Graph causal self-adaption-based high-end mechanical equipment distribution outside intelligent diagnosis method

Abstract

The invention discloses a graph causal-based adaptive high-end mechanical equipment distribution outside intelligent diagnosis method, and relates to the technical field of high-end mechanical equipment distribution outside intelligent diagnosis, and the method comprises the steps: collecting a vibration acceleration signal in a typical fault state through a typical fault vibration test experiment; constructing a multi-layer graph structure through Euclidean distance and cosine similarity, connecting nodes by using K-neighborhood, and performing time sequence signal analysis by using the K-neighborhood as input of a graph neural network; introducing a backdoor adjustment strategy and an approximate intervention method based on a pseudo environment label to eliminate interference of environmental factors; the data is generalized by multi-layer pseudo-environment representation in combination with a GCN or GAT encoder. According to the graph causal self-adaption-based high-end mechanical equipment distribution external intelligent diagnosis method provided by the invention, on the basis of the internal complexity of graph structure data, a pseudo environment label is used as an intermediate potential variable, and a node characteristic propagation path is dynamically adjusted, so that stable adaptation to diversified working conditions is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed out-of-distribution intelligent diagnosis for high-end mechanical equipment, and specifically to a method for distributed out-of-distribution intelligent diagnosis for high-end mechanical equipment based on graph causality adaptation. Background Art

[0002] In recent years, graph neural networks have shown great application potential in multiple fields due to their excellent performance in processing complex topological structures and non-Euclidean data. However, it is worth mentioning that most existing graph neural networks and other deep learning methods are based on the closed-set assumption of constant data distribution, that is, it is assumed that the training and test data come from the same distribution. When the data distribution changes due to factors such as environmental changes and sensor drift during the actual operation of mechanical equipment, the performance of existing models will drop significantly. Especially when facing out-of-distribution data (OOD), existing deep learning models usually become overconfident, resulting in inaccurate predictions, thus affecting their reliability in critical diagnosis tasks.

[0003] Based on the existing predicament, experts naturally think about a problem involving GNNs: "How to prepare for sudden scenarios (out-of-distribution data) during peacetime (in-distribution data), that is, endow GNNs with the ability of extrapolation." Existing methods mostly deal with distribution shift by improving model robustness or introducing domain adaptation training, but these methods often rely on explicit environmental labels or assume known environmental distributions, and it is difficult to adapt to the situation of lack of prior knowledge in real scenarios. In addition, these methods are prone to capturing unstable environment-sensitive features, resulting in limited generalization ability of the model in practical applications. Especially in a complex and changeable application scenario such as mechanical equipment, existing solutions face problems such as model overfitting or being overly sensitive to specific environments in practical applications. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the technical problems solved by the present invention are: existing methods for distributed out-of-distribution intelligent diagnosis of high-end mechanical equipment lack effective extrapolation ability, lack effective strategies for dealing with unlabeled or unknown environmental distributions, and how to endow graph neural networks with stronger generalization ability in complex environments, improve their robustness on out-of-distribution data, and at the same time avoid overfitting or excessive sensitivity to specific environments.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: A graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment, including collecting vibration acceleration signals in typical fault states through typical fault vibration test experiments; constructing a multi-layer graph structure through Euclidean distance and cosine similarity, using K-nearest neighbors to connect nodes, and using it as the input of a graph neural network for time series signal analysis; introducing a backdoor adjustment strategy and an approximate intervention method based on pseudo-environment labels to eliminate the interference of environmental factors; generalizing data through multi-layer pseudo-environment representation combined with GCN or GAT encoders.

[0007] As a preferred embodiment of the graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment according to the present invention, wherein: the collection of vibration acceleration signals in typical fault states includes conducting vibration test experiments to collect vibration acceleration signals in typical fault states. The single-state signals include inner race fault (IF), outer race fault (OF), rolling element fault (BF), and normal state (NA). The composite fault states include outer race + rolling element fault and inner race + rolling element fault. The vibration signals under each fault state are tested using five different-sized grooves.

[0008] As a preferred embodiment of the graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment according to the present invention, wherein: the construction of a multi-layer graph structure through Euclidean distance and cosine similarity includes setting the time series data collected by the mechanical equipment under n different domain rotational speeds and loss magnitude states ω n (n = 1, 2, …, n) as where k = 1, 2, …, K represents K working condition states under the nth domain, and the time series of each working condition state is marked L as the label set y (n) ∈{1, 2, …, K}, defining the time window length as Δt, and dividing each time series into non-overlapping subsequences, where L represents the length of the original time series, k Therefore, in the node generation stage, each subsequence is mapped to a node The node set is represented as and multi-dimensional time domain features are extracted to generate its feature vector to characterize the local dynamic behavior characteristics under a single working condition. In each domain layer ω n within, local connection relationships are constructed through similarity measurement between nodes. Given nodes and the similarity is measured by calculating the Euclidean distance, expressed as:

[0009]

[0010] Connect each node to its nearest neighbor in a K-nearest neighbor manner to form an intra-layer graph structure

[0011] When constructing a single-layer graph After that, define cross-layer connections to capture the global behavior patterns under multi-operation conditions. For any two nodes from different layers and Measure the similarity of cross-layer nodes by cosine similarity, expressed as:

[0012]

[0013] When the residual similarity is greater than the preset threshold η, that is, introduce a cross-layer edge connection between and Integrate the operation characteristics under different domain states, construct a graph structure with multi-level dependencies, and the adjacency matrix and the feature matrix constructed by intra-layer and cross-layer connections together constitute the graph structure data input into the GNN model

[0014] As a preferred solution of the graph causal adaptive intelligent diagnosis method for high-end mechanical equipment based on distribution outside the domain of the present invention, wherein: the backdoor adjustment strategy includes intervening on the environmental variable M through the introduction of the do-operation, and the do-operation eliminates the interference of the environmental variable M on the graph feature So that the model only captures the stable causal relationship q between θ (Y|do(G)), introduce the backdoor adjustment strategy based on the observed data, and approximate the causal intervention effect by solving the following formula:

[0015]

[0016] Among them, q0(M) represents the prior distribution of the environmental variable M;

[0017] Let the pseudo-environment estimator p Ω (M|G), infer the pseudo-environment variable m of node v according to the self-graph feature of the node, and input the inferred pseudo-environment m v and the self-graph feature v into the GNN predictor q (Y|G,M) for joint optimization to obtain: θ (Y|G,M) for joint optimization to obtain:

[0018]

[0019] Among them, is the supervised training loss, is the regularization loss.

[0020] As a preferred solution of the graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment according to the present invention, wherein: the pseudo-environment estimator includes the pseudo-environment estimator p Ω (M|G), infers the pseudo-environment representation during the feature aggregation process of each layer of the graph neural network, and the pseudo-environment is used as the latent variable of each layer, is inferred through the feature aggregation representation of node v, and is defined in the form of a Z-dimensional numerical vector, and the categorical distribution is used as the sampling basis, and the model conditional probability is represented based on the node embedding of the current layer as follows:

[0021]

[0022] wherein, W (l) represents the learnable weight matrix of the l-th layer. By introducing the Gumbel noise α z , the sampling process can approximate discretization while maintaining continuity, and is expressed as:

[0023]

[0024] wherein, α z is the random noise sampled from the Gumbel distribution, and λ is the hyperparameter.

[0025] As a preferred solution of the graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment according to the present invention, wherein: the GNN predictor includes an expert ensemble architecture based on the graph convolutional network, and the layer-by-layer feature update is expressed as:

[0026]

[0027] wherein, δ v and δ u are the degrees of nodes v and u, and W (l,z) are respectively the linear transformation matrices of the neighbor node information and the self-node information of the z-th expert branch of the l-th layer, is the activation function;

[0028] Construct an adaptive expert model based on the attention mechanism, and construct and model the pairwise interaction relationship between nodes, which is expressed as:

[0029]

[0030] wherein, represents the attention weight between nodes.

[0031] As a preferred solution of the graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment according to the present invention, wherein: generalizing data by combining multi-layer pseudo-environment representation with a GCN or GAT encoder includes combining a GCN or GAT as the backbone structure of the encoder, and using the feature extraction ability in the non-Euclidean space to improve the model's expression of high-dimensional topological data.

[0032] Another object of the present invention is to provide a graph-causal adaptive out-of-distribution intelligent diagnosis system for high-end mechanical equipment, which can, based on the principle of causal intervention, propose a new learning objective, and can, without environmental labels, infer pseudo-environment information, thereby effectively eliminating environmental confounding biases in the data, helping the model learn stable prediction relationships, and solving the problem that the current out-of-distribution intelligent diagnosis methods for high-end mechanical equipment lack effective strategies for dealing with unlabeled or unknown environmental distributions.

[0033] As a preferred solution of the graph-causal adaptive out-of-distribution intelligent diagnosis system for high-end mechanical equipment according to the present invention, wherein: it includes an experimental detection module, a layer construction module, an interference elimination module, and a combination optimization module; the experimental detection module is used to collect vibration acceleration signals in typical fault states through typical fault vibration test experiments; the layer construction module is used to construct a multi-layer graph structure through Euclidean distance and cosine similarity, connect nodes using K-nearest neighbors, and perform time-series signal analysis as the input of the graph neural network; the interference elimination module is used to introduce a backdoor adjustment strategy and an approximate intervention method based on pseudo-environment labels to eliminate the interference of environmental factors; the combination optimization module is used to generalize data by combining multi-layer pseudo-environment representation with a GCN or GAT encoder.

[0034] A computer device includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the steps of the graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment.

[0035] A computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, it implements the steps of the graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment.

[0036] Advantages of the present invention: The graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment provided by the present invention is based on the inherent complexity of graph-structured data. Using pseudo-environment labels as intermediate latent variables, by dynamically adjusting the node feature propagation path, it can robustly adapt to diverse working conditions. Based on the graph-causal inference theory, a pseudo-environment label estimator is designed. Through Gumbel-Softmax for differentiable sampling, the pseudo-environment latent variable is dynamically inferred from the observed node features, thereby eliminating the confounding effect brought by environmental changes. The present invention has achieved better results in terms of applicability, stability, and generalization. Brief Description of the Drawings

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0038] Figure 1 It is the overall flowchart of a graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment provided by the first embodiment of the present invention.

[0039] Figure 2 It is the experimental result graph of the out-of-distribution generalization performance of GCI-ODG of a graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment provided by the second embodiment of the present invention on two datasets.

[0040] Figure 3 It is the ablation experiment research and performance evaluation graph of the GCI-ODG model of a graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment provided by the second embodiment of the present invention.

[0041] Figure 4 It is the analysis graph of the hyperparameter Z and the performance of GCI-ODG of a graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment provided by the second embodiment of the present invention. Detailed Embodiments

[0042] To make the above objects, features, and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0043] Embodiment 1, refer to Figure 1, which is an embodiment of the present invention, provides a graph-causal adaptive intelligent diagnosis method for high-end mechanical equipment distribution, including:

[0044] S1: Collect vibration acceleration signals under typical fault conditions through typical fault vibration test experiments.

[0045] Furthermore, the vibration acceleration signals under typical fault conditions include: four single fault conditions, inner race fault (IF), outer race fault (OF), rolling element fault (BF), and normal condition (NA). In addition, compound fault conditions are also collected, including outer race + rolling element fault (CO-1) and inner race + rolling element fault (CO-2), for a total of 6 fault conditions. The vibration signals under each fault condition are tested with five different sizes of grooves, which are processed by wire cutting. The specific sizes are: 0.4×1mm (width 0.4mm, depth 1mm), 2×2mm, 2.8×3mm, 3.4×4, and 4×4mm. In order to simulate the changes in the environment under actual operating conditions, the vibration signals of the faulty bearing are collected at five different speeds (500r / min, 700r / min, 900r / min, 1000r / min, and 1100r / min) during the experiment.

[0046] S2: Construct a multi-layer graph structure through Euclidean distance and cosine similarity, use K-nearest neighbors to connect nodes, and use it as the input of the graph neural network for time series signal analysis.

[0047] Furthermore, in the field of intelligent diagnosis, especially when dealing with the identification and classification tasks of the operating conditions of high-end mechanical equipment, the data under multi-condition conditions is complex and high-dimensional. To address this challenge, this invention patent proposes a graph neural network (GNN) method based on a hierarchical graph structure. This method aims to capture local features while integrating global information through the structured representation of time series data, thereby enhancing the generalization ability of the model. The construction of the multi-layer graph structure through Euclidean distance and cosine similarity includes setting the time series data collected by the mechanical equipment under n different domain speeds and loss magnitude states ω n (n = 1, 2, …, n) as where k = 1, 2, …, K represents K operating condition states in the nth domain, and the time series of each operating condition state is marked as the label set y (n) ∈{1, 2, …, K}, define the time window length as Δt, and divide each time series into non-overlapping subsequences, where L represents the length of the original time series. Therefore, in the node generation stage, each subsequence is mapped to a node The node set is represented as Extract multi-dimensional time-domain features to generate its feature vector Characterize the local dynamic behavior characteristics under a single working condition, in each domain layer ω n Within, construct local connection relationships through similarity measures between nodes, that is, connect a given node within the layer (intra-layer) and Measure similarity by calculating the Euclidean distance, expressed as:

[0048]

[0049] Apply the K-nearest neighbor method to connect each node to its nearest neighbor node to form an intra-layer graph structure

[0050] After constructing a single-layer graph Define cross-layer connections to capture the global behavior pattern under multi-working condition conditions. For any two nodes from different layers and Measure the similarity of cross-layer nodes by cosine similarity, expressed as:

[0051]

[0052] When the residual similarity is greater than the preset threshold η, that is, introduce a cross-layer edge connection between and Integrate the working condition characteristics under different domain states, construct a graph structure with multi-level dependencies, and jointly form the graph structure data input into the GNN model through the adjacency matrix and the feature matrix The graph structure data is input into the GNN model for the classification and recognition of the operating states of mechanical equipment under multi-working conditions. This graph structure effectively captures the local features under each working condition state within a single layer through node connections, and at the same time utilizes the cross-layer connection mechanism to achieve the deep integration of different state information. This design enables the model to fully learn the complex data characteristics in a multi-level and multi-working condition environment during the training process, thereby significantly improving the generalization ability for unseen working conditions The graph result data is input into the GNN model for the classification and recognition of the operating states of mechanical equipment under multi-working conditions. This graph structure effectively captures the local features under each working condition state within a single layer through node connections, and at the same time utilizes the cross-layer connection mechanism to achieve the deep integration of different state information. This design enables the model to fully learn the complex data characteristics in a multi-level and multi-working condition environment during the training process, thereby significantly improving the generalization ability for unseen working conditions

[0053] S3: Introduce a backdoor adjustment strategy and an approximate intervention method based on pseudo-environment labels to eliminate the interference of environmental factors

[0054] Furthermore, intervene in the environmental variable M by introducing the "do-operation". The do-operation can eliminate the interference of the environmental variable M on the graph feature so that the model only captures the stable causal relationship q between θ(Y|do(G)), and is no longer affected by the noise and instability caused by environmental changes. Compared with the traditional conditional probability q θ Unlike (Y|G), the do-operation essentially removes the model's reliance on non-causal correlations in out-of-distribution samples by intervening variables, thereby improving the model's robustness in the context of distribution shift. However, it is worth noting that although q is directly calculated through physical intervention θ (Y|do(G)) is an ideal choice, but in practical applications, this method is often difficult to implement due to experimental cost and resource limitations. Therefore, a backdoor adjustment strategy is introduced based on observational data.

[0055] The backdoor adjustment strategy includes intervening the environment variable M by introducing the do-operation, and the do-operation eliminates the effect of the environment variable M on the graph feature. The interference makes the model only capture Stable causal relationship q with label variable Y θ (Y|do(G)), and is no longer affected by the noise and instability caused by environmental changes. Compared with the traditional conditional probability q θ Unlike (Y|G), the do-operation essentially removes the model's reliance on non-causal correlations in out-of-distribution samples by intervening variables, thereby improving the model's robustness in the context of distribution shift. However, it is worth noting that although q is directly calculated through physical intervention θ (Y|do(G)) is an ideal choice, but in practical applications this approach is often difficult to implement due to experimental cost and resource limitations.

[0056] Based on the observed data, the backdoor adjustment strategy is introduced to approximate the causal intervention effect by solving it, which is expressed as:

[0057]

[0058] Among them, q0(M) represents the prior distribution of the environmental variable M;

[0059] Assume that the pseudo environment estimator p Ω (M|G), based on the self-graph features of the nodes To infer the pseudo environment variable mv of node v, the inferred pseudo environment m v and self-image features Input to the GNN predictor q θ (Y|G,M) is jointly optimized to obtain:

[0060]

[0061] in, is the supervised training loss, is the regularization loss. This optimization objective decouples the dependence between environmental information and graph features, enabling the model to capture environment-insensitive causal patterns, thereby maintaining strong generalization ability in out-of-distribution tasks and effectively coping with data distribution shifts in different environments.

[0062] It should be noted that the method of the present invention does not rely on explicit environmental labels in the data, nor does it assume prior knowledge of the physical meaning of unobserved environments, and at the same time does not require the pseudo-environment to accurately reflect the actual context. Therefore, the pseudo-environment is modeled as a latent variable in this model and is represented by the embedding vector of node v. Although the pseudo-environment does not directly correspond to the real world, the present invention expects these representations to have sufficient information capacity to support the model to effectively extract key patterns from the observed data, thereby capturing more robust causal relationships and enhancing its generalization performance in out-of-distribution tasks. The distribution shift of graph-structured data is often closely related to the complex connection patterns between nodes. The structural features in the ego-graph may contain stable patterns crucial for generalization. Therefore, to further expand the learning ability of the model, the representation of the pseudo-environment is generalized to the embedding representation that runs through each layer of the graph neural network (GNN).

[0063] Specifically, the pseudo-environment as a latent variable for each layer is inferred through the feature aggregation representation of node v and is defined in the form of a Z-dimensional numerical vector. The pseudo-environment estimator includes the pseudo-environment estimator p Ω (M|G), which infers the pseudo-environment representation during the feature aggregation process of each layer of the graph neural network. The pseudo-environment as a latent variable for each layer is inferred through the feature aggregation representation of node v and is defined in the form of a Z-dimensional numerical vector. To achieve accurate inference of the pseudo-environment, the categorical distribution is used as the sampling basis, and the model conditional probability is represented based on the node embedding of the current layer as follows:

[0064]

[0065] where W (l) represents the learnable weight matrix of the l-th layer. By introducing Gumbel noise α z , the sampling process can approximate discretization while maintaining continuity, which is expressed as:

[0066]

[0067] where α z is the random noise sampled from the Gumbel distribution, and λ is a hyperparameter.

[0068] When approaching 0, the sampling result approaches a discrete value; while when it is larger, the sampling shows smooth continuous values. Therefore, by appropriately adjusting the value, we can achieve smooth continuous optimization in practice while ensuring the effectiveness of the model in inferring the pseudo - environment.

[0069] It should also be noted that in the out - of - distribution generalization (OOD) task, the adaptability and robustness of the model are crucial when dealing with changes in complex environments. For this reason, an Adaptive Expert Ensemble Graph Neural Network (Adaptive Expert Ensemble GNN) predictor is proposed, which can encode the input self - subgraph under the condition of the inferred pseudo - environment label. This pseudo - environment label is obtained through the graph - structure - based environment estimator p Ω (M|G). To more finely capture the layer - by - layer environmental information, the model introduces a layer - by - layer adaptive update mechanism to control the propagation of multiple expert units, which are instantiated through two different model instances. Specifically, for the expert ensemble architecture based on the graph convolutional network, its layer - by - layer feature update can be expressed as:

[0070]

[0071] where δ v and δ u are the degrees of nodes v and u, while and W (l,z) are the linear transformation matrices of the neighbor - node information and self - node information of the z - th expert branch in the l - th layer respectively. The activation function is introduced for non - linear transformation to ensure that complex interactions between node features can be fully captured. This framework can be regarded as a causal representation form of GCN, which adaptively propagates by dynamically selecting Z convolutional filters to guide the model.

[0072] To further improve the robustness and flexibility of the model, an adaptive expert model based on the attention mechanism is proposed, aiming to model the pairwise interaction relationships between nodes. Its updated model is expressed as:

[0073]

[0074] where, represents the attention weight between nodes. Therefore, the model generates a high - dimensional embedded representation of node v through L - layer adaptive information transfer and maps it to the final prediction y v of the node through a fully - connected layer. In the entire architecture, the environmental information of the node is obtained through the hidden vector inferred by the layer - by - layer adaptive expert ensemble model.Capture and express them to dynamically control the transmission and learning process of features. This design not only ensures the effective integration of local neighbor node information, but also adaptively adjusts the weights and feature paths of each layer through the inferred pseudo-environment context, thus significantly enhancing the generalization ability of the model to cope with diverse data distributions. Further, the inferred pseudo-environment provides rich context information for the information propagation of each layer of the graph neural network, enabling the model to adjust the propagation mechanism according to different environmental features, thus showing stronger generalization and robustness under complex structured data distributions. This architecture can exhibit excellent performance by capturing stable relationships in the graph structure, especially in node-level prediction tasks. At the same time, in the face of diverse distribution shifts, the model can maintain a high degree of prediction accuracy and stability, providing a theoretical basis and technical support for coping with complex scenarios in practical applications.

[0075] S4: Generalize the data by combining the multi-layer pseudo-environment representation with the GCN or GAT encoder.

[0076] Furthermore, generalizing the data by combining the multi-layer pseudo-environment representation with the GCN or GAT encoder includes combining the GCN or GAT as the encoder backbone structure and using the feature extraction ability in the non-Euclidean space to improve the expression of the model for high-dimensional topological data.

[0077] Example 2, referring to Figure 2 and Figure 3 , which is an embodiment of the present invention, provides a method for out-of-distribution intelligent diagnosis of high-end mechanical equipment based on graph causal adaptation. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0078] First, to deeply study the adaptability and robustness of the model, two types of experimental schemes are designed. The first type of experiment aims to divide the ID and OOD data by selecting data with different rotational speeds under the same fault size (2×2 mm), which we call BUCEA-R. The second type of experiment focuses on analyzing the influence of different fault sizes under the condition of a fixed rotational speed (900 r / min), which is called BUCEA-S. A more detailed experimental design is shown in Table 1.

[0079] Table 1 Specific experimental scheme design

[0080]

[0081]

[0082] To effectively simulate the complexity of data distribution drift in mechanical equipment during actual operation, a mixed dataset containing different rotational speeds and fault sizes was adopted in the experimental design. Specifically, the model training stage of each dataset included data in four different health states, all of which were ID (In-Distribution) data; in the testing stage, the dataset introduced a health state data that did not appear in the training stage, namely OOD (Out-Of-Distribution) data, aiming to evaluate the model's ability to identify and detect OOD data. In addition, to comprehensively verify the reliability and robustness of the model, five different health state data were used as OOD samples, and multiple rounds of training and testing were carried out respectively. By conducting experiments on different OOD detection benchmark methods, the superiority performance of the OOD detection method based on graph causal intervention was further analyzed and compared.

[0083] In the process of dataset selection and division, in order to extract the feature information of each sample, the vibration signal was cut by a sliding window with a length of 1024, and there was no overlap between adjacent time windows. The total signal length of each working condition was 102400, so the data of each working condition would be divided into 100 nodes. To enhance the generalization ability of the model, all ID node samples were randomly shuffled. Next, the ID data nodes were randomly divided into a training set, a validation set and a test set, and the division ratios were 60%, 10% and 30% respectively. Finally, all experiments adopted a unified learning strategy and experimental environment, where the learning rate was set to 0.001, the number of training epochs was 300, and all models were implemented based on the PyTorch framework to ensure the consistency and comparability of the experiments.

[0084] On the BUCEA-R dataset, the present invention investigated the influence of rotational speed change on the model performance under the same fault size (2×2mm), and the experimental results are as Figure 2(as shown in (a)). Research shows that when the ID data contains high - speed samples and the OOD data are low - speed samples, using GCN or GAT as the backbone model has a significantly higher classification accuracy on the OOD data than in the scenarios of low - speed ID data and high - speed OOD data. This phenomenon reveals that high - speed samples usually contain stronger and more significant vibration signal features. During the training process, the model can extract clearer fault patterns and capture high - frequency signal features. Therefore, even when facing low - speed OOD data in the test stage, the model can still identify potential fault features, thus achieving a relatively high classification accuracy. On the contrary, when the ID data mainly consists of low - speed samples, the model cannot comprehensively capture the diverse fault feature distributions during the training stage because the vibration signals of low - speed samples are weak and lack universality. In this case, the complex fault patterns and high - frequency vibration features contained in the high - speed OOD data exceed the feature distributions learned during model training, making it difficult for the model to adapt when facing distribution shift, resulting in a significant performance degradation. This result indicates that high - speed samples have important advantages in the model training process and emphasizes the importance of considering the intensity of sample vibration signal features in out - of - distribution generalization tasks.

[0085] The BUCEA - S experiment further reveals the impact of damage size changes on the model's generalization performance. Specifically, when the ID data contains damaged samples with larger sizes, the model's recognition ability on small - size damaged OOD data is significantly improved, and the experimental results are as Figure 2 (shown in (b)). Larger - size damages are usually accompanied by more obvious fault features, enabling the model to learn clear fault patterns during the training stage and maintain a relatively high recognition accuracy during the test stage. These two sets of experimental results together indicate that the intensity of the vibration signal and the significance of the fault features in the data play a crucial role in the model's generalization performance. High - intensity vibration signals and significant fault patterns can enhance the model's ability to capture key features, thereby improving the recognition performance of out - of - distribution samples.

[0086] A variety of classic and up-to-date out-of-distribution generalization methods were systematically compared. The experiments covered four vibration signal datasets (CWRU, JNU, BUCEA-R, and BUCEA-S), and two mainstream graph neural network (GNN) architectures, namely GCN and GAT, were respectively selected as the encoder backbones. The experimental results are shown in Table 2. The comparison methods included the classic empirical risk minimization (ERM) as the baseline, which uses a standard supervised loss function without adjusting for distribution drift. In addition, three mainstream OOD generalization models designed for independent samples were selected: IRM (Invariant Risk Minimization), DANN (Domain-Adversarial Neural Network), and Mixup. To adapt to the characteristics of graph-structured data, SRGNN was also introduced for comparison experiments. For a fair comparison, these methods were all experimented with GCN and GAT as the encoder backbones.

[0087] Table 2 Recognition accuracies of different algorithms for OOD data on four datasets

[0088]

[0089] The experimental results show that GCI-ODG (solving the out-of-distribution generalization problem based on graph causal inference) achieved the optimal OOD classification accuracy on all test datasets and two encoder architectures, significantly outperforming the existing mainstream methods. This excellent performance is mainly attributed to the application of causal inference and pseudo-environment label estimation strategies, which successfully eliminated the environmental confounding factors in the training data, enabling the model to robustly capture the stable feature relationships across working conditions. At the same time, on the BUCEA-R and BUCEA-S datasets with high heterogeneity, the OOD accuracies of the GCI-ODG method under the GAT architecture were 92.35% and 91.73% respectively, significantly better than the SRGNN and Mixup methods. Considering that these two datasets contain highly heterogeneous fault features and significant distribution changes, GCI-ODG effectively captured the key feature patterns under different working conditions through adaptive pseudo-environment label estimation, achieving efficient modeling and recognition of distribution drift data.

[0090] Meanwhile, when longitudinally analyzing the performance of different encoder architectures, it can be found that when GAT is selected as the backbone, the performance of all algorithms is generally better than that of GCN. This result can be attributed to the adaptive attention mechanism introduced by the GAT model. GCN uses graph convolution operations with fixed weights and cannot flexibly distinguish important features from noise features. Especially in low-speed samples, it is difficult to effectively capture key vibration signals, resulting in weak recognition ability for high-speed OOD data. In contrast, the GAT model adaptively adjusts the weights between nodes, giving priority to high-frequency vibration signals and significant fault patterns, significantly enhancing the adaptability and generalization ability of the model in complex environments.

[0091] Example 3, referring to Figure 3 and Figure 4 , which is an embodiment of the present invention, provides a graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment. To verify the beneficial effects of the present invention, scientific demonstrations are carried out through economic benefit calculations and simulation experiments.

[0092] The present invention proposes an intelligent diagnosis framework based on graph-causal intervention (referred to as GCI-ODG) for the out-of-distribution generalization problem in wind turbines.

[0093] In the present invention, to deeply evaluate the key contributions of each module in the proposed method, a series of ablation experiments are designed and carried out. The experimental results are as Figure 3 shown. The ablation experiments include three model configurations: First, the regularization loss term is removed, and only the standard supervised loss is relied on for training, which is called GCI-ODG-A; Second, the pseudo-environment representation of each layer is replaced with a globally shared single representation m v . At this time, the model degenerates into a simplified version GCI-ODG-B, losing the ability of hierarchical adaptive adjustment; Finally, the trainable environment estimator is replaced with a non-parametric mean pooling operation, that is, the mean pooling process is performed on the Z propagation branches of each layer to form GCI-ODG-C, thus removing the adaptive environment estimation mechanism.

[0094] The experimental results show that the GCI-ODG intelligent diagnosis framework significantly outperforms all ablation versions under all test datasets, verifying the scientificity and effectiveness of the method design. It is worth noting that on the BUCEA dataset, the performance of GCI-ODG-A significantly decreases, revealing the important role of the regularization loss term in highly heterogeneous and significantly distribution-shifted scenarios. The regularization term successfully suppresses the interference of environmental confounding bias by introducing the independence constraint between the pseudo-environment labels and graph features, promoting the model to learn stable causal relationships insensitive to the environment. The absence of this mechanism makes the model more dependent on the unstable features in the training data, resulting in obvious performance degradation on the BUCEA dataset with a large distribution shift, highlighting the indispensability of the regularization constraint for improving generalization ability. In contrast, the performance degradation of GCI-ODG-B further reveals the core value of the hierarchical pseudo-environment representation. The globally shared pseudo-environment representation m v cannot capture the complex changes between features at each level, resulting in obvious lack of flexibility in the model when aggregating multi-level features. The hierarchical adaptive pseudo-environment representation can dynamically perceive and adjust the environmental feature changes at different layers, effectively improving the generalization performance of the model on highly heterogeneous data and showing stronger adaptability and stability. In addition, in GCI-ODG-C, the adaptive environment estimator is replaced by a mean pooling operation. Although the computational complexity is simplified, the cost is a significant weakening of the model's classification ability on OOD data. Mean pooling ignores the feature differences between nodes and cannot flexibly adjust the propagation path and weight allocation. In contrast, the adaptive environment estimator can dynamically adjust the weights according to the importance of node features, effectively capturing key fault features and high-frequency vibration signals, thereby enhancing the robustness and resilience of the model under complex working conditions. The three work together to jointly construct a complete GCI-ODG framework, enabling it to exhibit excellent performance in out-of-distribution generalization tasks.

[0095] To deeply explore the influence of key hyperparameters in the model, the present invention respectively conducts a systematic analysis on the number of propagation branches Z and the Gumbel-Softmax parameter α z to reveal their mechanism of action on out-of-distribution generalization performance. The experiment is carried out on four datasets, verifying the sensitivity and influence trend of hyperparameter selection on the model performance under different environmental changes. The specific experimental results are as Figure 4 shown.

[0096] First, the analysis of the number of propagation branches \(Z\) shows that the model performance exhibits a significant non - monotonic change trend with respect to the value of \(Z\). When \(Z\) is small, the number of branches of the pseudo - environment estimator is insufficient to fully capture the diverse features in a complex environment, resulting in limited environmental perception ability of the model and showing weak generalization ability. As \(Z\) increases, the model performance gradually improves and reaches the best state when \(Z = 4\). At this time, the pseudo - environment estimator achieves a good balance in capturing key feature patterns and adjusting the propagation path, significantly enhancing the robustness of the model on heterogeneous data. However, further increasing the value of \(Z\) leads to a decline in model performance, especially on the BUCEA dataset with complex distribution, and this trend is particularly obvious. This phenomenon is attributed to the fact that too many branches introduce computational redundancy and feature noise, reducing the model's adaptability to environmental changes and degrading the generalization performance accordingly.

[0097] When analyzing the influence of the Gumbel - Softmax hyperparameter \(\lambda\), the experimental results also reveal the significant impact of the hyperparameter on the smoothness of the sampling distribution and the model performance. When the value of \(\lambda\) is small, the sampling process tends to be discretized, which helps the model capture more fine - grained features and improve the perception ability of key patterns in a complex environment. However, too low a value of \(\lambda\) may lead to an increase in the uncertainty of sample selection, causing large fluctuations in the model performance under different working conditions. As \(\lambda\) increases to a moderate level, an optimal balance is achieved between the smoothness and sharpness of the sampling results, enabling the model to maintain information richness while avoiding the instability caused by excessive discretization when capturing environmental features, significantly enhancing the generalization ability and robustness of the model.

[0098] It is worth noting that when the value of \(\lambda\) continues to increase, the model performance significantly decreases. A higher hyperparameter makes the Gumbel - Softmax output tend to an overly smooth uniform distribution, weakening the model's ability to distinguish key feature patterns. In this case, the model performs particularly poorly on the JNU and BUCEA datasets with high heterogeneity, indicating that excessive sampling smoothness will cause the model to lose its sensitive perception of environmental changes, thus reducing its robustness in a complex environment. Therefore, a reasonable selection of hyperparameters is crucial for improving the generalization ability of the model. When selecting moderate values of \(Z\) and \(\lambda\), a balance can be achieved between feature expression and computational complexity, significantly enhancing the environmental adaptability and robustness of the model.

[0099] Example 3, an embodiment of the present invention, provides a graph - causal adaptive out - of - distribution intelligent diagnosis system for high - end mechanical equipment, including an experimental detection module, a layer construction module, an interference elimination module, and a combination optimization module.

[0100] Among them, the experimental detection module is used to collect vibration acceleration signals in typical fault states through typical fault vibration test experiments; the layer construction module is used to construct a multi-layer graph structure through Euclidean distance and cosine similarity, connect nodes using K-nearest neighbors, and use it as the input of the graph neural network for time series signal analysis; the interference elimination module is used to introduce a backdoor adjustment strategy and an approximate intervention method based on pseudo-environment labels to eliminate the interference of environmental factors; the combined optimization module is used to generalize data by combining multi-layer pseudo-environment representations with GCN or GAT encoders.

[0101] If a function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0102] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or in combination with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device.

[0103] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection (electronic device) having one or more wirings, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which a program can be printed, as the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or, if necessary, other suitable processing, and then stored in a computer memory.

[0104] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, the multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having suitable combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), and the like. It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

[0105] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A graph-causality adaptive intelligent diagnosis method for high-end mechanical equipment in the distribution outside, characterized in that, Including: Collect vibration acceleration signals in typical fault states through typical fault vibration test experiments; Construct a multi-layer graph structure through Euclidean distance and cosine similarity, use K-nearest neighbors to connect nodes, and perform time-series signal analysis as the input of the graph neural network; Introduce a backdoor adjustment strategy and an approximate intervention method based on pseudo-environment labels to eliminate the interference of environmental factors; Generalize data through multi-layer pseudo-environment representation combined with GCN or GAT encoders.

2. The distributed external intelligent diagnosis method based on graph causality self-adaptation for high-end mechanical equipment according to claim 1, characterized in that: The collection of vibration acceleration signals in typical fault states includes conducting vibration test experiments to collect vibration acceleration signals in typical fault states. The single-state signals include inner-race fault (IF), outer-race fault (OF), rolling-element fault (BF), and normal state (NA). The compound fault states include outer-race + rolling-element fault and inner-race + rolling-element fault. The vibration signals in each fault state are tested with five different-sized grooves.

3. The distributed external intelligent diagnosis method for high-end mechanical equipment based on graph causality adaptation according to claim 2, wherein: The construction of the multi-layer graph structure by Euclidean distance and cosine similarity includes setting the rotational speed and loss magnitude state ω of the mechanical equipment in n different domains n (n = 1, 2, …, n), and the time series data collected under each state is where k = 1, 2, …, K represents K working condition states in the nth domain, and the time series of each working condition state is marked as the label set y (n) ∈{1, 2, …, K}. Define the time window length as Δt, and divide each time series into non-overlapping subsequences, where L represents the length of the original time series. Therefore, in the node generation stage, each subsequence is mapped to a node k The node set is represented as and multi-dimensional time domain features are extracted to generate its feature vector to characterize the local dynamic behavior characteristics under a single working condition. In each domain layer ω within, local connection relationships are constructed through similarity measurement between nodes. Given nodes n and and the similarity is measured by calculating the Euclidean distance, which is expressed as: Connect each node to its nearest neighbor in a K-nearest neighbor manner to form an intra-layer graph structure After constructing a single-layer graph cross-layer connections are defined to capture the global behavior patterns under multi-condition scenarios. For any two nodes from different layers and the similarity between cross-layer nodes is measured by cosine similarity and expressed as: When the remainder similarity is greater than the preset threshold η, that is, introduce cross-layer edge connections between and to integrate the working condition characteristics in different domain states, construct a graph structure with multi-level dependencies, and form the graph structure data input into the GNN model jointly by the adjacency matrix and the feature matrix ​ 4. The method for distributed external intelligent diagnosis of high-end mechanical equipment based on graph causality adaptation according to claim 3, characterized in that: The backdoor adjustment strategy includes intervening in the environmental variable M through the introduction of the do-operator. The do-operator eliminates the interference of the environmental variable M on the graph features so that the model only captures the stable causal relationship q θ (Y|do(G)) between the label variable Y. Based on the observed data, a backdoor adjustment strategy is introduced, and the causal intervention effect is approximated by solving the following formula: Among them, q0(M) represents the prior distribution of the environmental variable M; Set the pseudo-environment estimator p Ω (M|G), according to the self-graph features of the nodes to infer the pseudo-environment variable m of node v v , and input the inferred pseudo-environment m v and the self-graph features into the GNN predictor q θ (Y|G,M) for joint optimization to obtain: Among them, is the supervised training loss, is the regularization loss.

5. The distribution external intelligent diagnosis method for high-end mechanical equipment based on graph causality self-adaptation according to claim 4, wherein: The pseudo-environment estimator includes a pseudo-environment estimator p Ω (M|G), which infers a pseudo-environment representation during the feature aggregation process of each layer of the graph neural network. The pseudo-environment serves as a latent variable for each layer, is inferred through the feature aggregation representation of node v, and is defined in the form of a Z-dimensional numerical vector, using a categorical distribution as the sampling basis, and the model conditional probability is represented based on the node embedding of the current layer as follows: Among them, W (l) represents the learnable weight matrix of the l-th layer. By introducing Gumbel noise α z , while keeping the sampling process continuous and approximately discretized, it is expressed as: where α z is the random noise sampled from the Gumbel distribution, and λ is the hyperparameter.

6. The distributed external intelligent diagnosis method for high-end mechanical equipment based on graph causality adaptation according to claim 5, characterized in that: The GNN predictor includes an expert ensemble architecture based on a graph convolutional network, and the layer-by-layer feature update is expressed as: where, δ v and δ u are the degrees of nodes v and u, and W (l,z) are the linear transformation matrices of the neighbor node information and the self-node information of the z-th expert branch in the l-th layer, respectively, is the activation function; Construct an adaptive expert model based on the attention mechanism, and construct the pairwise interaction relationship between modeling nodes, expressed as: Among them, represents the attention weight between nodes.

7. The distribution external intelligent diagnosis method for high-end mechanical equipment based on graph causality adaptation according to claim 6, wherein: The generalization of data through multi-layer pseudo-environment representation combined with GCN or GAT encoders includes combining GCN or GAT as the encoder backbone structure, and using the feature extraction ability in the non-Euclidean space to improve the model's expression of high-dimensional topological data.

8. A system adopting the graph-causality adaptive distribution external intelligent diagnosis method for high-end mechanical equipment as described in any one of claims 1 to 7, characterized in that: Including an experimental detection module, a layer construction module, an interference elimination module, and a combination optimization module; The experimental detection module is used to collect vibration acceleration signals in typical fault states through typical fault vibration test experiments; The layer construction module is used to construct a multi-layer graph structure through Euclidean distance and cosine similarity, use K-nearest neighbors to connect nodes, and perform time-series signal analysis as the input of the graph neural network; The interference elimination module is used to introduce a backdoor adjustment strategy and an approximate intervention method based on pseudo-environment labels to eliminate the interference of environmental factors; The combination optimization module is used to generalize data through multi-layer pseudo-environment representation combined with GCN or GAT encoders.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the graph-causal adaptive out-of-distribution intelligent diagnosis method for high-end mechanical equipment according to any one of claims 1 to 7.

Images