Rotating machinery fault diagnosis method and system based on gaussian boundary constraint network

Abstract

The application discloses a rotating machinery fault diagnosis method and system based on a Gaussian boundary constraint network, relates to the technical field of fault diagnosis, and obtains vibration signal data of rotating machinery under a single source domain working condition to form a source domain training set with labels; based on the source domain training set, a Gaussian boundary constraint network is trained to obtain a trained Gaussian boundary constraint network; the Gaussian boundary constraint network comprises a feature extractor, a feature decoupling module, a Gaussian boundary constraint module and a known category prediction head, the Gaussian boundary constraint module comprises a Gaussian classifier, and the Gaussian classifier trains the input first features through a negative log-likelihood loss and a confidence-aware boundary loss; vibration signal data of the rotating machinery is obtained, input into the trained Gaussian constraint network for fault diagnosis, and a fault diagnosis result is output. The application can process various label inconsistency scenarios, and realizes accurate identification of shared categories and robust rejection of unknown categories.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, and in particular to a rotating machinery fault diagnosis method and system based on a Gaussian boundary constraint network. BACKGROUND

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute the prior art.

[0003] Rotating machinery is a key component in mechanical systems and is widely used in the core transmission link of modern industry. Its running state is directly related to production efficiency. However, when the equipment is running in a complex, variable and harsh working environment for a long time, it is prone to failure, which seriously interferes with the normal production process and even causes huge economic losses and safety hazards. Therefore, the development of an efficient fault diagnosis system is of great significance to improve system reliability and reduce maintenance costs.

[0004] With the acceleration of industrial informatization, a large amount of data is generated during the operation of mechanical equipment, which provides new development opportunities for data-driven fault diagnosis methods. Among many methods, deep learning has become a research hotspot because it does not need to rely on prior physical models and can automatically learn effective features. At present, various deep network architectures such as convolutional neural network (CNN), recurrent neural network (RNN) and Transformer have been successfully applied to fault diagnosis tasks. However, the effectiveness of these methods is usually based on two key assumptions: one is that there is a sufficient amount of training data, and the other is that there is a consistent data distribution between the training set and the test set. However, in actual applications, rotating machinery often faces changes in operating conditions such as different speeds and loads, which can cause significant distribution differences between vibration signals collected under different working conditions. In addition, it is costly and difficult to label all working condition data, and models trained based on single working condition data often experience a sharp decline in performance when transferred to unknown working conditions, which is known as domain shift.

[0005] To alleviate the influence of domain shift, researchers have proposed methods such as domain adaptation (DA) and domain generalization (DG). The goal of domain adaptation is to align the feature distributions between the labeled source domain and the unlabeled target domain, thereby improving the model's performance on the target domain. According to the relationship between the label spaces of the source and target domains, DA methods are generally divided into closed-set DA (CDA), partial-set DA (PDA), and open-set DA (ODA). Among them, CDA assumes that the source and target domains have the same label set, and there are various methods based on maximum mean discrepancy (MMD) and adversarial training that have been widely applied in the field of fault diagnosis. PDA assumes that the target domain only contains a subset of the source domain's label set, while ODA considers the presence of unknown classes in the target domain that are not present in the source domain. PDA focuses on avoiding aligning target samples to private classes in the source domain, while ODA aims to accurately identify shared class samples and uniformly determine unknown class samples as "unknown classes."

[0006] Although the above methods have achieved certain results in specific tasks, they generally rely on prior knowledge of the relationship between the label spaces of the source and target domains. In actual industrial scenarios, the label space of the target domain is often unknown, making it difficult to choose an appropriate adaptation strategy. In addition, current research has paid less attention to the more challenging open-partial-set DA (OPDA) problem. To address such problems, the concept of universal domain adaptation (UniDA) has been proposed. "Universal" reflects the possibility of the target domain containing both shared classes and private unknown classes. UniDA can handle various DA scenarios such as CDA, PDA, ODA, and OPDA, and its goal is to accurately identify shared class samples while effectively determining unknown classes. However, achieving efficient UniDA under the condition of complete unknown target domain still faces significant challenges.

[0007] In contrast, domain generalization (DG) aims to train a model using multiple source domain data only to achieve good generalization performance in unseen target domains. To further cope with more complex scenarios where the class space is completely uncertain, universal domain generalization (UniDG) is proposed. Currently, UniDG has taken shape in the field of computer vision, but its research in mechanical fault diagnosis is still relatively scarce. To fill this gap, the prior art proposes an adaptive intermediate class distribution alignment (AICDA) method that aligns the features of source and target domains using an intermediate distribution. The prior art also proposes a domain-private suppression meta-recognition network (DPSMR) that suppresses domain-specific noise through channel-level transferability estimation and strengthens domain-invariant feature learning; it also introduces a meta-recognition calibration mechanism to alleviate the overconfidence of the model when dealing with new classes, and combines a double-consistency loss to improve the stability of feature expression. However, both AICDA and DPSMR rely on multiple source domains or synthetic target domains, which makes them difficult to deploy in a single source domain data context. SUMMARY

[0008] To overcome the above-mentioned deficiencies of the prior art, the present application provides a rotating machinery fault diagnosis method and system based on a Gaussian boundary constraint network, proposes a single-source universal domain generalization method based on a Gaussian boundary constraint network, which can handle multiple label inconsistency scenarios and achieve accurate identification of shared classes and robust rejection of unknown classes.

[0009] To achieve the above-mentioned purpose, one or more embodiments of the present application provide the following technical solutions:

[0010] In a first aspect, the present application provides a rotating machinery fault diagnosis method based on a Gaussian boundary constraint network, comprising:

[0011] Obtaining vibration signal data of a rotating machinery under a single source domain working condition to form a labeled source domain training set;

[0012] Training a Gaussian boundary constraint network based on the source domain training set to obtain a trained Gaussian boundary constraint network; the Gaussian boundary constraint network includes a feature extractor, a feature decoupling module, a Gaussian boundary constraint module, and a known class prediction head, the Gaussian boundary constraint module includes a Gaussian classifier, and the Gaussian classifier trains the input first feature through a negative log-likelihood loss and a confidence-aware boundary loss;

[0013] Obtaining vibration signal data of a rotating machinery, inputting it into the trained Gaussian constraint network for fault diagnosis, and outputting a fault diagnosis result.

[0014] In a further technical solution, the feature decoupling module receives the global features output by the feature extractor, and uses a gating mechanism to decompose the global features into a first gating feature and a second gating feature. After being mapped by a fully connected layer, the first feature and the second feature are obtained respectively. The first feature is input into the Gaussian boundary constraint module, and the second feature is input into the known category prediction head.

[0015] A further technical solution involves using covariance loss as a constraint between the first and second features during the training of the Gaussian boundary constraint network.

[0016] A further technical solution is that the negative log-likelihood loss encourages the first feature to follow a Gaussian distribution that satisfies the constraints, expressed as:

[0017]

[0018] in, Indicates the first One source domain sample, Representation of features The One dimension, Indicates the first The features of the source domain sample of the first One dimension, Expression characteristics The true label, Indicates belonging to a category The The mean of a Gaussian distribution Indicates belonging to a category The The variance of a Gaussian distribution.

[0019] A further technical solution is that the distribution range and location of the first feature constrained by the confidence-aware boundary loss are expressed as follows:

[0020]

[0021]

[0022]

[0023]

[0024] in, Indicates the built-in confidence boundary threshold. Indicates the first Normalized Mahalanobis distance for each sample Indicates the outer confidence boundary threshold. Indicates the first The feature vector of each sample a mean vector representing a class to which a first sample belongs, a mean vector representing a class to which a first sample belongs, a mean vector representing a class to which a first sample belongs, a standard deviation vector representing a class to which a first sample belongs, a standard deviation vector representing a class to which a first sample belongs, a standard deviation vector representing a class to which a first sample belongs, a cumulative distribution function representing a standard normal distribution, a confidence coefficient representing an inner boundary, a confidence coefficient representing an outer boundary.

[0025] In a further aspect, the total loss of the Gaussian boundary constraint network training is:

[0026]

[0027] wherein, , , a loss function weight coefficient, a training loss of the Gaussian classifier, a binary cross-entropy, a covariance loss.

[0028] In a further aspect, the failure diagnosis process of the trained Gaussian constraint network is:

[0029] a known class prediction head performs preliminary classification on the vibration signal data of the rotating machinery to obtain a candidate known class;

[0030] a Gaussian mean vector and a standard deviation vector corresponding to the candidate known class are obtained based on the candidate known class;

[0031] a normalized distance is calculated using the first feature of the vibration signal data and the Gaussian mean vector and the standard deviation vector of the candidate known class;

[0032] the normalized distance is compared with a confidence threshold, if the normalized distance is greater than the confidence threshold, it is determined that the vibration signal data is an unknown class, otherwise, the predicted label output by the known class prediction head is taken as the final classification result.

[0033] In a second aspect, the present application provides a rotating machinery failure diagnosis system based on a Gaussian boundary constraint network, comprising:

[0034] a training set construction module configured to: acquire vibration signal data of the rotating machinery under a single source domain working condition to form a labeled source domain training set;

[0035] a model training module configured to train the Gaussian boundary constraint network based on a source domain training set to obtain a trained Gaussian boundary constraint network, the Gaussian boundary constraint network comprising a feature extractor, a feature decoupling module, a Gaussian boundary constraint module, and a known class prediction head, the Gaussian boundary constraint module comprising a Gaussian classifier trained on an input first feature by a negative log-likelihood loss and a confidence-aware boundary loss;

[0036] a model prediction module configured to obtain vibration signal data of the rotating machinery, input the vibration signal data into the trained Gaussian boundary constraint network for fault diagnosis, and output a fault diagnosis result.

[0037] In a third aspect, the present application provides a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the rotating machinery fault diagnosis method based on the Gaussian boundary constraint network according to the first aspect.

[0038] In a fourth aspect, the present application provides a computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the rotating machinery fault diagnosis method based on the Gaussian boundary constraint network according to the first aspect when executing the program.

[0039] The above one or more technical solutions have the following beneficial effects:

[0040] The present application realizes unified and robust diagnostic capability for universal domain generalization (UDG) scenarios, innovatively applies the Gaussian boundary constraint network to single source domain fault diagnosis, and effectively solves the generalization problem in various complex scenarios such as closed set (CDG), partial set (PDG), open set (ODG), and open-partial set (OPDG) in a unified framework through the unique structure design and loss function of the Gaussian boundary constraint network. The model does not need to know the scene type of the target domain in advance, and can automatically adapt and make reliable decisions, improving the practicality and universality.

[0041] The present application completes the training based on only the source domain labeled data without relying on the prior of target domain data and labels, and provides a unified "known / unknown" judgment process in the inference stage, which is convenient for reproduction and deployment in actual monitoring systems.

[0042] The application jointly adopts negative log-likelihood (NLL) and confidence-aware boundary (CAB) loss in a Gaussian boundary-constrained network (GBCN), jointly constrains to reduce overconfidence and misjudgment of unknown samples in a target domain, suppresses feature over-clustering and keeps cautious to out-of-bound samples, reduces the risk of misjudging unknown categories as known categories from the mechanism, and enables the Gaussian boundary-constrained network to have high-confidence rejection ability when both shared and unknown categories exist.

[0043] The application improves feature separation effect by gating feature decoupling and covariance-based statistical independent regularization, and solves the problem of insufficient class separability caused by deep feature semantic coupling.

[0044] The application shapes feature distribution under the condition of no target prior by means of Gaussian boundary constraint, unifies and is compatible with closed-set domain generalization (CDA), partial domain generalization (PDA), open-set domain adaptation (ODA) and open-partial set domain generalization (OPDA) and the like, realizes accurate identification of shared categories and robust rejection of unknown categories, and solves the generalization bottleneck when only single source is available and multi-source alignment cannot be used. BRIEF DESCRIPTION OF DRAWINGS

[0045] The drawings accompanying the specification of the application form part of the application and serve to provide further understanding of the application, the illustrative embodiments of the application and the description thereof serve to explain the application and do not constitute an improper limitation of the application.

[0046] Figure 1 is a structural diagram of the Gaussian boundary-constrained network of the embodiment of the application;

[0047] Figure 2 is a confidence-aware boundary loss schematic diagram of the Gaussian boundary-constrained module of the embodiment of the application;

[0048] Figure 3 is a graph of the influence of ACA and H-score in the experiment of the embodiment of the application;

[0049] Figure 4 is a graph of the average accuracy of CDG, PDG, ODG and OPDG tasks under different confidence intervals in the embodiment of the application.​ DETAILED DESCRIPTION

[0050] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0051] It is also important to note that the terms used herein are not intended to limit the exemplary embodiments to the specific embodiments which are described herein. Rather, it is contemplated that the exemplary embodiments are intended to encompass all embodiments which are within the four corners of the appended claims. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising", when used in this specification, specify the presence of stated features, steps, operations, devices, components and / or combinations thereof, but do not preclude the presence or addition of one or more other features, steps, operations, devices, components and / or combinations thereof.

[0052] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0053] Embodiment One

[0054] As shown in the following table, the embodiment discloses a rotating machinery fault diagnosis method based on Gaussian boundary constraint network, which comprises the following steps: Figure 1

[0055] S1: Obtain the vibration signal data of the rotating machinery under a single source domain working condition to form a labeled source domain training set;

[0056] In the rotating machinery fault diagnosis, the Universal Domain Generalization (UDG) aims to learn a model that can be generalized to unseen domains in the case of both distribution shift and label mismatch. Under the setting of Single-source Universal Domain Generalization (SUDG), the model is trained based on a single labeled source domain .

[0057] ​The present application faces the actual industrial scene of rotating machinery fault diagnosis, and the source domain dataset and the target domain sample are both derived from mechanical vibration signals. Specifically, the source domain dataset can include labeled vibration signal data collected under a certain working condition (for example, rotating speed 900 rpm, load 0 hp, room temperature environment), which can cover typical categories such as healthy state, outer ring fault, inner ring fault, rolling body fault and mixed fault. These data can be derived from a test bench, an industrial production line or a public rotating machinery fault diagnosis dataset (such as CWRU, PU, etc.), to ensure representativeness and verifiability. The target domain sample corresponds to the unlabeled vibration signal collected under the working condition different from the source domain (such as rotating speed 1500 rpm, load 1 hp, temperature 30℃), which can contain the signal of the existing categories in the source domain, and can also appear unknown fault categories that do not exist in the source domain. Through such design, the cross-domain generalization ability of the present application under complex working conditions can be embodied, and the actual needs of unknown category recognition in industrial field can also be met.

[0058] Model is expected to generalize to an unseen target domain , whose data comes from a different distribution than the source domain , i.e. , and has . The target domain can contain both shared classes with the source domain and its private classes. Let and be the label sets of the source and target domains, respectively, and the shared class set be defined as ; and and represent the source and target private class sets, respectively. To quantify the overlap between them, the Jaccard similarity coefficient (J) is introduced:

[0059]

[0060] The lower the value, the greater the label difference, and the more difficult the model generalizes to the target domain.

[0061] UDG involves generalization under different class transfer scenarios, including closed-set domain generalization (CDG), partial-set domain generalization (PDG), open-set domain generalization (ODG) and open-partial-set domain generalization. ; Partial-set domain generalization (PDG): ; Open-set domain generalization (ODG): Open-Partial Domain Generalization (OPDG): , and .

[0062] The goal of SUDG is to train a model using a source domain such that it satisfies: (1) during inference, each sample from the target domain is either correctly classified into a known class of the source domain label space or identified as an unknown class; (2) the model is robust to both domain distribution shift and label space shift without using any target domain data during training.

[0063] The vibration signal data of the rotating machinery under a single source domain working condition is collected to form a labeled source domain training set, wherein each sample is labeled with a fault class label. The original vibration signal is preprocessed, including noise reduction, standardization, and segmentation processing.

[0064] S2: based on the source domain training set, the Gaussian boundary-constrained network is trained to obtain a trained Gaussian boundary-constrained network;

[0065] In this embodiment, the goal of the Gaussian boundary-constrained network (Gaussian Boundary-Constrained Network, GBCN) is to realize feature decoupling and optimize the decision boundary under the condition of unknown classes, so that the model can resist both domain migration and label distribution difference when deployed. As shown in Figure 1 The Gaussian boundary-constrained network includes a feature extractor, a feature decoupling module, a Gaussian boundary-constrained module, and a known class prediction head. The feature decoupling module separates the global features extracted by the feature extractor into two branches, one of which inputs the Gaussian boundary-constrained module for modeling the Gaussian distribution and imposing a confidence boundary constraint, and the other of which inputs the known class prediction head for regular classification of known classes. Covariance loss is applied between the two branches to constrain, reduce redundancy, and ensure that the "Gaussian features" and "supplementary features" are independent of each other. In this way, the discriminability of the traditional classification head for known classes is retained, and the robust rejection of unknown classes is realized by the Gaussian boundary-constrained module.

[0066] The Gaussian boundary-constrained module is trained with a joint loss, specifically: composed of a negative log-likelihood (Negative Log-Likelihood, NLL) loss and a confidence-aware boundary (Confidence-Aware Boundary, CAB) loss. NLL aligns the features and class-conditional Gaussian distribution; CAB limits the features within the confidence boundary region to avoid excessive aggregation to the class center and damage the generalization ability. Through the dual constraints of "distribution alignment + geometric boundary", the Gaussian classifier is more robust to domain migration and label space migration.

[0067] (I) Feature decoupling module

[0068] The feature decoupling module is composed of a trainable gating mechanism and two independent branches. The feature decoupling module adopts a trainable gating mechanism, and the feature extractor adopts a convolutional neural network for feature extraction. The feature vector output by the convolutional neural network (CNN) is decomposed into two branches, i.e., global features. Each branch is modulated by a Sigmoid-activated gate, enabling the network to map features to different subspaces.

[0069] First, the global feature vector output by the feature extractor is input into a fully connected layer (in this embodiment, the number of nodes is 2, and the activation function is Sigmoid) to form a gating unit, obtaining two sets of gating coefficients, i.e., first gate and second gate . Subsequently, the first gate feature and the second gate feature are obtained by element-wise multiplication of and , respectively. That is, the global feature is scaled by two independent gates, i.e., first gate and second gate , to obtain the first gate feature and the second gate feature . Then, , are further mapped by a fully connected layer (in this embodiment, the number of nodes is 256, and the activation function is ReLU) to obtain the first feature and the second feature . Finally, covariance loss is introduced to and to constrain the statistical independence of the two branches, reduce redundancy, and enhance diversity. To reduce redundancy between branches and improve representation diversity, covariance loss is used between and

[0070] :

[0071]

[0072] wherein denotes the covariance loss, denotes the Frobenius norm (i.e., the square root of the sum of the squares of each element of the matrix, used to measure the size of the matrix), denotes the covariance matrix (i.e., the covariance of the two sets of features in each dimension, used to characterize the statistical correlation between features).

[0073] ​​The network adopts a multi-branch binary classification strategy based on Sigmoid, each output independently predicts whether the corresponding class exists, thereby obtaining a decision boundary with more uncertainty perception. The classification loss is defined as binary cross entropy:

[0074]

[0075] wherein, represents binary cross entropy, represents the true label of the th source domain sample, represents the predicted probability of the th source domain sample.

[0076] Through the above technical solution, the feature decomposition is performed according to the trainable gating coefficient. This method is equivalent to obtaining two different features from the same input feature , so that the network can automatically learn a branch biased towards Gaussian distribution constraint ( ) and a branch biased towards known class prediction ( ).

[0077] The first branch mainly faces the Gaussian boundary constraint module, which is used to extract features aligned with the Gaussian distribution. The negative log likelihood (NLL) and confidence boundary (CAB) loss are used to constrain it to make it more consistent with the Gaussian probability space, so as to realize accurate modeling of known classes and robust rejection of unknown classes. The second branch mainly faces the known class prediction head, which is used to retain supplementary information closer to traditional discriminative features. The cross-entropy loss is used to maintain the classification ability of known classes. The difference between the two branches is that the first branch emphasizes probability distribution and boundary constraint, and the feature is more biased towards “Gaussianizable” semantic components; the second branch emphasizes discriminativeness and supplementary information, and the feature is more biased towards regular classification discrimination. By introducing covariance loss to constrain the two branches, it encourages them to be statistically independent, reduces redundancy, and forms a complement.

[0078] (II) Gaussian boundary constraint module

[0079] The first feature is the input of the Gaussian classifier (Gaussian boundary constraint module), and the Gaussian classifier includes the class mean vector and variance vector of each class ( is the feature dimension).

[0080] The training target includes negative log likelihood loss NLL and confidence-aware boundary loss CAB.

[0081] (1) Negative log likelihood loss

[0082] The encouraging feature is subject to a Gaussian distribution that satisfies the constraint condition:

[0083]

[0084] wherein, denotes the th source domain sample, denotes the th dimension of the feature, denotes the th dimension of the feature of the th source domain sample, denotes the th dimension of the feature , and denotes the th Gaussian distribution mean value belonging to the class , denotes the th Gaussian distribution variance belonging to the class .

[0085] (2) Confidence-aware boundary loss

[0086] The distribution range and position of the feature are constrained:

[0087]

[0088]

[0089]

[0090]

[0091] wherein, denotes the inner confidence boundary threshold, which is used to limit the minimum range of the sample distance from the class center; denotes the normalized Mahalanobis distance of the th sample, i.e., the distance between the sample feature and its class center; denotes the outer confidence boundary threshold, which is used to limit the maximum range of the sample distance from the class center; denotes the feature vector of the th sample; denotes the mean vector of the class to which the th sample belongs; denotes the standard deviation vector of the class to which the th sample belongs; CDF, cumulative distribution function, representing the standard normal distribution, confidence coefficient representing the inner boundary, used to calculate ; confidence coefficient representing the outer boundary, used to calculate .

[0092] threshold value determined by confidence (such as 85%, 90%, 95% confidence interval), scaling coefficient used to tighten the inner boundary, construct the confidence interval. restrict the features within the annular zone around the class center defined by the confidence: as shown in Figure 2 , points located within the inner boundary are considered overconfident and are penalized; points that cross the outer boundary are considered uncertain samples and are also penalized. This constraint helps to maintain a semantically reasonable, confidence-controllable feature space, avoiding over-clustering or divergence, and promoting robust recognition and rejection of unknown classes.

[0093] The confidence-aware boundary loss determines the double confidence boundary through Gaussian distribution inversion, simultaneously incorporates intra-class convergence and inter-class rejection into the constraint, thereby significantly enhancing the stability of known class discrimination and the ability to reject unknown classes without relying on artificial thresholds, which is fundamentally different from existing single threshold or cross-entropy loss.

[0094] The training loss of the Gaussian classifier G is:

[0095] .

[0096] The total loss of the Gaussian boundary constraint network training is:

[0097]

[0098]

[0099] wherein, , , is the normalized loss function weight coefficient, ,

[0100] , , is the learnable loss weight coefficient.

[0101] is input into the Gaussian boundary constraint module, which trains through negative log likelihood (NLL) and confidence boundary (CAB) loss, and the gradient of the probability modeling and geometric boundary constraint is passed to Make its characteristics gradually consistent with the statistical properties of Gaussian distribution, so as to realize accurate modeling of known classes and robust rejection of unknown classes.

[0102] (Three) known category prediction head

[0103] The input known category prediction head maintains the discriminability of known classes through the conventional classification loss. The two branches complement each other, on the one hand, to ensure Align the distribution in the Gaussian probability space and improve the rejection ability of unknown classes, and on the other hand, to ensure Retain discriminative features to maintain the classification accuracy of known classes.

[0104] S3: Obtain the vibration signal data of the rotating machinery, input it to the trained Gaussian constraint network for fault diagnosis, and output the fault diagnosis result.

[0105] During prediction, the model aggregates the outputs of the known category classification network and the Gaussian classifier for multi-branch decision. For a given input, the feature extractor generates two routes of gated features, one for the uncertainty-aware Gaussian score and the other for standard classification prediction.

[0106] The Gaussian classifier determines whether the target sample belongs to the known class by calculating the standardized distance between the target sample and the mean of each class. In the prediction stage, first, the known category prediction head classifies the target domain sample, outputs the prediction probability of each class , and selects the class with the highest probability As the candidate known class of the sample. According to this candidate class, the corresponding Gaussian mean vector And the standard deviation vector Can be directly obtained. Then, the standardized distance Is calculated using the first feature of the sample .

[0107] For class , define:

[0108]

[0109] Where, is the feature of the target domain sample.

[0110] Compare this standardized distance with the confidence threshold set in the training stage , if the minimum Gaussian distance of the target domain sample is greater than the confidence threshold , it is determined that the sample is an unknown class, if It is less than or equal to If the sample is determined to be of the known class, the class prediction label output by the known class prediction head is directly taken as the final classification result, denoted as:

[0111]

[0112] wherein, denotes the prediction probability of

[0113] In summary, to overcome the challenges of the prior art, the present application faces the single-source universal domain generalization (SUniDG) task and proposes a Gaussian boundary constraint network (GBCN) to effectively identify rotating machinery faults in unknown target domains. The model consists of two core modules: a feature decoupling module and a Gaussian boundary constraint module. The feature decoupling module uses a learnable gating mechanism to decompose global features into independent semantic representations for use by the known class prediction head and the Gaussian boundary module. The feature decoupling module achieves statistical decorrelation and effective decoupling between features under the guidance of covariance loss. The Gaussian boundary constraint module further optimizes the structure of the feature space by introducing a confidence-aware boundary mechanism. The intra-class distribution is constructed within the defined confidence boundary range, and the joint optimization of the negative log-likelihood loss (NLL) and the confidence-aware boundary loss (CAB) is used to introduce probability geometric constraints, effectively suppressing intra-class over-clustering and improving the robustness of the model to unknown class interference.

[0114] The following describes the experiments.

[0115] The experiments were performed on a computer equipped with an NVIDIA RTX 4090D Ti GPU. Keras was used for deep learning tasks, and TensorFlow served as the Keras backend. The Adam optimizer was used for optimization, with an initial learning rate of , which was reduced by 10% every 100 epochs. The batch size was set to 512. The number of training epochs was 1500. Early stopping was used for model training, with 10% of the source domain dataset used as the validation dataset. Classification accuracy and H-score were used as evaluation metrics for performance comparison. For known fault classes, the new fault type was classified as the Unknown class, indicating unknown category.

[0116] (1) Comparison methods

[0117] To evaluate the effectiveness of the proposed framework, GBCN is compared with the state-of-the-art methods.

[0118] Adaptive intermediate class-wise distribution alignment (AICDA) (Q. Qian, J. Luo, Y. Qin, Adaptive intermediate class-wise distribution alignment: a universal domain adaptation and generalization method for machine fault diagnosis, IEEE transactions on neural networks and learning systems, 2024): AICDA introduces a class alignment mechanism that adaptively selects intermediate domain representations to minimize class mismatch between source and target domains.

[0119] Domain-private-suppress meta-recognition network (DPSMR) (M. Xu, Y. Zhang, B. Lu, Z. Liu, Q. Sun, A novel domain-private-suppress meta-recognition network based universal domain generalization for machinery fault diagnosis, Knowledge-Based Systems 309 (2025) 112775): DPSMR adopts a meta-learning strategy to suppress domain-private features while enhancing the discriminative ability of domain-shared representations.

[0120] GBCN without covariance loss (GBCN-NC): This variant removes the covariance loss component to separate and evaluate its contribution to overall performance.

[0121] GBCN without confidence-aware boundary loss (GBCN-NCAB): This ablation variant removes the confidence-aware boundary loss (CAB), which aims to enhance class boundary clarity and reduce misclassification of unknown target samples.

[0122] (2) Evaluation methods

[0123] The present application uses two indicators to evaluate the effectiveness of various methods: average class accuracy and H-score.

[0124] Average Class Accuracy (ACA): To ensure fair evaluation in the presence of class imbalance, the present application calculates the average class accuracy of all target domain classes (including unknown classes). The definition is as follows:

[0125]

[0126] where, is the classification accuracy of class , represents the set of source domain classes, represents the set of target domain known classes, represents the set of target domain unknown classes in ODG and OPDG. H-score: To better reflect the trade-off between known class and unknown class recognition in ODG and OPDG, the present application also uses H-score, defined as the harmonic mean between the average accuracy of known classes (OS) and the accuracy of unknown classes (UK):

[0127]

[0128] where,

[0129] is the average accuracy of known classes, is the accuracy of unknown classes. (3) CWRU dataset

[0130] The present application uses the Case Western Reserve University (CWRU) bearing dataset to evaluate the performance of the proposed method under different domain conditions. This dataset contains ten health states, including one normal state and nine fault types based on fault location (inner ring, ball, outer ring) and severity (0.007 inches, 0.014 inches, 0.021 inches). As shown in Table 1, all fault types are labeled from 0 to 9.

[0131] Table 1 Label of CWRU dataset

[0132]

[0133]

[0134] ​Each signal was acquired at a sampling rate of 48 kHz. The signal was segmented into overlapping windows, each containing 4096 samples with a stride of 256. The segmented samples were saved into four domain-specific subsets: D0 to D3, each representing a different running condition: 1796 rpm, 1775 rpm, 1755 rpm, and 1732 rpm. For each domain, 400 samples were randomly selected from each class to ensure class balance. These domains simulated distribution shifts and were used as source or target domains in various transfer tasks.

[0135] Table 2 shows the task settings for SUDG on the CWRU dataset. A total of 14 tasks (C1-C14) were constructed, divided into four scenarios: CDG, PDG, ODG, and OPDG. Each task consists of a source-target domain pair ( The source and target domains are defined, along with their corresponding label sets and Jaccard similarities. The CDG task (C1-C3) assumes the source and target domains share the exact same label set, with a Jaccard similarity of 1.00. The PDG task (C4-C6) considers partial label overlap, where the target domain contains only a subset of the source domain labels. Jaccard similarities range from 0.40 to 0.67, reflecting varying degrees of local alignment between the source and target domains. The ODG task (C7-C10) introduces unobserved, unknown categories from the source domain into the target domain. This setting tests the model's ability to identify both seen and unseen categories simultaneously. The OPDG task (C11-C14) combines partial and open set conditions and is the most challenging. The target domain includes unseen and partially overlapping categories, with Jaccard similarities ranging from 0.20 to 0.71.

[0136] Table 2 shows the different UDG task settings and Jaccard similarity on the CWRU dataset.

[0137]

[0138] (4) Impact on model accuracy

[0139] In order to study The impact on model performance, this invention in the confidence region With the percentage fixed at 95%, the following was evaluated. Accuracy values ​​ranging from 0.5 to 0.9. Different The value leads to a significant change in classification performance. Figure 3 It can be seen that, =0.9 achieved the highest average accuracy and H-score.

[0140] (5) Confidence region Impact on model accuracy

[0141] To investigate the sensitivity of GBCN under different confidence thresholds The performance of GBCN is compared in the four UDG scenarios shown in Table 2. These comparisons are made in the confidence regions ranging from 45% to 95%, and in the experiments, .

[0142] As can be seen from Figure 4 the PDG tasks achieve 100% accuracy even at lower thresholds, indicating that they are relatively less affected by domain shift and unknown class interference. In contrast, the ODG and OPDG scenarios show more pronounced improvement as the confidence region increases, indicating that the boundary is crucial for rejecting unknown samples.

[0143] ACA steadily increases with increasing confidence threshold and peaks at 95%. In the rest of the experiments of the present invention, = 95% and = 0.9 are used.

[0144] (6) Performance analysis on CWRU dataset

[0145] Tables 3 and 4 show the performance comparison of 14 UDG tasks on the CWRU dataset. The evaluation includes four domain generalization scenarios: CDG, PDG, ODG, and OPDG. The present invention compares GBCN with its ablation variants (GBCN-NC and GBCN-NCAB) and two state-of-the-art baseline methods AICDA and DPSMR.

[0146] In terms of ACA (Table 3), GBCN achieves the highest overall performance in all scenarios, with an average improvement of 5%-49% over the best-performing baseline. GBCN achieves 100% accuracy on 8 out of 14 tasks, demonstrating strong generalization ability even in the presence of domain shift and label inconsistency.

[0147] Table 3 Average class accuracy on CWRU dataset (tasks C1-C14) under UDG settings

[0148]

[0149] For H-score (Table 4), which jointly evaluates the performance of known and unknown classes, GBCN leads with an average accuracy of 96.93%, significantly outperforming all baseline methods. The ablation variant GBCN-NCAB performs poorly on open-domain tasks, with H-score dropping to 0% in ODG and OPDG tasks. This highlights the crucial role of confidence-aware boundary loss (CAB) in handling unknown classes.

[0150] The results demonstrate that GBCN provides robust domain generalization and effective unknown class rejection capability, especially in ODG and OPDG scenarios.

[0151] Table 4 H-score on 12k-CWRU dataset (tasks C7-C14) in UDG setting

[0152]

[0153] Embodiment Two

[0154] The embodiment discloses a rotating machinery fault diagnosis system based on a Gaussian boundary constraint network, and comprises:

[0155] A training set construction module configured to: acquire vibration signal data of the rotating machinery under a single source domain working condition, and form a labeled source domain training set.

[0156] A model training module configured to: train the Gaussian boundary constraint network based on the source domain training set, to obtain a trained Gaussian boundary constraint network; the Gaussian boundary constraint network comprises a feature extractor, a feature decoupling module, a Gaussian boundary constraint module, and a known class prediction head, the Gaussian boundary constraint module comprises a Gaussian classifier, and the Gaussian classifier trains the input first feature through a negative log-likelihood loss and a confidence-aware boundary loss.

[0157] A model prediction module configured to: acquire vibration signal data of the rotating machinery, input the vibration signal data into the trained Gaussian constraint network for fault diagnosis, and output a fault diagnosis result.

[0158] Embodiment Three

[0159] The embodiment aims to provide a computing device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of embodiment one.

[0160] Embodiment Four

[0161] The embodiment aims to provide a computer-readable storage medium, a computer-readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to perform the steps of the method of embodiment one.

[0162] The steps and methods involved in the apparatuses of Embodiments 3 and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0163] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0164] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0165] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of rotating machinery based on Gaussian boundary constraint networks, characterized in that, include: Vibration signal data of rotating machinery under single source domain conditions are acquired to form a labeled source domain training set; Based on the source domain training set, the Gaussian boundary constraint network is trained to obtain the trained Gaussian boundary constraint network. The Gaussian boundary constraint network includes a feature extractor, a feature decoupling module, a Gaussian boundary constraint module, and a known class prediction head. The Gaussian boundary constraint module includes a Gaussian classifier, which is trained on the first input feature through negative log-likelihood loss and confidence-aware boundary loss. The vibration signal data of rotating machinery is acquired, input into a trained Gaussian constraint network for fault diagnosis, and the fault diagnosis results are output. The feature decoupling module receives the global features output by the feature extractor, and uses a gating mechanism to decompose the global features into a first gating feature and a second gating feature. After being mapped by a fully connected layer, the first feature and the second feature are obtained respectively. The first feature is input into the Gaussian boundary constraint module, and the second feature is input into the known category prediction head. The negative log-likelihood loss encourages the first feature to follow a Gaussian distribution that satisfies the constraints, expressed as: in, This represents the negative log-likelihood loss. Indicates the first One source domain sample, Representation of features The One dimension, Indicates the first The features of the source domain sample of the first One dimension, Expression characteristics The true label, Indicates belonging to a category The The mean of a Gaussian distribution Indicates belonging to a category The The variance of a Gaussian distribution; The distribution range and location of the first feature constrained by the confidence-aware boundary loss are expressed as follows: in, This represents the confidence-perceived boundary loss. Indicates the built-in confidence boundary threshold. Indicates the first Normalized Mahalanobis distance for each sample Indicates the outer confidence boundary threshold. Indicates the first The feature vector of each sample Indicates the first The category to which each sample belongs The mean vector, Indicates the first The category to which each sample belongs The standard deviation vector, The cumulative distribution function represents the standard normal distribution. The confidence coefficient represents the inner boundary. This represents the confidence coefficient for the outer boundary.

2. The rotating machinery fault diagnosis method based on Gaussian boundary constraint network as described in claim 1, characterized in that, When training a Gaussian boundary-constrained network, covariance loss is used to constrain the first and second features.

3. The rotating machinery fault diagnosis method based on Gaussian boundary constraint network as described in claim 1, characterized in that, The total loss during training the Gaussian boundary constraint network is: in, , , Represents the weight coefficients of the loss function. This represents the training loss of the Gaussian classifier. Represents the binary cross-entropy. This represents the covariance loss.

4. The rotating machinery fault diagnosis method based on Gaussian boundary constraint network as described in claim 1, characterized in that, The fault diagnosis process of the trained Gaussian constrained network is as follows: The vibration signal data of rotating machinery is initially classified using a known category prediction head to obtain candidate known categories; Based on the candidate known categories, obtain their corresponding Gaussian mean vector and standard deviation vector; The standardized distance is calculated using the first feature of the vibration signal data and the Gaussian mean vector and standard deviation vector of the candidate known categories. The standardized distance is compared with the confidence threshold. If the standardized distance is greater than the confidence threshold, the vibration signal data is determined to be of an unknown category. Otherwise, the predicted label output by the known category prediction head is used as the final classification result.

5. A rotating machinery fault diagnosis system based on Gaussian boundary constraint networks, employing the rotating machinery fault diagnosis method based on Gaussian boundary constraint networks as described in any one of claims 1-4, characterized in that, include: The training set construction module is configured to acquire vibration signal data of rotating machinery under a single source domain condition and form a labeled source domain training set. The model training module is configured to: train the Gaussian boundary constraint network based on the source domain training set to obtain a trained Gaussian boundary constraint network; the Gaussian boundary constraint network includes a feature extractor, a feature decoupling module, a Gaussian boundary constraint module, and a known class prediction head; the Gaussian boundary constraint module includes a Gaussian classifier, which is trained on the first input feature using negative log-likelihood loss and confidence-aware boundary loss; The model prediction module is configured to: acquire vibration signal data of rotating machinery, input it into a trained Gaussian constraint network for fault diagnosis, and output fault diagnosis results.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the rotating machinery fault diagnosis method based on Gaussian boundary constraint network as described in any one of claims 1-4.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the rotating machinery fault diagnosis method based on Gaussian boundary constraint network as described in any one of claims 1-4.

Images