence loss function. Based on the evaluation results of the query set in the outer loop, a joint loss value is output to update the model's global parameters. Specifically: Within the meta-learning inner loop, the support set is used as input, the support set augmentation sample set generated in s2 is called, and the domain combination classification loss constructed in s3 is called. The fast adaptation function of the inner loop is: , Internal learning rate To support the enhancement of the sample set; Through this rapid adaptation process, the model can form a preliminary ability to discriminate the current working condition distribution with limited support set sample information. At the same time, the IAC module makes controlled adjustments to the general discrimination boundary through the instance bias compensation mechanism to avoid intra-class structure compression and discrimination degradation caused by simple feature alignment.

[0059] Furthermore, using query sets Updated parameters for the inner ring To evaluate cross-domain generalization performance, a generalization task loss function is constructed, which is as follows:

[0060] For query set; This loss is used to measure the model's ability to classify unseen work condition samples after completing rapid adaptation within the domain, and to guide the model to learn parameter initializations with cross-domain generalization potential from the task level.

[0061] Furthermore, to avoid optimization oscillations and generalization performance degradation caused by inconsistent gradient directions among different source domains during multi-source domain joint training, a gradient consistency alignment loss function is introduced in the outer loop:

[0062] For the first The gradient corresponds to each source operating condition domain. The average gradient; This constraint suppresses conflicts in update directions between source domains from an optimization perspective, enabling the model to gradually converge to update directions that are stable for continuous multi-source conditions, thereby improving the stability of the cross-domain training process.

[0063] To prevent semantic representations of the same fault category from drifting during multi-source domain training, a semantic centroid convergence loss function is introduced in the outer loop:

[0064] For the first In the first source operating condition domain Centroid of class sample features This is the global centroid of the category; The task loss function, gradient consistency alignment loss function, and semantic centroid convergence loss function mentioned above are unified into a joint multi-constraint objective function:

[0065] This is a balancing coefficient used to adjust the intensity of the influence of gradient consistency constraints and semantic structure constraints on the overall optimization process.

[0066] In the outer loop, a meta-optimization update is performed on the model based on the joint multi-constraint objective function. This meta-optimization update uses stochastic gradient descent to update the global parameters.

[0067] in The external learning rate is used to obtain a fault diagnosis model with stable cross-domain generalization ability under unknown operating conditions.

[0068] In this application, within the inner loop, enhanced samples are generated and rapidly adapted using a multi-source continuous operating condition model with reliability constraints. Classification is then completed through an adaptive decision module that combines general discriminant and instance-specific bias. Specifically, within the inner loop, using the support set as input, a sample adaptive enhancement module generates structured enhanced samples under multi-source continuous operating condition conditions, and combines sample reliability constraints to eliminate noise interference. Subsequently, the instance adaptive decision module dynamically generates classifier bias terms based on the feature distribution of the current instance and combines them with shared general classifier weights, transforming the traditional fixed decision boundary into an elastic boundary formed by the superposition of general failure modes and instance-specific biases. This mechanism enables the model to perform preliminary modeling of the continuous discriminant structure formed by potential operating condition changes under limited source domain information when rapidly adapting to the support set, while retaining key domain-specific information during parameter updates.

[0069] In the outer loop, the model parameters updated via the inner loop are used for forward inference and cross-domain generalization performance evaluation of the query set samples. At this point, the instance-adaptive classification structure introduced by the sample adaptive enhancement module participates in the overall optimization of the meta-gradient as part of the cross-domain generalization constraint. By jointly considering the generalization task loss function, gradient alignment constraint function, and semantic centroid convergence loss function in the outer loop loss function, the model is guided at the optimization level to maintain consistency in the update direction under different source domain adaptation paths, thereby stabilizing the cross-domain shared semantic structure and significantly improving the model's cross-domain diagnostic robustness and generalization performance under unknown conditions.

[0070] That is, through the above-mentioned collaborative strategy of "rapid adaptation of inner loop + multi-constraint meta-optimization of outer loop", the model can not only complete the rapid adaptation of known working conditions in the support set stage, but also explicitly constrain the cross-domain discrimination structure, optimization path consistency and category semantic stability in the external meta-optimization process, thereby achieving more stable and robust cross-domain fault diagnosis under unknown working conditions.

[0071] To comprehensively verify the cross-domain generalization ability of the proposed method under different operating conditions, this paper selects two representative publicly available datasets for rotating machinery bearing fault diagnosis for experimental validation: the SDUST dataset and the PU dataset. The two datasets have significant differences in experimental platform structure, operating condition variations, and fault type complexity, which can evaluate the robustness and generalization performance of the model under unknown operating conditions from different perspectives.

[0072] To comprehensively verify the cross-domain generalization ability of the proposed method under different operating conditions, this paper selects two representative publicly available datasets for rotating machinery bearing fault diagnosis for experimental validation: the SDUST dataset (Shandong University of Science and Technology bearing fault dataset) and the PU dataset (University of Paderborn, Germany bearing fault dataset). These two datasets are publicly available existing technology datasets in the field of rotating machinery fault diagnosis, and they differ significantly in experimental platform structure, operating condition variations, and fault type complexity. They can evaluate the robustness and generalization performance of the model under unknown operating conditions from different perspectives, as shown in Tables 1 and 2 below: Table 1. Comparison of average diagnostic accuracy of various methods on the SDUST bearing dataset

[0073] Table 2. Comparison of average diagnostic accuracy of various methods on the PU bearing dataset

[0074] For the SDUST dataset, the IRM method (an existing technique in the field of domain adaptation and generalization in machine learning, widely used in multi-source domain generalization and cross-condition fault diagnosis, and well-known to those skilled in the art, so it will not be elaborated further) suffers from severe loss of key discriminative information when faced with complex non-stationary signals from rotating machinery due to its strong constraint on the causal invariance of feature representations under different conditions. This results in the worst overall performance across all experiments. The CORAL method (a classic domain adaptation method that reduces data distribution shift under different conditions by aligning the covariance matrices of features from the source and target domains, thereby improving cross-domain fault diagnosis performance; an existing technique, so it will not be elaborated further) alleviates the distribution shift problem to some extent compared to IRM by aligning the second-order statistics of features from different source domains. In the single-condition shift scenario of the SDUST dataset, the intra-class structure differences of similar fault samples are small, and the alignment of second-order statistics can better maintain the discriminative structure in the feature space. Therefore, CORAL outperforms MLDG, which focuses on parameter generalization, on this dataset.

[0075] However, in the PU dataset, which features more complex operating conditions, different operating conditions can easily lead to morphological differences in the feature space of similar fault samples. In this scenario, CORAL's diagnostic accuracy decreased by 7.11% compared to its performance on the SDUST dataset. MLDG (a meta-learning-based domain generalization method that constructs meta-training and meta-testing tasks on multiple source domains, enabling the model to learn feature representations that can generalize to unknown operating condition domains) improved the accuracy of the PU dataset to 88.10% by simulating the cross-domain meta-optimization process in meta-learning. MGADA (a multi-gradient adaptive domain adaptation method that improves the generalization ability of cross-operating condition fault diagnosis by adaptively aligning the distribution differences between multiple domains) combined with Mixup augmentation (a common data augmentation method in existing technologies that constructs virtual training samples by linearly interpolating and mixing two sets of samples and their labels to enhance the model's generalization ability and regularization effect, and suppress overfitting) achieved suboptimal results of 91.08% and 88.35% on the SDUST and PU datasets, respectively. However, the results from the sub-tasks show that both MLDG and MGADA exhibit varying degrees of performance fluctuations in tasks P2 and P3 (which, in the PU bearing fault dataset, are strong transfer tasks with greater operating condition differences, more significant feature distribution shifts, and more complex discrimination boundaries) where the operating conditions differ more significantly and the discrimination boundaries are more complex. The method presented in this section achieves the highest average accuracy of 94.41% and 92.00% on the SDUST and PU datasets, respectively, and demonstrates good stability across all operating condition tasks, significantly outperforming benchmark methods such as IRM, CORAL, MLDG, and MGADA.

[0076] Depend on Figure 5 The visualization results show that while the IRM method exhibits some clustering trends in the feature space for different fault categories, the distribution of samples within each category is relatively loose. In particular, samples from different source domains remain significantly separated within the same category, and target domain samples struggle to form unified clusters with source domain samples. This indicates that under strong invariance constraints, it suppresses some condition-related but discriminative features, resulting in weak cross-domain alignment. The MLDG method presents elongated features; although samples of the same category are roughly clustered together, there is still some overlap between different categories, and the inter-class discriminative boundaries are not clear enough. The CORAL method aligns different categories using second-order statistics, making the boundaries relatively clear. However, observing within a single fault category reveals that samples from different domains are still distributed in different regions of the feature cluster, with a significant distance between target and source domain samples. This indicates that it only achieves alignment at the global statistical level and fails to fully characterize the fine-grained changes in intra-class structure under complex operating conditions. The MGADA method, combining meta-learning and Mixup data augmentation strategies, improves intra-class compactness to some extent, exhibiting a more pronounced clustering trend for similar samples compared to IRM, MLDG, and CORAL. However, because the hybrid enhancement process does not explicitly model sample reliability and domain differences, samples from different domains within the same category still exhibit a relatively scattered distribution, with limited inter-domain fusion and slightly blurred internal structures in some categories. In contrast, the proposed method demonstrates the best visualization effect in the feature space. Different fault categories form feature clusters with clear boundaries and significant intervals, reflecting strong inter-class separability. Simultaneously, samples from the same fault category are highly compact in the feature space, and samples from different source and target domains achieve good fusion within clusters without significant inter-domain separation. This indicates that the proposed method can effectively suppress cross-domain interference while retaining key discriminative information in complex operating scenarios, thereby effectively improving the cross-domain fault diagnosis capability under unseen operating conditions.

[0077] A fault diagnosis device for rotating machinery applicable to various operating conditions, comprising: The data acquisition module is used to collect vibration signals and source operating condition information of rotating machinery and perform preprocessing. The data storage module is used to store raw data, preprocessed data, soft tag data, and source operating condition metadata; The diagnostic module is used to diagnose the fault status of rotating machinery and output the fault category and its corresponding confidence level according to the preset diagnostic report template.

[0078] A storage medium storing a computer program thereon, which, when executed by a processor, enables the implementation of the rotating machinery fault diagnosis method applicable to varying operating conditions as described above.

[0079] A computer program product, wherein when the instructions in the computer program product are executed by a processor of an electronic device, the electronic device performs the rotating machinery fault diagnosis method applicable to varying operating conditions as described above.

[0080] The present invention has been described above by way of example, but the present invention is not limited to the specific embodiments described above. Any modifications or variations made based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for diagnosing faults in rotating machinery applicable to various operating conditions, characterized in that, Applied to rotating machinery, including the following steps: S1. Construct a dual-loop cross-domain fault diagnosis framework based on meta-learning. The dual-loop cross-domain fault diagnosis framework is designed such that, in each training round, at least a support domain and a query domain are divided from the source working condition domain set, and a support set and a query set are constructed respectively. The dual-loop cross-domain fault diagnosis framework includes at least an inner loop and an outer loop. In the inner loop, the parameters in the dual-loop cross-domain fault diagnosis framework are updated with a finite number of steps using the support set as input to obtain fast weight parameters. In the outer loop, the cross-domain generalization performance of the model updated by the inner loop is evaluated using the query set as input, and the global parameters are meta-updated. S2. Construct a multi-source continuous operating condition model based on reliability constraints. The multi-source continuous operating condition model is designed to perform reliability screening on multi-source continuous operating condition samples and use Dirichlet multi-source hybrid generation to generate enhanced samples and their soft labels. S3. Call the enhanced sample, extract features from the enhanced sample, and obtain the enhanced sample features; Based on the enhanced sample features, and combined with the adaptive decision module that combines general discrimination and instance bias, the support set enhanced samples are initially classified to generate a support set enhanced sample set, and the model is quickly adapted in the current domain. S4. Construct a joint multi-constraint objective function to achieve stable cross-domain fault diagnosis. The joint multi-constraint objective function includes a joint generalization task loss function, a gradient consistency alignment loss function, and a semantic centroid convergence loss function. The joint loss value is output based on the evaluation results of the query set in the outer loop to update the global parameters of the model.

2. The method for diagnosing rotating machinery faults applicable to varying operating conditions according to claim 1, characterized in that, The support domain and query domain come from different source operating condition domains. The source operating condition domains are independent sets of operating conditions with different operating conditions and different data distributions, in order to simulate the diagnostic scenario of generalizing from seen operating conditions to unseen operating conditions.

3. The method for diagnosing rotating machinery faults applicable to varying operating conditions according to claim 1, characterized in that, The multi-source continuous operating condition model based on reliability constraints includes the following steps: reliability screening of multi-source continuous operating condition samples, and generation of enhanced samples and their soft labels using Dirichlet multi-source hybrid generation. S21. For any sample Feature extraction was performed to obtain The classifier outputs the number of fault categories. Corresponding prediction probability distribution ; S22. Measuring sample uncertainty through information entropy: Calculate the sample reliability weight based on the information entropy: ; S23, When the sample reliability weight When the sample is added to the candidate sample pool, the sample is added to the candidate sample pool. Complete reliability screening. This is the reliability threshold; S24, from the candidate sample pool Extracted from Samples from different source operating conditions and corresponding tags The mixing coefficient vector of the sampled Dirichlet distribution , These are the parameters of the Dirichlet distribution; S25. Correct the mixing coefficient vector by combining the sample reliability weights, where k represents the index of the kth source operating condition sample, with a value ranging from 1 to K: Enhanced samples are generated based on the corrected mixing coefficients: Soft labels for enhanced samples are generated based on the corrected mixing coefficients, where y is the sample label corresponding to the source operating condition domain: 。 4. The method for diagnosing rotating machinery faults applicable to varying operating conditions according to claim 1, characterized in that, The adaptive decision-making module that combines general discrimination and instance bias includes a general classifier and a bias prediction network; The general classifier has a weight matrix, which is: For the set of real numbers, The dimension of the feature vector output by the feature extractor. The number of fault categories; the weight matrix is ​​shared among the various source operating condition domains; For input samples Feature extraction is performed to obtain the corresponding instance features. ; The bias prediction network takes instance features as input and outputs a bias term corresponding to the classifier: ; The deviation term is normalized, and a preset scaling factor is introduced. By applying constraints, adaptive classification weights for instances are constructed based on the normalized bias term and the weight matrix of the general classifier: ; Instance features are processed based on instance adaptive classification weights to obtain sample classification predictions: .

5. The method for diagnosing rotating machinery faults applicable to varying operating conditions according to claim 4, characterized in that, In the inner loop, enhanced samples are generated and quickly adapted using a multi-source continuous operating condition model with reliability constraints, and classification is completed through an adaptive decision module that combines general discrimination with instance bias. The inner loop fast adapt function is: , Internal learning rate To support the enhancement of the sample set; In the outer loop, the model is updated using meta-optimization based on the joint multi-constraint objective function. This meta-optimization update uses stochastic gradient descent to update the global parameters. in External learning rate; To obtain a fault diagnosis model with stable cross-domain generalization ability under unknown operating conditions.

6. The method for diagnosing rotating machinery faults applicable to varying operating conditions according to claim 4, characterized in that, Based on the standard cross-entropy loss, a regularization constraint on the bias modulus is introduced to construct a domain-based combined classification loss, avoiding the model's over-reliance on instance-specific information during cross-domain meta-optimization: For standard cross-entropy loss, This is the regularization coefficient, used to balance classification accuracy and bias constraint strength.

7. The method for diagnosing rotating machinery faults applicable to varying operating conditions according to claim 6, characterized in that, The loss function for the generalization task is: For query set; The gradient consistency alignment loss function is: For the first The gradient corresponds to each source operating condition domain. The average gradient; The semantic centroid convergence loss function is: For the first In the first source operating condition domain Feature centroids of class samples The global centroid of this category across all source domains is: The joint multi-constraint objective function is: This is the balance coefficient.

8. A fault diagnosis device for rotating machinery applicable to various operating conditions, characterized in that, include: The data acquisition module is used to collect vibration signals and source operating condition information of rotating machinery and perform preprocessing. The data storage module is used to store raw data, preprocessed data, soft tag data, and source operating condition metadata; The diagnostic module outputs the fault category and its corresponding confidence level based on the preset diagnostic report template.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it can implement the rotating machinery fault diagnosis method applicable to varying operating conditions as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, When the instructions in the computer program product are executed by the processor of the electronic device, the electronic device performs the rotating machinery fault diagnosis method applicable to varying operating conditions as described in any one of claims 1 to 7.

Images