A physical information graph network and agent variable working condition migration fault diagnosis method

Abstract

The application discloses a physical information graph network and intelligent agent variable working condition migration fault diagnosis method, belonging to the technical field of fault diagnosis. The method first constructs a multi-working condition migration data set and normalizes preprocessing, and obtains initial features through a convolution feature extraction module; then, through a physical information residual gate graph fusion mechanism, the data driven graph and the physical information graph are adaptively fused, and a fusion topology is output; subsequently, a double-path Transformer sampling aggregation graph network is constructed, and global context and sparse topology correlation are jointly modeled to obtain high-level discriminative features; finally, a double-agent reinforcement learning collaborative domain adaptive framework is introduced, cross-working condition distribution migration is realized through sample weighted screening and feature disturbance alignment, and target domain fault diagnosis is completed. The application fuses physical mechanism and data correlation, adopts double-path graph network and double-agent collaborative optimization, improves feature representation and cross-domain generalization ability under complex variable working conditions, and is suitable for unsupervised variable working condition migration fault diagnosis of mechanical equipment.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and more specifically, to a method for fault diagnosis of variable operating condition migration in physical information graph networks and intelligent agents. Background Technology

[0002] Mechanical equipment is a crucial foundation of modern industrial systems, and its operational status directly impacts production efficiency, equipment reliability, and production safety. Due to prolonged operation under high loads, strong interference, and complex, variable conditions, mechanical equipment is prone to wear, degradation, and even sudden failures. Furthermore, monitoring signals under different operating conditions often exhibit significant distributional differences, making it difficult for traditional intelligent fault diagnosis methods to achieve stable and effective cross-condition identification. In addition, in real-world industrial scenarios, data under the target operating condition is often difficult to accurately label, especially fault sample labeling, which is costly and challenging. This makes it difficult to directly apply traditional supervised learning methods to cross-condition diagnostic tasks.

[0003] While existing unsupervised transfer fault diagnosis methods have made some progress in cross-condition knowledge transfer, the following issues still require further research.

[0004] (1) Traditional UDA methods mostly focus on global distribution alignment, and do not fully explore the potential correlation and structural information between samples, resulting in limited ability to characterize fault features under complex working conditions; (2) Although existing graph neural network methods have enhanced the ability to learn cross-domain sample structure relationships, most of them still rely on data-driven graph construction, which does not make sufficient use of the physical mechanism and state evolution law of mechanical system, and lacks flexible adaptive control of dynamic domain alignment process.

[0005] (3) Existing reinforcement learning-assisted domain adversarial methods mostly adopt a single agent decision-making mode, which makes it difficult to simultaneously take into account source domain sample quality control and adversarial alignment intensity adjustment, thus limiting the model's collaborative optimization capability and cross-domain generalization performance in complex and variable working conditions. Summary of the Invention

[0006] 1. The technical problem that the invention aims to solve In view of the shortcomings of the existing technology, this invention provides a method for fault diagnosis of variable working conditions using a physical information graph network and an agent (EIPGFRLDA method). First, a physical information residual gated graph fusion mechanism is designed. By fusing physical information and data relationships, an adaptive topology is constructed, making feature representations more physically consistent while preserving sample correlation, thereby improving the model's cross-working-condition generalization ability. Second, a dual-path Transformer sampling aggregation graph network is constructed based on this graph to dynamically model intra-sample feature interactions and inter-sample topological relationships, improving the model's adaptability to structural correlation changes under complex working conditions. Finally, a dual-agent reinforcement learning collaborative domain adaptive framework is proposed. Through the collaborative optimization of sample selection and feature perturbation, the model's dynamic perception and adaptive control ability to complex working-condition changes is enhanced.

[0007] 2. Technical Solution To achieve the above objectives, the technical solution provided by the present invention is as follows: The present invention provides a method for fault diagnosis of variable operating condition migration in physical information graph networks and intelligent agents, comprising the following steps: Step 1: Construct a migration dataset containing vibration signals under multiple working conditions. Divide the multi-working-condition dataset into source domain data and target domain data according to the differences in working conditions, and perform normalization preprocessing to obtain preprocessed vibration signal samples. Step 2: Input the preprocessed vibration signal samples into the convolutional feature extraction module to extract the initial feature representation of the vibration signal samples; Step 3: Construct a physical information residual gating graph fusion mechanism. Input the initial features extracted in Step 2, construct and adaptively fuse the data-driven graph and the physical-driven graph through the physical information residual gating graph fusion mechanism, and output the fused topology structure. Step 4: Construct a dual-path Transformer sampling aggregation graph network. Using the initial features and fused topology as input, jointly model the global contextual dependencies and sparse topological associations between samples to obtain high-level discriminative features. Step 5: In the training phase, a dual-agent reinforcement learning collaborative domain adaptive framework is introduced. Through collaborative optimization of sample weighted selection and feature perturbation alignment, feature distribution transfer across working conditions is achieved. Step 6: Based on the trained model, identify faults in the target domain samples and output the diagnostic results.

[0008] Furthermore, the physical information residual gating graph fusion mechanism includes a data-driven branch, a physical information branch, and a residual gating unit. The initial features extracted by the convolutional feature extraction module are input into the data-driven branch and the physical information branch for parallel processing, respectively, to obtain the data-driven graph and the physical-driven graph. The residual gating unit performs adaptive weighted fusion of the data-driven graph and the physical-driven graph, and outputs the final fused topology structure.

[0009] Furthermore, the data-driven branch constructs a data association graph based on the feature similarity between samples. First, it performs linear mapping on the initial features to obtain feature representations suitable for relationship modeling. Then, it introduces a multi-head attention mechanism to perform interactive modeling on the mapped sample features. Based on the cosine similarity between sample pairs, it calculates the initial similarity matrix. After processing by the filtering and enhancement unit, it performs softmax normalization to obtain the data-driven graph.

[0010] Furthermore, the filtering and enhancement unit includes ReLU activation, Top-r filtering, self-loop enhancement, and symmetry operation. ReLU activation is used to suppress negatively correlated responses, Top-r filtering is used to retain the most representative adjacency relationship of each node, self-loop enhancement is used to strengthen the ability of nodes to maintain their own characteristics during graph propagation, and symmetry operation is used to ensure that the resulting graph structure satisfies the topological properties of an undirected graph.

[0011] Furthermore, the physical information branch combines the physical constraint relationships during equipment operation to construct a physical prior graph. First, the deep features of the samples are mapped to implicit dynamic state representations through a physical state discriminator. Then, physical constraints are constructed by referencing the dynamic relationships of a single-degree-of-freedom damped system. The cosine similarity between sample pairs is calculated, and the physical driving graph is obtained by temperature-scaled softmax normalization. The physical state discriminator includes a first linear transformation layer, a first ReLU activation layer, a first Dropout layer, a second linear transformation layer, a second ReLU activation layer, and a second Dropout layer arranged sequentially, which gradually maps the input deep features into the corresponding implicit dynamic state representation.

[0012] Furthermore, the residual gating unit uses the norm of the physical equation residuals to assign adaptive gating weights to each sample and adopts a domain-aware graph selection strategy. The source domain samples use a fused graph, while the target domain samples only use a data-driven graph, and the two domains share the same graph convolution weights.

[0013] Furthermore, the dual-path Transformer sampling aggregation graph network includes a Transformer path and a graph attention path. The Transformer path generates node representations with global context enhancement, while the graph attention path adaptively constructs a sparse dynamic topology and completes message propagation on the graph. The topology information is injected into the Transformer representation through the SAGE bridging mechanism, and unified modeling is achieved through the feature fusion transformation module.

[0014] Furthermore, the SAGE bridging mechanism samples, aggregates, and fuses the features of associated nodes within the neighborhood of the target node, injecting local neighborhood topology information from the graph structure into the global feature representation generated by the Transformer, thereby achieving collaborative modeling of global dependency information and local structural information.

[0015] Furthermore, the dual-agent reinforcement learning collaborative domain adaptive framework introduces two reinforcement learning agents based on proximal policy optimization: a sample selection agent and a feature perturbation agent. The collaborative optimization of the two agents is achieved through a three-stage progressive training strategy.

[0016] Furthermore, the sample selection agent takes high-level discriminative features as state input, outputs beta distribution parameters through a three-layer fully connected strategy network, samples to obtain sample weights, and multiplies the sample weights by the classification loss of the corresponding sample to suppress the gradient contribution of low-quality samples. The feature perturbation agent uniformly divides the high-level feature space into multiple subspaces, extracts the statistics of each subspace as the state representation, outputs the mean of the Gaussian distribution through the policy network, samples to obtain the low-dimensional core action vector, maps it through the perturbation projector to generate the perturbation vector, and uses norm constraints to limit the perturbation within a set proportion.

[0017] 3. Beneficial effects Compared with the prior art, the technical solution provided by this invention has the following advantages: (1) The present invention designs a Physical Information Residual Gated Graph Fusion Mechanism (PIRGF), which adaptively fuses the physical information graph and the data-driven graph through a residual gating strategy, and constructs a dynamic topology structure that takes into account both physical constraints and data association, thereby improving the physical consistency of feature representation and the cross-condition generalization ability of the model.

[0018] (2) The present invention constructs a dual-path Transformer sampling aggregation graph network (DP-TSGN), which integrates the global modeling capability of Transformer with the local aggregation capability of Sample and Aggregate (SAGE) to realize dynamic joint modeling of graph structure relationship and feature representation, thereby improving the model's adaptability to structural relationship changes under complex working conditions.

[0019] (3) This invention proposes a dual-agent collaborative domain adaptive framework (DACDA), which achieves dynamic optimization of the sample selection and domain alignment process by having two agents collaboratively perform sample quality assessment and feature perturbation alignment, thereby enhancing the model's cross-domain transfer capability and diagnostic stability under complex working conditions. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a structural diagram of the Physical Information Residual Gated Graph Fusion Mechanism (PIRGF) of the present invention; Figure 3 This is a structural diagram of the Dual-Path Transformer Sampling Aggregation Graph Network (DP-TSGN) of the present invention; Figure 4 This is a structural diagram of the dual-agent cooperative domain adaptive framework (DACDA) of the present invention. Detailed Implementation

[0021] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments.

[0022] Example 1 The effectiveness of the proposed method was validated by applying two widely used mechanical fault diagnosis datasets.

[0023] Firstly, the HUST bearing fault dataset provided by Huazhong University of Science and Technology (HUST) is used. This dataset includes nine bearing operating states: normal, inner ring fault, outer ring fault, ball fault, and inner-outer ring combined fault. Each fault state includes both moderate and severe faults. To verify the applicability of the proposed method under different operating conditions, this embodiment of the invention selects data corresponding to severe faults for analysis. Four operating conditions—30 Hz, 40 Hz, 60 Hz, and 0-40-0 Hz—are used as the source and target domains, respectively, to construct 12 cross-condition transfer diagnostic tasks to verify the diagnostic performance and generalization ability of the proposed method.

[0024] Secondly, the XJTUSuprGear dataset provided by Xi'an Jiaotong University (XJTU) was used. This dataset includes five gear operating states, namely the normal state and four fault states corresponding to different crack depths: 0.2 mm, 0.6 mm, 1.0 mm, and 1.4 mm. To further verify the applicability and generalization ability of the proposed method under different speed conditions, three operating conditions—900 r / min, 1200 r / min, and 0-1200-0 r / min—were used as the source and target domains, respectively, to construct six cross-condition transfer diagnostic tasks to verify the cross-domain diagnostic performance of the proposed method under different operating conditions.

[0025] Combination Figure 1 This embodiment presents a fault diagnosis method for complex operating conditions using a physical information graph network and intelligent agents. The method comprises a convolutional feature extraction module, a physical information residual gated graph fusion mechanism, a dual-path Transformer sampling aggregation graph network, and a dual-agent collaborative domain adaptive framework. The convolutional feature extraction module extracts local temporal features from the original vibration signal to obtain an initial feature representation. The physical information residual gated graph fusion mechanism constructs and adaptively fuses the data-driven graph and the physical-driven graph based on the initial features, outputting a fused topology. The dual-path Transformer sampling aggregation graph network takes the initial features and the fused topology as input, jointly models the global contextual dependencies and sparse topology between samples, obtaining high-level discriminative features. The dual-agent collaborative domain adaptive framework performs sample selection and feature perturbation co-optimization on the high-level discriminative features, achieving cross-operating condition feature distribution alignment, ultimately improving the model's fault diagnosis performance under complex operating conditions. The steps are as follows: Step 1: Construct a vibration signal migration dataset containing multiple operating conditions. The constructed multi-condition dataset is divided into source domain data and target domain data according to the differences in operating conditions, and then normalized preprocessing is performed to obtain preprocessed vibration signal samples, providing a reliable data foundation for subsequent model training and performance evaluation.

[0026] Step 2: Input the preprocessed vibration signal sample into the convolutional feature extraction module to extract the initial feature representation of the input vibration signal sample.

[0027] Step 3: Construct a physical information residual gating graph fusion mechanism based on the initial features. Construct and adaptively fuse the data-driven graph and the physical-driven graph through the physical information residual gating graph fusion mechanism, and output the fused topology structure. To improve the model's ability to model the correlation of fault features under multiple operating conditions, this invention designs a Physical Information Residual Gated Graph Fusion Mechanism (PIRGF). This mechanism constructs a data-driven graph and a physical information graph, and adaptively fuses the two graph structures by combining physical equation loss. While ensuring the flexibility of feature representation, it enhances the model's ability to represent the dynamic mechanisms of equipment. Its structure is as follows: Figure 2 As shown.

[0028] Specifically, the PIRGF includes a data-driven branch, a physical information branch, and a residual gating unit. Features X extracted by the convolutional feature extraction module are input into the data-driven branch and the physical information branch for parallel processing. The residual gating unit is used to adaptively weight and fuse the two types of graph structures, outputting the final fused adjacency matrix for use by the subsequent graph learning module. Furthermore, there is a correspondence between the adjacency matrix and the graph structure: given a graph, its adjacency matrix can be determined. Conversely, given the adjacency matrix, the graph's topology can be completely reconstructed.

[0029] The data-driven branch is used to construct a data association graph based on the feature similarity between samples to characterize the implicit structural relationships between faulty samples. Specifically, it first processes the initial features of the samples extracted in step two. X A linear mapping is performed to obtain a feature representation suitable for relation modeling. Subsequently, a multi-head attention mechanism is introduced to interactively model the mapped sample features, and an initial similarity matrix is ​​calculated based on the cosine similarity between sample pairs. Building upon this, to reduce the interference of redundant connections on the graph structure representation and enhance the effectiveness of adjacency relationships, this invention sets up a filtering and enhancement unit to perform nonlinear screening and structural constraint processing on the initial similarity matrix. The filtering and enhancement unit includes ReLU activation, Top-r filtering, self-loop enhancement, and symmetry operations. ReLU activation is used to suppress negatively correlated responses. Top-r filtering is used to retain the most representative adjacency relationships for each node to improve the sparsity of the graph structure; self-loop enhancement is used to strengthen the ability of nodes to maintain their own features during graph propagation; and symmetry operations are used to ensure that the resulting graph structure satisfies the topological properties of an undirected graph. Finally, the matrix processed by the filtering and enhancement unit is softmax normalized to obtain a data-driven adjacency matrix, which is further used to reconstruct the data-driven graph. This is used for subsequent graph feature learning and fault diagnosis. Its formula is defined as:

[0030] in, and The first The query matrix and key matrix of each attention head are defined as follows:

[0031] in, For the number of attention heads, For each head dimension. , The input features are mapped to query, key, and value representations respectively by a learnable projection matrix. and Used to calculate the similarity between samples to construct a graph topology. (Molecule) Calculate all sample pairs at the 1st The inner product similarity under the head, divided by the corresponding The cosine similarity is obtained by multiplying the norms. Take the average of the individual sizes. for Sparsification operation. This means averaging the matrix and its transpose to ensure the undirected symmetry of the graph structure. This is an identity matrix used to add self-loops so that each node retains its own information during aggregation. The outermost layer... This indicates that probability normalization is performed on each row, where the zero position is masked before normalization. .

[0032] The physical information branch is used to construct a physical prior graph by combining the physical constraints during equipment operation, thereby introducing prior information related to the system's dynamic state. Specifically, a physical state discriminator is first set up to map the deep features of the samples into implicit dynamic state representations, used to characterize the potential physical states of the samples during equipment operation. The physical state discriminator consists of multiple feedforward mapping units, specifically including a first linear transformation layer, a first ReLU activation layer, a first Dropout layer, a second linear transformation layer, a second ReLU activation layer, and a second Dropout layer arranged sequentially, used to progressively map the input deep features into corresponding implicit dynamic state representations. For any sample feature, its corresponding physical state can be represented as:

[0033] in, , and These correspond to implicit representations of displacement, velocity, and acceleration, respectively. Represents a physical state discriminator. This refers to the state dimension. To give these implicit state quantities physical meaning, this invention draws on the dynamic relationships of a single-degree-of-freedom damped system and constructs the following physical constraints:

[0034] Among them, quality Damping stiffness The parameters are initialized in the form of a learnable diagonal matrix, with each diagonal element initially assigned the values ​​1.0, 0.1, and 1.0, respectively, and the remaining elements set to 0. , , Let these represent displacement, velocity, and acceleration, respectively. Therefore, the physical equation residual is defined for each sample:

[0035] in, This indicates element-wise multiplication. The closer the state variables and parameters learned by the network are to the actual dynamic relationship, the better. The closer it gets to the zero vector. To utilize this fit to construct a graph structure for physical perception, the three sets of state variables are concatenated and normalized:

[0036] Then, the cosine similarity between sample pairs is calculated, and the physical information adjacency matrix is ​​obtained through temperature-scaled softmax normalization, which is then used to reconstruct the physical driving graph. :

[0037] in, Number of samples per batch, temperature coefficient This controls the sharpness of the similarity distribution. For the first The implicit dynamic state vector of a sample is composed of three sets of characteristics: displacement, velocity, and acceleration, which comprehensively characterize the complete motion state of the sample in the implicit physical space. Indicates sample With sample In the normalized physical state space, the larger the cosine similarity value, the closer the implicit dynamic behaviors of the two samples are.

[0038] because , , and , and All are randomly initialized learnable quantities, with the physics-driven graph in the early stages of training. This is unreliable. Therefore, this invention constructs a residual gating unit, utilizing residual... The norm is used to assign adaptive gating weights to each sample.

[0039]

[0040] When the sample residuals are small When the residuals are large, the model tends to retain structural information from the physics-driven graph. The smaller the graph, the more the model relies on the data-driven graph. Therefore, the final fused adjacency matrix can be represented as:

[0041] in, As a scalar, it represents the value of the sample. All connections in the same row are weighted uniformly. This residual gating strategy ensures that the introduction of physical information is gradual, so that the physical driving graph will only substantially participate in shaping the graph structure once the network has achieved sufficient accuracy in fitting the physical relationships.

[0042] Furthermore, in cross-domain migration scenarios, the generalization of physical parameters is difficult to guarantee because the source and target domains may correspond to different operating conditions. To avoid the source domain learning... , , In cases where erroneous graph structures are generated in the target domain, leading to negative transfer, this invention employs a domain-aware graph selection strategy, namely, using fused graphs for source domain samples. The target domain samples only use data-driven graphs. The two domains share the same graph convolution weights to achieve feature space alignment, but the way the graph structure is constructed varies from domain to domain, thereby utilizing physical information while avoiding the risk of cross-domain physical mismatch.

[0043] Through the above methods, the present invention can simultaneously take into account data representation capability and physical consistency under complex working conditions, effectively improving the accuracy and robustness of graph structure modeling, and providing a more discriminative topological representation for subsequent fault diagnosis.

[0044] Step 4: Construct a dual-path Transformer sampling aggregation graph network to dynamically and jointly model the global contextual dependencies and sparse topological associations between samples.

[0045] This invention provides a Dual-Path Transformer Sample-and-Aggregate Graph Network (DP-TSGN), the structure of which is as follows: Figure 3 As shown. The specific network includes a Transformer path and a graph attention path. This network uses the sample features extracted in step two. As input, two complementary paths are constructed: a Transformer path for generating globally context-enhanced node representations, and a graph attention path for adaptively constructing sparse dynamic topologies and completing message propagation on the graph. Finally, the topological information is injected into the Transformer representation through the SAGE bridging mechanism, and unified modeling is achieved in conjunction with the feature fusion transformation module, thereby realizing deep coupling and collaborative representation of the two paths at the structural level.

[0046] The Transformer path constructs a node representation for each sample that incorporates global interaction information. First, the features... Projecting into the hidden space yields Subsequently, the Transformer encoder aggregates intra-batch contextual information for each sample under a global attention mechanism, resulting in the encoded feature representation. Simultaneously, local information from the projection stage and global semantics from the encoding stage are preserved, and the two are concatenated to obtain the node representation:

[0047] For the graph attention path, the structural relationships between samples are first modeled based on the proposed Physics-Informed Residual-Gated Graph Fusion (PIRGF) mechanism, and the resulting adjacency matrix is ​​then used to... (source domain Target domain ) and node features after graph attention propagation Based on this, a self-loop is applied to the adjacency matrix and then normalized to obtain the normalized adjacency matrix. .

[0048] Furthermore, to achieve a normalized adjacency matrix Representation of Transformer path nodes To effectively integrate these elements, this invention designs SAGE-style neighborhood aggregation as a bridging mechanism between two paths. Specifically, this is achieved by representing nodes on a graph topology information structure. The SAGE Bridge Mechanism is used to aggregate neighborhood information and concatenate features to obtain a structure-enhanced feature representation.

[0049] in, Represents the learnable feature transformation weights. This step involves layer normalization. Through this process, the representation of each node not only includes the global semantic information provided by the Transformer path but also encodes the neighborhood structure information characterized by graph topological relationships. The SAGE bridging mechanism establishes a feature transfer channel between the Transformer path and the graph attention path. Through the neighborhood sampling and aggregation operations of GraphSAGE, local neighborhood topological information in the graph structure is injected into the global feature representation generated by the Transformer, thereby achieving collaborative modeling of global dependency information and local structural information. Specifically, using the feature representation output by the Transformer path as input, the neighboring nodes of the target node are sampled according to the constructed graph adjacency relationships. The sampled neighborhood features are then aggregated and fused with the target node's own features to obtain an enhanced feature representation containing local topological relationships. This invention, through the above method, retains the Transformer path's ability to model long-range dependencies while enhancing the model's ability to perceive local association patterns and topological changes between samples, thus providing a more discriminative representation for subsequent feature fusion and fault identification.

[0050] In obtaining graph attention path features With structural enhancement features Subsequently, the Feature Fusion Transformation Module (FFTM) designed in this invention is used to uniformly model the two types of features. This module first concatenates features from different paths and adaptively adjusts different feature channels through a learnable gating mechanism. Then, a multilayer perceptron is used to perform a nonlinear mapping on the fused features, and representations from the Transformer are introduced. Linear residual connections are used to enhance the stability of model training, resulting in the final output representation:

[0051] FFTM stands for Feature Fusion Transformation Module. This module first concatenates features from the graph attention path and the SAGE bridging mechanism, and then adaptively adjusts different feature channels through a learnable gating mechanism to dynamically balance the contributions of the two types of information. The SAGE bridging mechanism is used to sample, aggregate, and fuse features of associated nodes within the neighborhood of the target node, thereby obtaining node representations containing local topological relationship information. Subsequently, a two-layer fully connected network is used to perform a nonlinear transformation on the fused features, and node representations from the Transformer branch are introduced. Linear residual connections are used to enhance the model's expressive power and alleviate gradient degradation during deep network training.

[0052] After completing convolutional feature extraction, physical information graph fusion, and feature modeling of the dual-path Transformer graph network, the model can obtain high-level discriminative features that combine local temporal information, global contextual dependencies, and cross-sample topological associations. However, due to significant differences in data distribution between different working conditions, directly using source domain knowledge for fault identification in the target domain can easily lead to a decline in the model's generalization performance. Therefore, it is necessary to further dynamically align and optimize the cross-working-condition feature distribution through a domain adaptation mechanism to improve the model's robustness and diagnostic accuracy in complex working-condition migration scenarios. Based on this, this invention introduces a dual-agent reinforcement learning collaborative domain adaptation framework after the feature learning stage to jointly optimize the cross-domain sample selection and feature perturbation process.

[0053] Step 5: In the training phase, a dual-agent reinforcement learning collaborative domain adaptive framework is introduced. Through collaborative optimization of sample weighting and feature perturbation alignment, cross-condition feature distribution transfer is achieved; for example... Figure 4 As shown, this invention proposes a dual-agent cooperative domain adaptation (DACDA) framework for reinforcement learning, which introduces two reinforcement learning agents based on proximal policy optimization (PPO) to undertake sample selection and feature perturbation tasks respectively, and achieves their collaboration through a three-stage progressive training strategy.

[0054] Specifically, Agent-S (the sample selection agent) aims to evaluate the contribution of each sample from the source domain to cross-domain migration and assign it a weight between 0 and 1. Agent-S takes the high-level feature representation output from the dual-path Transformer sampling aggregation graph network in step four as its state input, and outputs two parameters of the beta distribution through a three-layer fully connected policy network. and Sample weights are obtained by sampling from this distribution. This weight is directly multiplied by the classification loss of the corresponding sample, thus suppressing the gradient contribution of low-quality samples.

[0055] in and The first One source domain sample and its label, For the classifier output, Let cross-entropy be the loss function. This represents the number of samples in the source domain.

[0056] In addition, Agent-S's reward signal comprehensively considers classification fidelity. and combat confusion :

[0057]

[0058] in, Indicates the first The confidence level that a source domain sample is correctly classified by the classifier. This represents the predicted probability of the classifier for the correct category after perturbation. It takes the maximum value of 1 when the classifier is completely correct and the minimum value of -1 when it is completely wrong. The domain discriminator represents the first domain. The output probability of each source domain sample Let be the output probability of the domain discriminator. During training, the source domain label is set to 0 and the target domain label is set to 1. Both are mapped to the interval [ ]. [1,1], respectively measuring the classification correctness and domain transfer potential of the samples. To prevent the policy from degenerating to discarding all samples, an intra-batch average weight is introduced. The anti-collapse penalty. The reward function of Agent-S is defined as:

[0059] This reward design makes the agent tend to assign high weights to samples that are correctly classified and easy to transfer, while being penalized when the overall weight is too low.

[0060] The goal of Agent-F (Feature Perturbation Agent) is to apply controllable adversarial perturbations to source domain features, generating virtual cross-domain features to increase the training difficulty of the domain discriminator. Specifically, Agent-F first uniformly divides the high-level feature space output from step four into... Each subspace is used to extract statistics (mean, standard deviation, maximum, and minimum) as state representations, which are then output through a policy network. The mean of a Gaussian distribution is used to sample and obtain a low-dimensional core action vector. This action is then mapped back to the original feature dimension via a perturbation projector to generate a perturbation vector. To ensure the perturbation amplitude is controllable, a norm constraint is used to limit the perturbation to a value of the original characteristic energy. Within the proportion ( Virtual features after perturbation The data is fed into the domain discriminator via a gradient reversal layer, and adversarial training is driven by a binary cross-entropy loss. The rewards for Agent-F are designed separately for the source and target domains:

[0061]

[0062] in, Predict the entropy for the target domain, superscript , Representing the source domain and the target domain respectively. Intra-batch perturbation vector The mean of the norm. The source domain reward requires the perturbation to confuse the discriminator while maintaining classification fidelity with double the weight; the target domain reward starts from the reverse confusion and additionally penalizes cases with excessive prediction uncertainty; both include sparse penalty terms to constrain the perturbation magnitude. At the same time, a consistency constraint is introduced into the main network loss, requiring the classifier output distribution before and after the perturbation to remain consistent through KL divergence, preventing the perturbation from destroying the learned discriminative structure.

[0063] To ensure that the capabilities of the two agents match the learning progress of the main network, this invention designs a three-stage progressive training strategy.

[0064] (1) In stage 0 (warm-up period, Epoch0-10), only the feature extraction backbone network constructed in steps 2 to 4 is trained, and the extracted features are classified and output through the classification head. At the same time, the dual agent parameters are frozen so that the network can learn stable fault discrimination features first, thereby providing reliable state input for subsequent agents.

[0065] (2) Phase 1 (sample screening period, Epoch 10–30) Activate Agent-S and domain discriminator based on Phase 0, but turn off the perturbation function of Agent-F. Agent-S gradually reduces the weights of noisy samples and outliers to near zero, while the domain discriminator establishes preliminary domain discrimination ability on the original features.

[0066] (3) Phase 2 (Feature Refinement Period, Epoch 30 to End): Based on the stable sample selection results, the Phase 1 weights (S1 weights) are loaded to activate Agent-F, forming a dynamic adversarial game between the feature extractor and Agent-F. In this phase, based on the domain adversarial loss, a perturbation consistency constraint is further introduced. And multi-granularity alignment loss, encompassing four classic alignment losses: target domain entropy minimization, class-level feature center cosine alignment, CORAL second-order statistic alignment, and MMD. This collaboratively promotes the alignment of cross-domain feature distributions from multiple levels. The total loss for Stage 2 is:

[0067] in, The classification loss is weighted by the Agent-S sample weights. For domain adversarial binary cross-entropy loss with sample weight balancing:

[0068] in, For the domain label (source domain is 0, target domain is 1). The number of samples in the target domain. For sample balancing weights, the source domain samples' The target domain samples are obtained by scaling the Agent-S weights using a balancing coefficient. . The representation domain discriminator is used to characterize the distribution and belonging probability of features between the source and target domains.

[0069] The weighting coefficient for the physical consistency loss. The residual constraint loss of the physical equations, Consistent constraints on the KL divergence of the classifier output distribution before and after perturbation:

[0070] in, The raw features output by the feature extractor. The perturbation features generated for Agent-F For category labels, and These are the soft label probability distributions output by the classifier for the features before and after the perturbation, respectively. Let KL divergence be the constraint. This constraint requires that the classifier predictions remain consistent before and after the perturbation, preventing Agent-F from disrupting the learned class discrimination structure while increasing domain confusion.

[0071] The weighted combination of multi-granularity alignment losses is shown in the formula.

[0072]

[0073] in, This represents the loss that minimizes the prediction entropy of high-confidence samples in the target domain. This represents the class-level feature center cosine alignment loss based on pseudo-labels. The Frobenius norm difference represents the covariance moments between the source and target domains. MMD first-moment matching loss represents the mean features of the source and target domains.

[0074] Step 6: Based on the trained model, perform fault identification on the target domain samples and output the diagnostic results.

[0075] To verify the fault diagnosis performance and generalization ability of the proposed EIPGFRLDA method, comparative experiments were conducted on the HUST bearing dataset and the XJTUSpurGear dataset. The comparison methods included five representative or state-of-the-art fault diagnosis models: DDTLN, UGTNH, TF-SupCon, MSM-SWCNN, and DSMDTN. The experimental results of each method under different transfer task settings on the HUST bearing dataset and the XJTUSpurGear dataset are shown in Tables 1 and 2. The results show that the proposed EIPGFRLDA method outperforms the aforementioned methods in both diagnostic accuracy and generalization ability.

[0076] Table 1 Comparative experimental results based on the HUST bearing dataset

[0077] Table 2 Comparative experimental results based on the XJTUSpurGear dataset

[0078] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention, and are not actually limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for fault diagnosis of variable operating condition migration in a physical information graph network and an intelligent agent, characterized in that, Includes the following steps: Step 1: Construct a migration dataset containing vibration signals under multiple working conditions. Divide the multi-working-condition dataset into source domain data and target domain data according to the differences in working conditions, and perform normalization preprocessing to obtain preprocessed vibration signal samples. Step 2: Input the preprocessed vibration signal samples into the convolutional feature extraction module to extract the initial feature representation of the vibration signal samples; Step 3: Construct a physical information residual gating graph fusion mechanism. Input the initial features extracted in Step 2, construct and adaptively fuse the data-driven graph and the physical-driven graph through the physical information residual gating graph fusion mechanism, and output the fused topology structure. Step 4: Construct a dual-path Transformer sampling aggregation graph network. Using the initial features and fused topology as input, jointly model the global contextual dependencies and sparse topological associations between samples to obtain high-level discriminative features. Step 5: In the training phase, a dual-agent reinforcement learning collaborative domain adaptive framework is introduced. Through collaborative optimization of sample weighted selection and feature perturbation alignment, feature distribution transfer across working conditions is achieved. Step 6: Based on the trained model, identify faults in the target domain samples and output the diagnostic results.

2. The method for fault diagnosis of variable operating condition migration in a physical information graph network and an agent according to claim 1, characterized in that, The physical information residual gating graph fusion mechanism includes a data-driven branch, a physical information branch, and a residual gating unit. The initial features extracted by the convolutional feature extraction module are input into the data-driven branch and the physical information branch for parallel processing to obtain the data-driven graph and the physical-driven graph, respectively. The residual gating unit performs adaptive weighted fusion of the data-driven graph and the physical-driven graph to output the final fused topology.

3. The method for fault diagnosis of variable operating condition migration in a physical information graph network and an agent according to claim 2, characterized in that, The data-driven branch constructs a data association graph based on the feature similarity between samples. First, it performs linear mapping on the initial features to obtain feature representations suitable for relationship modeling. Then, it introduces a multi-head attention mechanism to perform interactive modeling on the mapped sample features. Based on the cosine similarity between sample pairs, it calculates the initial similarity matrix. After filtering and enhancement unit processing, it is normalized by softmax to obtain the data-driven graph.

4. The method for fault diagnosis of variable operating condition migration in a physical information graph network and an intelligent agent according to claim 3, characterized in that, The filtering and enhancement unit includes ReLU activation, Top-r filtering, self-loop enhancement, and symmetry operation. ReLU activation is used to suppress negatively correlated responses, Top-r filtering is used to retain the most representative adjacency relationship of each node, self-loop enhancement is used to strengthen the ability of nodes to maintain their own characteristics during graph propagation, and symmetry operation is used to ensure that the resulting graph structure satisfies the topological properties of an undirected graph.

5. A method for fault diagnosis of variable operating condition migration in a physical information graph network and an agent according to any one of claims 2-4, characterized in that, The physical information branch combines the physical constraint relationships during equipment operation to construct a physical prior graph. First, the deep features of the samples are mapped to implicit dynamic state representations through a physical state discriminator. Then, physical constraints are constructed by referencing the dynamic relationships of a single-degree-of-freedom damped system. The cosine similarity between sample pairs is calculated, and the physical driving graph is obtained by temperature-scaled softmax normalization. The physical state discriminator includes a first linear transformation layer, a first ReLU activation layer, a first Dropout layer, a second linear transformation layer, a second ReLU activation layer, and a second Dropout layer arranged sequentially, which gradually maps the input deep features into the corresponding implicit dynamic state representation.

6. The method for fault diagnosis of variable operating condition migration in a physical information graph network and an intelligent agent according to claim 5, characterized in that, The residual gating unit uses the norm of the physical equation residuals to assign adaptive gating weights to each sample and adopts a domain-aware graph selection strategy. The source domain samples use a fused graph, while the target domain samples only use a data-driven graph. The two domains share the same graph convolution weights.

7. The method for fault diagnosis of variable operating condition migration in a physical information graph network and an intelligent agent according to claim 6, characterized in that, The dual-path Transformer sampling aggregation graph network includes a Transformer path and a graph attention path. The Transformer path generates node representations with global context enhancement, while the graph attention path adaptively constructs a sparse dynamic topology and completes message propagation on the graph. The topology information is injected into the Transformer representation through the SAGE bridging mechanism, and unified modeling is achieved through the feature fusion transformation module.

8. The method for fault diagnosis of variable operating condition migration in a physical information graph network and an intelligent agent according to claim 7, characterized in that, The SAGE bridging mechanism samples, aggregates, and fuses the features of associated nodes within the neighborhood of the target node, injecting local neighborhood topology information from the graph structure into the global feature representation generated by the Transformer, thereby achieving collaborative modeling of global dependency information and local structural information.

9. The method for fault diagnosis of variable operating condition migration in a physical information graph network and an intelligent agent according to claim 8, characterized in that, The dual-agent reinforcement learning collaborative domain adaptive framework introduces two reinforcement learning agents based on proximal policy optimization: a sample selection agent and a feature perturbation agent. The collaborative optimization of the two agents is achieved through a three-stage progressive training strategy.

10. The method for fault diagnosis of variable operating condition migration in a physical information graph network and an intelligent agent according to claim 9, characterized in that, The sample selection agent takes high-level discriminative features as state input, outputs beta distribution parameters through a three-layer fully connected strategy network, samples sample weights, multiplies the sample weights by the classification loss of the corresponding samples, and suppresses the gradient contribution of low-quality samples. The feature perturbation agent uniformly divides the high-level feature space into multiple subspaces, extracts the statistics of each subspace as the state representation, outputs the mean of the Gaussian distribution through the policy network, samples to obtain the low-dimensional core action vector, maps it through the perturbation projector to generate the perturbation vector, and uses norm constraints to limit the perturbation within a set proportion.

Images