A method for diagnosing single-source open-set topology stability faults in rotating machinery.

Abstract

This invention belongs to the field of intelligent fault diagnosis technology for rotating machinery. Addressing the challenge of coordinating cross-condition distribution drift and open-set fault identification under single-source training, it constructs a hierarchical, progressive, and collaborative optimization framework consisting of latent condition view generation, intra-class Laplace consistency learning, and structure-aware open-set inference. This invention elevates cross-condition invariance from feature coordinate alignment to Laplace operator-level topological alignment. Based on a stable topological template obtained through training, it achieves operator-level open-set inference. Requires only single-source condition training data, significantly improving the model's classification accuracy for known faults and its rejection accuracy for unknown faults under unknown conditions. It exhibits excellent robustness and engineering practicality, and can be widely applied to fault diagnosis of various rotating machinery such as bearings, gearboxes, fans, and pumps.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology for rotating machinery, specifically to a method for topology stability fault diagnosis of rotating machinery applicable to single-source training, cross-condition distributed drift, and open-set fault identification scenarios. Background Technology

[0002] Rotating machinery is a core piece of equipment in industrial production and other fields. The operating status of its key components directly determines the safety and reliability of the system, and fault diagnosis technology is the core of equipment health management. Deep learning-based fault diagnosis methods have been widely used, but they still face many core challenges in real-world industrial scenarios: models trained under single-source conditions struggle to cope with the distributional drift of unknown conditions and have poor cross-condition generalization ability; during the testing phase, they are prone to misjudging unknown faults, and traditional closed-set diagnostic methods cannot effectively distinguish between known and unknown faults; in single-source training scenarios, coordinate-level feature alignment suffers from ill-posedness, and existing open-set diagnostic methods do not fully exploit the topological structure of the fault manifold, resulting in insufficient robustness.

[0003] Existing technologies suffer from significant technical shortcomings: most domain generalization methods rely on multi-source training data, and methods in single-source scenarios fail to exploit topological invariance; open-set diagnostic methods are susceptible to operating condition drift and are not optimized in conjunction with cross-operating condition generalization capabilities; some topology learning methods are detached from the physical characteristics of rotating machinery operating condition changes; and the lack of an integrated collaborative optimization framework limits the overall model performance. In summary, existing technologies cannot address the collaborative challenges of cross-operating condition distribution drift and open-set fault identification under single-source training, making it difficult to meet the practical application needs of industrial sites. There is an urgent need to develop an integrated fault diagnosis method that simultaneously improves the model's cross-operating condition generalization and unknown fault rejection capabilities. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a single-source open-set topology stability fault diagnosis method for rotating machinery. By constructing a hierarchical, progressive, and collaborative optimization framework of "latent working condition view generation - intra-class Laplace consistency learning - structure-aware open-set inference," it simulates physically reliable cross-working-condition disturbances, stabilizes the inherent manifold topology of fault categories, and achieves operator-level open-set fault identification. This effectively solves the collaborative technical challenges of cross-working-condition distribution drift and open-set fault identification under single-source training, improves the model's classification accuracy for known faults and its rejection accuracy for unknown faults under unknown working conditions, and meets the application needs of practical industrial scenarios.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: This invention constructs a three-stage collaborative framework of "Low-Condition View Generation (LCVG) - Intra-Class Laplace Consistency Learning (IL-SCL) - Structure-Aware Open Set Inference (SA-OSI)," the specific technical solution of which is as follows: (I) Mechanism for Generating Physically Reliable Potential Working Condition Views This mechanism combines the physical characteristics of the changing operating conditions of rotating machinery to construct a set of physical basis perturbation operators, thereby achieving accurate simulation of potential operating condition changes. In the absence of target domain data, it transforms single-source operating condition training data into multi-view potential operating condition perturbation data, providing a structured training foundation for subsequent topological stability learning.

[0006] (II) Intraclass Laplace Structural Consistency Learning Mechanism This mechanism elevates cross-condition invariance from the feature coordinate level to the intrinsic manifold topology level, constructing intra-class Laplace consistency constraints. By building a class-adaptive weighted adjacency matrix for each potential condition view feature of each fault category, the normalized graph Laplace operator for each view is calculated to encode the intrinsic manifold topology of the fault category. Then, the average Laplace operator of the category is used as a stable topology template. The Laplace structure consistency loss aligns the Laplace operators of each view to the average template, achieving cross-condition stability of the intra-class manifold topology. Simultaneously, the classification loss is fused to ensure inter-class separability and avoid embedding space degradation and collapse.

[0007] (III) Operator-level open set reasoning mechanism with structure awareness This mechanism is based on a stable intra-class Laplacian topological template to achieve operator-level open-set fault reasoning. A trained class-averaged Laplacian operator is used as the stable manifold template for that class. The class Laplacian operator is reconstructed using a substitution method for test samples. The deviation between the reconstructed Laplacian operator and the stable template is calculated to measure the operator-level perturbation of the known class topology by the test samples. Then, an open-set threshold is determined based on the statistical characteristics of the topological deviation after training. Finally, known fault classification or unknown fault rejection is performed based on the comparison between the deviation and the threshold.

[0008] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages compared with the prior art: (1) Strong generalization ability across working conditions: Based on physical characteristics, latent working condition disturbances are constructed, and the cross-working condition invariance is improved to the topological level. The classification accuracy can reach more than 96%, which is at least 12% higher than the existing mainstream methods.

[0009] (2) High robustness of feature representation: Stable faulty manifold topology is learned through intraclass Laplace structure consistency learning. Ablation experiments show that the accuracy drops by more than 6% after removing the IL-SCL module.

[0010] (3) High accuracy of unknown fault rejection: The open set reasoning based on the Laplace operator bias metric can achieve a rejection rate of over 93%, which is at least 19% higher than the existing open set methods.

[0011] (4) Fast convergence of collaborative optimization: The three-layer module collaborative optimization and the unified objective function avoid gradient conflicts. The complete method achieves an accuracy of over 96% on multiple datasets.

[0012] (5) High engineering practicality: Only single-source working condition training data is required, and it can be adapted to vibration, current and other signals. It can be deployed on rotating mechanical equipment such as bearings, gearboxes, fans and pumps.

[0013] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the overall implementation process of the fault diagnosis method described in this invention. Figure 2 This is a schematic diagram of the structure and hyperparameters of the shared parameter convolutional feature extraction network used in this invention; Figure 3 This is a comparison chart of the ROC curves for open set recognition of the method of this invention and existing mainstream methods. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] It should be noted that the terms "vertical," "horizontal," "up," "down," "left," "right," and similar expressions used in this article are for illustrative purposes only and do not represent the only possible implementation.

[0017] 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 invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1As shown, the fault diagnosis model of this invention adopts an end-to-end collaborative optimization training method. The overall training process is as follows: sample standardization → generation of potential working condition view → shared feature extraction → intra-class Laplacian topological stability learning → joint loss backpropagation → saving topological template and open set threshold; the diagnosis process is as follows: test sample standardization and feature extraction → reconstructing the Laplacian operator and calculating topological deviation → comparing the maximum deviation with the threshold → outputting the diagnosis result. The specific implementation steps are as follows: (a) Step 1: Single-mode signal standardization processing For single-mode monitoring signals of rotating machinery (taking vibration signals as an example, unit) (Time series signal), source domain data is collected under a single operating condition (fixed speed, fixed load); the single source signal is independently subjected to min-max normalization processing, mapping all feature values ​​to the [0,1] interval. The normalization formula is: in, The original signal value, , These are the minimum and maximum values ​​of a single-source signal, respectively. This is the standardized signal value.

[0020] Implementation details: Standardization processing is performed on the entire single-source signal, preserving the signal's inherent fault characteristics; the processed signal is organized into a tensor with dimensions [number of samples, 1, sequence length], and the sequence length is set to... Preferred This provides a standardized input basis for subsequent convolutional feature extraction.

[0021] Objective: To eliminate dimensional differences in signal amplitude and baseline offset interference, ensure the stability of feature learning, and avoid topology learning bias caused by differences in signal amplitude.

[0022] Technical benefits: Improves the stability of model training and provides input data of a uniform scale for subsequent potential working condition view generation and feature extraction.

[0023] (ii) Step 2: Generation of physically reliable potential working condition view Four types of physical basis perturbation operators are constructed using the LCVG module to perturb the standardized single-source signal, generating four latent operating condition views, corresponding to amplitude scaling, frequency domain scaling, time-frequency distortion, and combined perturbation views, respectively. Each perturbation parameter is randomly sampled from a probability distribution consistent with the actual operating conditions of rotating machinery, ensuring the physical reliability and diversity of the perturbations. The specific implementation of each operator is as follows: Amplitude scaling operator: simulates the amplitude modulation of vibration signals caused by load changes, the formula is: Among them, the amplitude scaling factor Random sampling from a truncated uniform distribution corresponds to changes in operating conditions with decreasing and increasing load, preserving the fault frequency structure; Frequency domain scaling operator: simulates the linear shift of fault characteristic frequencies caused by changes in rotational speed, first... The spectrum is obtained by performing a Fourier transform. Then perform frequency domain scaling. Finally, an inverse Fourier transform is performed to obtain the time-domain signal. ,in The frequency domain scaling factor is used to randomly sample from a truncated uniform distribution, corresponding to the operating conditions of decreasing and increasing rotational speed, while maintaining harmonic relationships. Time-frequency distortion operator: simulates local time-frequency distortion caused by changes in structural stiffness, and the formula is: ,in It is Gaussian white noise. The noise figure is randomly sampled from a uniform distribution, and the noise is spectrally constrained (only the frequency band near the fault characteristic frequency is retained) to preserve the periodicity of the fault impulse; Combined perturbation operator: Integrating amplitude scaling and frequency domain scaling, time-frequency distortion and amplitude scaling to construct a combined operator. , It simulates the changes in operating conditions caused by the coupling of multiple factors.

[0024] Implementation details: The parameters of all disturbance operators are set according to the actual operating conditions of the rotating machinery, preferably with a load change of ±30% and a speed change of ±20%, to ensure the physical reliability of the disturbance; the dimensions of the generated multi-view signal are consistent with the original standardized signal, which is [number of samples, 1, sequence length].

[0025] Objective: In the absence of target domain data, transform single-source training data into structured multi-view potential working condition disturbance data to simulate potential cross-working condition changes and provide a foundation for subsequent intra-class topological stability learning.

[0026] Technical effect: It realizes the structured augmentation of single-source data, and the simulated potential working condition changes are highly consistent with the actual industrial scenario, effectively solving the problem of insufficient samples for cross-working condition generalization under single-source training.

[0027] (III) Step 3: Construction of Multi-View Shared Feature Extraction Network like Figure 2 As shown, a convolutional feature extraction network with shared parameters is constructed. This is an end-to-end deep convolutional neural network, employing a combination of four convolutional layers, pooling layers, and adaptive max-pooling layers. The specific structure and hyperparameters are as follows: Convolutional layer 1: 1 input channel, 32 output channels, kernel size 15, stride 1, padding='same', followed by batch normalization (BN) and ReLU activation function; Convolutional layer 2: 32 input channels, 64 output channels, kernel size 3, stride 1, padding='same', followed by batch normalization (BN) and ReLU activation function; Convolutional layer 3: 64 input channels, 128 output channels, kernel size 3, stride 1, padding='same', followed by batch normalization (BN) and ReLU activation function; Convolutional layer 4: 128 input channels, 256 output channels, kernel size 3, stride 1, padding='same', followed by batch normalization (BN) and ReLU activation function; Pooling layer: Max pooling is used with a kernel size of 2 and a stride of 2, which is combined with convolutional layers to compress the feature space; then adaptive max pooling (AMP) is used with an output dimension of 256 to adaptively compress the spatial dimension and retain key local fault features.

[0028] The signals from four potential working condition views are input into this convolutional feature extraction network. All views share the network parameters, realizing a unified feature mapping of multi-view signals and outputting a high-dimensional feature representation corresponding to each view, with a feature dimension of 256.

[0029] Simultaneously, a classifier is constructed, consisting of three fully connected layers: 512→256, 256→128, and 128→C (where C is the number of known fault categories). The last layer is connected to a Softmax activation function, which outputs the probability that each sample belongs to a different known fault category.

[0030] Implementation details: All learnable parameters of the convolutional feature extraction network and classifier are initialized using Kaiming to ensure the initial stability of model training.

[0031] Objective: To establish a unified embedding space for multi-view features, extract fault features from each potential working condition view, and provide a high-dimensional feature foundation for intra-class Laplace structure consistency learning.

[0032] Technical effect: It realizes unified feature mapping of multi-view signals, ensures the comparability of features of different potential working condition views, and the extracted high-dimensional features can effectively characterize the essential features of the fault.

[0033] (iv) Step 4: Intraclass Laplace structural consistency constraint The core function of the IL-SCL module is to achieve cross-view stability of the fault category topology through intra-class Laplace structure consistency learning. Specific operations include: Feature segmentation and adjacency matrix construction: The 256-dimensional high-dimensional features of the four views are segmented according to the fault category, and for each category... Each view Feature set ( For category (Number of samples), construct class-adaptive weighted adjacency matrix The elements are: in, For class adaptive scaling parameters.

[0034] Normalized graph Laplacian operator computation: Calculate the degree matrix corresponding to the adjacency matrix. (Diagonal matrix, diagonal elements are) (the sum of rows), and then calculate the normalized graph Laplace operator: in, It is the identity matrix. The view is encoded Next category The intrinsic manifold topology.

[0035] Construction of the average Laplacian operator: for each category Calculate the average operator of the four Laplacian operators for each view, and use it as the stable topological template for this class: ,in, Number of views for potential operating conditions.

[0036] Joint Loss Function Construction and Optimization: Constructing the Laplace structural consistency loss to minimize the Frobenius bias between the Laplace operator for each view and the average template: Simultaneously, cross-entropy classification loss is incorporated to ensure inter-class separability: ,in, For the sample The true label, For the classifier to classify samples Category Predicted probabilities; Construct joint loss function: ,in, and The weighting coefficients balance intra-class topological stability with inter-class semantic separation.

[0037] Implementation details: The Laplacian operator is computed within each training batch to ensure computational efficiency; the Frobenius bias is computed without eigenvalue decomposition, reducing computational complexity.

[0038] Objective: To achieve cross-view stability of the inherent manifold topology of fault categories, elevate cross-condition invariance from the feature coordinate level to the operator level, and overcome the ill-posed problem of coordinate-level feature alignment in single-source scenarios; at the same time, to ensure inter-class separability and avoid feature embedding space degradation.

[0039] Technical effectiveness and irreplaceability: Intra-class topology effectively stabilizes the essential characteristics of fault categories; if this module is skipped, subsequent open-set inference will fail due to the lack of a stable topological template and will be unable to cope with cross-condition distribution drift.

[0040] (v) Step 5: Save the stable topology template with known fault categories After the model training converges, perform the following operations: Extract the average Laplace operator for each known fault category As a stable manifold topological template for this category, it is saved to the model library to provide operator-level decision criteria for subsequent open set inference; For each known fault category Each diving condition view Based directly on the feature set of all training samples in this view Following the consistent graphing rules and computation methods used during the training phase, construct the view-level Laplacian operator corresponding to the current view. Calculate the Frobenius deviation between the view-level Laplacian operator and the corresponding class-stable topological template, as the topological deviation of the view: ; Iterate through all known fault categories and all potential operating condition views, collect the topology deviations corresponding to all views, and form a global topology deviation set. Calculate the expected value of this set of deviations. with standard deviation The open set recognition threshold is determined by using a statistical method that superimposes the mean and a fixed multiple of the standard deviation. , ,in By pre-setting fixed coefficients, a statistically robust and engineering-adaptable criterion is provided for open-set inference in scenarios with limited samples.

[0041] Implementation details: View-level topology bias is calculated directly based on the view features and corresponding Laplacian operators generated after training, without the need for repeated feature extraction and model iteration. It can be completed in a single traversal with no additional training overhead. The open set threshold is automatically determined by the statistical features of the global topology bias set, without relying on manual parameter tuning or quantile estimation. It still has stable and reliable judgment capabilities in scenarios with a small number of samples.

[0042] Objective: To provide a category-level stable topology template for subsequent open set inference, and an open set judgment threshold determined based on the inherent fluctuation range of the fault topology, thereby achieving open set fault identification through operator-level topology matching.

[0043] Technical effects: The stable topology template can accurately depict the inherent manifold topology of each known fault category, unaffected by fluctuations in the operating condition view; the threshold obtained based on global topology deviation statistics can objectively characterize the topology fluctuation boundary of known faults, and still has strong robustness and statistical reliability under small sample conditions, providing a stable, accurate and reproducible basis for judgment in open set inference.

[0044] (vi) Step 6: Structure-aware operator-level open set fault reasoning The SA-OSI module is used to diagnose faults in test samples, achieving integrated classification of known faults and rejection of unknown faults. This inference process is based on the aforementioned construction of stable topology templates for known fault categories, calculation of view-level topology deviation, and determination of global thresholds. It ensures consistent topology operator dimensions and rigorous decision logic throughout the process, adapting to real-world application scenarios without view augmentation during the testing phase. Specific operations include: 1. Preparation before reasoning (closed loop with the training phase) Retrieve stable topology templates corresponding to all known fault categories saved to the model library during the training phase, denoted as set. ,in Indicates different known fault categories, For category The average Laplacian operator (i.e., the class-stable topological template) has a dimension of ( (This refers to the total number of samples in this category during the training phase, which should be consistent with the mapping rules for the test samples during the subsequent inference phase). Retrieve the open set recognition topology decision threshold determined during the training phase. This threshold is formed by summing the view-level topology deviations corresponding to all known fault categories and all potential operating conditions into a global deviation set, and is calculated using "mean + preset fixed coefficient". Calculated using the method of "multiple standard deviations" ( (A preset fixed coefficient, preferably 2 or 3).

[0045] 2. Sample preprocessing and construction of the new Laplacian operator The original feature data of the fault samples to be identified is collected, and the original feature data is preprocessed in the same way as the training stage to obtain standardized test feature data, ensuring the consistency between the features of the test samples and the features of the training samples, and avoiding identification bias caused by feature differences. For each known fault category Retrieve the stable topology template corresponding to this category. (Constructed from training samples of this category), a stable topology template is randomly selected. A sample feature in the corresponding training sample set is replaced with the preprocessed feature of the test sample to obtain a new sample set containing the feature of the test sample. Using the same graph construction rules as in the training phase (such as constructing a weighted adjacency matrix based on Euclidean distance), a new Laplacian operator is calculated on the aforementioned new sample set, denoted as... ; Iterate through all known fault categories, repeating the above steps. For each known fault category, perform the "sample replacement - new Laplace operator calculation" operation to obtain the set of new Laplace operators corresponding to all known fault categories. .

[0046] 3. Calculation of topological differences between the test sample and each known fault category For each known fault category The new Laplace operator corresponding to this category With this category of stable topology templates The comparison is used to characterize the degree of influence of the test sample on its inherent topology after it is incorporated into the known fault category. The topological difference between the two is calculated using the Frobenius norm, thus obtaining the relationship between the sample under test and the known fault category. Topological difference The calculation formula is as follows: ,in, The Frobenius norm is used to accurately characterize the degree of difference between the nascent Laplacian operator and the category-stable topological template (i.e., two topological structures). The larger the difference, the greater the destruction of the inherent topological structure of the sample after its features are incorporated into the known fault category, that is, the greater the difference between the sample and the known fault category. Traverse all known fault categories Repeat the above steps to obtain the topological differences between the sample to be tested and all known fault categories, denoted as the difference set. .

[0047] 4. Open set fault identification and judgment For the set of differences Perform statistical analysis and extract the maximum difference in the set, denoted as . The maximum difference represents the maximum difference between the topology of the sample under test and all known fault categories, that is, the maximum degree of disruption to the inherent topology of the sample under test after it is incorporated into each known fault category. Maximum difference With the preset topology determination threshold The following comparison is performed, and the following judgment rules are applied: like This indicates that after the test sample is integrated into all known fault categories, it does not exceed the normal fluctuation range of the topology of that category. The test sample matches the known fault category, and the test sample is determined to be a known fault. Then, based on the output probability distribution of the classifier, the known fault category is determined. like This indicates that after the sample to be tested is incorporated into one (or more) known fault categories, the disruption to its inherent topology exceeds the normal fluctuation boundary. The sample to be tested does not match any of the known fault categories, and is therefore determined to be an unknown open set fault, that is, it does not belong to any trained known fault category.

[0048] 5. Output and Feedback of Reasoning Results Output the identification result of the sample to be tested (known fault / unknown open set fault). If it is a known fault, output its corresponding fault category simultaneously. If it is an unknown open set fault, it can trigger the anomaly feedback mechanism to prompt staff to conduct further fault investigation and sample replenishment.

[0049] 6. Key Explanations All operations in the inference phase (feature preprocessing, graph construction rules, Laplacian operator calculation, and difference calculation method) are completely consistent with those in the training phase, ensuring the accuracy and repeatability of the judgment results and avoiding misjudgments caused by differences in operations. The judgment process uses the "maximum difference" as the core criterion, and accurately judges the matching ability by measuring the degree of damage to the topology of the sample to be tested after it is integrated into a known fault category.

[0050] (vii) Step 7: Model Unified Optimization Convergence This step integrates intra-class topological stability and inter-class semantic separation, constructing a joint loss function to achieve global optimization convergence of the model. Specific operations include: Optimizer and training parameter settings: The Adam optimizer is preferred for the joint loss function. To minimize the problem, the optimizer's learning rate is set to 0.001, batch size to 32, and maximum number of training epochs to 200. Training and validation set partitioning: The raw data of the single-source working condition is randomly divided into a training set and a validation set in a ratio of 8:2. The training set is used for updating model parameters, and the validation set is used to monitor the model convergence status. Early stopping mechanism setting: When the accuracy of known fault classification on the validation set no longer improves for 10 consecutive rounds, training is stopped early to avoid model overfitting and obtain the final fault diagnosis model; Parameter synchronization update: Through backpropagation, all learnable parameters of the shared feature extraction network and classifier are updated synchronously to achieve collaborative optimization of each module.

[0051] During each training cycle, the LCVG module generates multi-view perturbation data, the shared feature extraction network extracts multi-view features, the IL-SCL module performs topology consistency learning, and all parameters are updated synchronously based on backpropagation of the joint loss function; the validation set monitors the classification accuracy, and an early stopping mechanism is triggered if the accuracy does not improve for several consecutive rounds.

[0052] Core objective: To achieve collaborative optimization of intra-class Laplace structure consistency learning and inter-class separation, ensuring that the model has both a stable intra-class topology and good inter-class separation under single-source training, thereby improving the model's cross-condition generalization ability and open set recognition ability.

[0053] VI. Overall Collaborative Implementation Logic The fault diagnosis model of this invention adopts an end-to-end collaborative optimization training method, with each step forming a close dependency and collaborative optimization mechanism. The overall implementation logic of each stage is as follows: Training phase: Sample standardization → LCVG generation of multiple views → Shared feature extraction → IL-SCL topological stabilization → Joint loss backpropagation → Saving template and threshold; Diagnostic phase: Standardize test samples and extract shared features → Reconstruct the Laplacian operator using SA-OSI and calculate the test sample bias → Compare the maximum bias with the threshold and output the result.

[0054] The inventiveness of this invention lies in constructing a closed-loop optimization mechanism for single-source to multi-view conversion, coordinate-level to operator-level conversion, and feature matching to topology matching, specifically embodied as follows: LCVG transforms single-source training data into structured multi-view data based on physical properties, providing a foundation for topologically stable learning; IL-SCL elevates cross-condition invariance from the feature coordinate level to the operator-level topology level, solving the ill-posed problem of coordinate-level alignment in single-source scenarios; SA-OSI achieves operator-level open set inference based on stable topological templates, transforming the criterion for open set recognition from feature matching to topological matching. This forms a closed-loop system: single-source dimensionality expansion → feature extraction → topological stability → operator-level open set inference, a closed-loop mechanism that does not exist in existing technologies.

[0055] Existing methods only achieve single-source sample amplification, coordinate-level feature alignment, or distance-based open set identification, but they do not incorporate the physical characteristics of rotating machinery to construct a perturbation model, nor do they perform synergistic optimization of cross-condition generalization and open set identification based on topology. Compared with existing methods, this method has significant inventiveness and novelty.

[0056] VII. Experimental Verification and Effect Comparison To verify the effectiveness, universality, and robustness of the bearing fault diagnosis method of this invention, two sets of internationally recognized bearing benchmark datasets were selected for comparative and ablation experiments. Through quantitative indicators and qualitative analysis of ROC curves, the technical advantages, inventiveness, and engineering application value of this invention were fully demonstrated.

[0057] (I) Experimental Setup 1. Experimental Hardware and Software Environment The experimental hardware uses a mainstream deep learning computing platform: an Intel Core i9-10900K processor and an NVIDIA RTX 3090 24GB graphics card; the software environment is based on Windows 11, using Python 3.8, the PyTorch deep learning framework, and the CUDA 11.3 parallel computing environment to ensure the reproducibility of the experiment.

[0058] 2. Experimental Dataset This invention uses two sets of internationally recognized bearing benchmark datasets for verification, covering multiple operating conditions and multiple fault types, and verifying the universality and generalization ability of the verification method.

[0059] The first set of bearing datasets Four typical operating conditions were set up. The experiment adopted a single-source condition training and cross-condition testing paradigm: Condition A was selected as the source domain training set, containing known categories of normal, inner ring fault, and outer ring fault; Conditions B, C, and D were selected as the target test set, introducing composite faults as unknown category samples in addition to known categories. Vibration acceleration signals were collected at a sampling frequency of 64kHz, with a single sample length of 1024, and were used for model training and testing after normalization preprocessing.

[0060] The second set of bearing datasets Four variable speed and variable load operating conditions were set up, using the same experimental paradigm as the first dataset: condition 1 was used as the source domain training set, and conditions 2, 3, and 4 were used as the target test sets; the known categories were normal, inner race fault, and outer race fault, and the unknown category was set as rolling element fault. The signal sampling frequency was 48kHz, the single sample length was 1024, and the same normalization preprocessing method was used to ensure uniform experimental conditions.

[0061] 3. Comparison Methods and Evaluation Indicators Four mainstream industry methods were selected for comparison: Method 1: Traditional CNN closed set fault diagnosis method; Method 2: A DANN method based on depth domain adaptation; Method 3: Traditional open set recognition method based on feature distance; Method 4: Open set identification method based on classifier confidence.

[0062] All methods use the same training hyperparameters and the same hardware and software environment to ensure fairness in the comparison. The experiment uses classification accuracy (Acc), macro-average F1-score, unknown fault rejection rate (RR), ROC curve, and AUC value as comprehensive evaluation indicators.

[0063] (II) Experimental Results and Analysis 1. Comparison and Analysis of Experimental Results The performance results of each method on the two datasets are shown in Table 1.

[0064] Table 1. Experimental results of various methods on two sets of bearing datasets. As shown in Table 1: Known fault classification performance: The Acc and F1-score of this invention are significantly better than those of the comparison methods. On the first dataset, the Acc and F1-score of this method are improved by up to 14.47% and 14.13% respectively compared with the existing methods; on the second dataset, the Acc and F1-score are improved by up to 12.33% and 12.78% respectively compared with the existing methods, demonstrating significant advantages in cross-condition feature learning and known fault identification capabilities.

[0065] Unknown fault rejection performance: Methods 1 and 2 are closed-set architectures and lack the ability to identify unknown faults; Methods 3 and 4 have limited open-set rejection performance. On the first dataset, our method achieved a rejection rate of 93.38%, a maximum improvement of 33.13% compared to existing open-set methods; on the second dataset, the rejection rate reached 95.61%, a maximum improvement of 19.87% compared to existing open-set methods, demonstrating a significant improvement in unknown fault rejection performance.

[0066] Dataset adaptability: This invention maintains high stability and performance on two different benchmark datasets, and has good versatility and adaptability to different working conditions.

[0067] 2. Ablation Experiment Results and Analysis This invention comprises a collaborative framework consisting of an LCVG potential working condition view generation module, an IL-SCL topology consistency constraint module, and an SA-OSI operator-level open set inference module. Ablation experiments were conducted by removing each core module one by one to verify the necessity of each module; the results are shown in Table 2.

[0068] Table 2 Ablation experiment results of the method of the present invention on two sets of bearing datasets. Ablation results show that the LCVG module can effectively supplement the missing features of single-source operating conditions and improve the generalization ability across operating conditions; the IL-SCL constraint can maintain the topological stability of fault features and suppress operating condition drift interference; the SA-OSI module is the core unit for realizing open-set rejection of unknown faults, and its removal results in the loss of open-set recognition function. All modules work together and are indispensable, jointly constituting the complete technical architecture of this invention.

[0069] 3. ROC curve experimental results and analysis like Figure 3 As shown, the ROC curve is plotted with the false positive rate (FPR) on the horizontal axis and the true positive rate (TPR) on the vertical axis, and the AUC value represents the overall performance of open-set recognition. Since methods 1 and 2 cannot identify unknown faults, only the ROC curves of the method of this invention are compared with those of methods 3 and 4. The AUC value of this invention on the second dataset is 0.989, and the AUC value on the first dataset is 0.974, representing an improvement of at least 18.4% compared to other traditional methods. This maintains a high recognition rate while controlling the risk of false positives, further verifying the stability and superiority of the cross-condition open-set diagnosis of this invention.

[0070] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Anyone skilled in the art can make various modifications and alterations without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the claims.

Claims

1. A method for diagnosing single-source open-set topology stability faults in rotating machinery, comprising the following steps: S1. Acquire the monitoring signal of the rotating machinery under single-source operating conditions, and perform standardized preprocessing on the monitoring signal; S2. Perform multi-view amplification on the preprocessed monitoring signal to obtain multiple submersible condition view signals; S3. Extract high-dimensional fault features corresponding to the multiple potential working condition view signals through a deep convolutional neural network with shared parameters; Its characteristic is that it further includes the following steps: S4. Intra-class Laplacian Topology Alignment Training: For the high-dimensional fault features of each potential working condition view of each known fault category, construct the normalized graph Laplacian operator corresponding to each view, and use the average Laplacian operator of all views of the same fault category as the stable topology template of that category; construct the Laplacian structure consistency loss, align the normalized graph Laplacian operator of each view to the stable topology template of the corresponding category, and at the same time fuse the classification loss to construct the joint loss function to complete the collaborative optimization training of the model; S5. Template and threshold determination: Save the stable topology templates of each known fault category after training, collect the topology deviations corresponding to all known fault categories and all potential working condition views after training to form a global topology deviation set, and determine the open set identification threshold based on the statistical characteristics of the global topology deviation set. S6. Operator-level open set reasoning: For the test sample to be tested, after standardization preprocessing and feature extraction, for each known fault category, the new Laplace operator of the category is reconstructed based on the high-dimensional fault features of the test sample, and the topological deviation between the new Laplace operator and the corresponding category stable topological template is calculated. S7. Diagnostic result output: Compare the maximum value of the topological deviation of all categories with the open set identification threshold. If the maximum value does not exceed the open set identification threshold, the sample to be tested is determined to be a known fault and the corresponding fault category is output; if the maximum value exceeds the open set identification threshold, the sample to be tested is determined to be an unknown fault.

2. The method for diagnosing single-source open-set topology stability faults in rotating machinery according to claim 1, characterized in that, In step S4, the normalized graph Laplacian operator is calculated based on the class-adaptive weighted adjacency matrix and degree matrix of the view features corresponding to the fault category, and is used to encode the intrinsic manifold topology of the corresponding fault category samples.

3. The method for diagnosing single-source open-set topology stability faults in rotating machinery according to claim 1, characterized in that, In step S4, the Laplace structural consistency loss is used to constrain the structural deviation between the normalized graph Laplace operator of each view and the stable topology template of the same fault category, thereby achieving cross-condition stability of the manifold topology inherent in the fault category.

4. The method for diagnosing single-source open-set topology stability faults in rotating machinery according to claim 1, characterized in that, In step S5, the open set identification threshold is calculated based on the mean and standard deviation of the global topological deviation set using the method of "mean + preset fixed coefficient times standard deviation".

5. The method for diagnosing single-source open-set topology stability faults in rotating machinery according to claim 1, characterized in that, In step S6, the specific method for reconstructing the new Laplacian operator is as follows: for each known fault category, a sample feature in the training sample set corresponding to the stable topology template of the category is replaced with the high-dimensional fault feature of the test sample to be tested, resulting in a new sample set containing the test sample. The new Laplacian operator corresponding to the category is calculated using the same graphing rules as in the training phase.

6. The method for diagnosing single-source open-set topology stability faults in rotating machinery according to claim 1, characterized in that, The multi-view augmentation in step S2 is implemented by constructing a perturbation operator based on the physical characteristics of the changing operating conditions of rotating machinery equipment. This is used to simulate potential cross-operating condition changes, and the generated potential operating condition view signals have the same dimension as the preprocessed monitoring signals.

7. A single-source open-set topology stability fault diagnosis system for rotating machinery, characterized in that, The system includes a signal preprocessing module, a multi-view augmentation module, a shared feature extraction module, a topology consistency learning module, a template and threshold management module, an operator-level open set inference module, and a diagnostic result output module; the system is used to execute the single-source open set topology stability fault diagnosis method for rotating machinery equipment as described in any one of claims 1 to 6.

Images