Adaptive, extra-distributed intelligent diagnostic method based on graphene causality for high-quality mechanical devices

Description

Y20P5BE-YUV005BEL02.04.2025 2 which leads to a limited generalizability of the model in real-world applications, particularly in a complex and changing application scenario such as mechanical devices, where existing solutions are confronted with the problem of model overfitting or sensitivity to a particular environment in real-world applications.5 CONTENT OF THE PRESENT INVENTION In consideration of the above problems, the present invention is made available. Therefore, the technical problem addressed by the present invention is that extra-distributed intelligent diagnostic methods for high-quality mechanical devices suffer from a lack of effective extrapolation capability, the absence of an effective strategy to cope with unlabeled or unknown environmental distributions, and the problem of how to equip the graphical neural network with a stronger generalization capability in complex environments and improve its robustness to data outside the distribution.while avoiding the problem of overfitting or oversensitivity to certain environments. To solve the above problems, the present invention provides the following technical solution: an adaptive, extra-distributed intelligent diagnostic method based on graph causality for high-quality mechanical devices, characterized in that it comprises: collecting vibration acceleration signals under typical fault conditions by means of a vibration test experiment of typical faults; creating a multi-layered graph structure by means of Euclidean spacing and cosine similarity, connecting nodes using K-neighbors and analyzing time-series signals as input for a graphical neural network; introducing a "backdoor" adaptation strategy and an approximate intervention method based on labels of pseudo-environments,to eliminate interference from environmental factors; generalizing the data through a multi-layered representation of the pseudo-environment in combination with a GCN or GAT encoder. As an advantageous solution of the adaptive, extra-distributed intelligent diagnostic procedure 30 based on graphene causality for high-quality mechanical devices, wherein the BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 3 Collecting vibration acceleration signals under typical fault conditions, the collection of vibration acceleration signals under typical fault conditions by performing vibration test experiments, wherein a single-state signal comprises an inner ring fault IF, an outer ring fault OF, a rolling element fault BF and a normal state NA, and whereas a composite fault condition comprises an outer ring fault and a rolling element fault, an inner ring fault and a rolling element fault, and wherein the vibration signal in each fault condition is tested using five different sizes of recesses. As an advantageous solution of the adaptive,extra-distributed intelligent diagnostic procedure based on graph causality for high-quality mechanical devices, wherein the creation of a multi-layered graph structure using Euclidean spacing and cosine similarity, the setting of the time series data collected by the mechanical device at different domain speeds and loss quantity states (=1,2,⋯,) as () , where =1,2,⋯, operating condition states in K under the nth domain, where the time series of each operating condition state is denoted as a set of labels () ∈ {1,2,⋯,}, where the time window length is defined as , and where each time series () is subdivided into non-overlapping subseries, where L represents the original time series length, so that in the node generation phase each subseries is assigned to a node, () is mapped, where the node set is represented as =, ,, ,⋯, , and where multidimensional 20 time domain features are extracted to generate the feature vector h, ().which maps the local dynamic behavioral properties under a single operating condition, where within each domain layer local connection relationships are created by similarity measures between the nodes, where the nodes are given, () and, () , where the similarity is calculated by calculating the Euclidean distance25, which is represented as follows: (, () ,, () )=‖, () −, () ‖ where each node is connected to its nearest neighbor node by K-neighbors, BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 4 to form a layer-internal graph structure () = () ,ℰ (); where, after creating the single-layer graph (), the cross-layer connections are defined to capture a global behavioral model under different operating conditions, where for any two nodes, () ∈ and, () ∈ from different layers, the similarity of the cross-layer nodes is measured by the cosine similarity, which is represented as follows: cos_sim, () ,, () = , () ⋅, () ‖, () ‖‖, () ‖ where,If the cosine similarity is greater than a preset threshold, i.e., a cross-layer edge connection is introduced between, () and, (), the features of the operating conditions in different domain states are integrated, 10 a graph structure with multilayer dependencies is created, and an adjacency matrix and a feature matrix, created by the layer-intralayer and cross-layer connections, together form the graph structure data = (,) that are fed into the GNN model. As an advantageous solution of the adaptive, extra-distributed intelligent diagnostic procedure 15 based on graph causality for high-value mechanical devices, wherein the "backdoor" adaptation strategy includes an intervention on an environment variable M by introducing a "do" operation, where the "do" operation removes the disturbance of the environment variable M on the graph feature,so that the model captures only a stable causal relationship |() between and the identifier variable Y20; where a “backdoor” adjustment strategy is introduced based on the observed data to approximate the causal intervention effect by solving the following formula: (|())=()[(|,)]25 where (() represents the previous distribution of the environment variable M; where a pseudo-environment estimator is set as (|) such that the BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 5 pseudo-environment variable mv of the node is derived on the basis of the node's self-graph feature, where the derived pseudo-environment variable mv and the self-graph feature are fed into the GNN predictor Qg to perform the joint optimization, which is represented as follows: (|do())≥(|)[(|,)]−((|)∥())5 where ℒ is the monitoring training loss and ℒ is the regularization loss. As an advantageous solution of the adaptive, extra-distributed intelligent diagnostic procedure based on graphene causality for high-quality mechanical devices,where the pseudo-environment estimator comprises a pseudo-environment estimator(|) that derives a pseudo-environment representation during the aggregation process of the features of each layer in the graphical neural network, and where the pseudo-environment () serves as a potential variable in each layer, where the aggregation representation of the features of node v serves for the derivation, where a definition in the form of a Z-dimensional numerical vector is made, where the category distribution () is used as the sampling basis, and the model-related probability () on the basis of the node embedding () on the current layer is represented as follows: () = () () where () represents the learnable weight matrix of the l-en layer, where the sampling process can be discretized while maintaining continuity by Gumbel noise is introduced, which is represented as follows: 20 () = ( () +) / ∑ ∑( () ) / Gumbel where the random noise is obtained by sampling from the Gumbel distribution,and a hyperparameter. As an advantageous solution of the adaptive, extra-distributed intelligent diagnostic procedure based on graph causality for high-quality mechanical devices, wherein the 25 GNN predictor comprises an expert integration architecture based on graph convolutional networks, where the layer-wise feature update is represented as follows BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 6: () =, () 1 ∈ () (), () () + (), () () where and are the degrees of the nodes v and u, (), and (), are the linear transformation matrices of the neighbor node information and the self-node information of the s-th expert branch of the l-th layer, respectively, and the 5 activation function is; where an adaptive expert model based on the attentional mechanism is created to generate the pairwise interaction relationship between modeled nodes, represented as follows: () =, () (,) ∈ () (,) () + (,) () 10 where (,) represents the attentional weight between nodes. As an advantageous solution of the adaptive,extra-distributed intelligent diagnostic procedure based on graph causality for high-quality mechanical devices, wherein generalizing data by a multi-layered representation of the pseudo-environment in combination with a GCN or GAT encoder includes combining GCN or GAT as the main structure of the encoder, using feature extraction capability in a non-Euclidean space, and improving model expression for high-dimensional topological data. Another purpose of the present invention is to provide an adaptive, extra-distributed intelligent diagnostic system based on graph causality for high-value mechanical devices, which is capable of proposing a new learning objective through the principle of causal intervention, and which is capable of effectively eliminating environmental disturbances in the data by deriving pseudo-environmental information without requiring environmental labels. It can help the model learn a stable predictive relationship that solves the problem.that the 25 current extra-distributed intelligent diagnostic procedures for high-quality mechanical devices do not contain effective strategies for dealing with unlabeled or unknown environmental distributions. BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 7 As an advantageous solution of an adaptive, extra-distributed intelligent diagnostic system based on graph causality for high-quality mechanical devices, comprising an experimental detection module, a layer creation module, a fault elimination module, and a combination optimization module, wherein the experimental detection module serves to collect vibration acceleration signals under a typical fault condition by means of a vibration test experiment of typical faults, wherein the layer creation module serves to create a multi-layered graph structure by means of Euclidean spacing and cosine similarity, to connect nodes using K-neighbors, and to analyze time-series signals as input for a graphical neural network,wherein the interference elimination module serves to introduce a “backdoor” adaptation strategy and an approximate intervention method based on pseudo-environment labels to eliminate interference caused by environmental factors, wherein the combination optimization module serves to generalize data by a multi-layered representation of the pseudo-environment in combination with a GCN or GAT encoder.15 The present invention further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the processor executes the computer program that implements the steps of the adaptive, extra-distributed intelligent diagnostic procedure based on graph causality for high-quality mechanical devices.20 The present invention further provides a computer-readable storage medium on which a computer program is stored, wherein the computer program, when read by is executed by a processor, the steps of the adaptive,extra-distributed intelligent diagnostic procedure based on graph causality for high-quality mechanical devices implemented.25 The present invention has the following advantageous effects: The present invention provides an adaptive, extra-distributed intelligent diagnostic procedure based on graph causality for high-quality mechanical devices, which is based on the intrinsic complexity of graph-structured data, uses pseudo-environmental markers as latent intermediate variables, and achieves robust 30 BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 8 adaptation to different operating conditions by dynamically adjusting the propagation paths for node features. Based on the theory of graph causality inference, an estimator for pseudo-environmental markers was developed,to dynamically infer hidden pseudo-environmental variables from the observed node features using Gumbel-Softmax with microsampling and thus eliminate the disturbance effect caused by environmental changes. The invention leads to better results in terms of applicability, stability, and generalizability. BRIEF DESCRIPTION OF THE DRAWINGS 10 To illustrate the technical solutions of the embodiments of the present utility model more clearly, the attached drawings used in the description of the embodiment examples are briefly described below. Of course, the attached drawings in the following description are only some of the embodiments of the present utility model, and for a person skilled in the art, 15 other attached drawings can be created on the basis of these drawings without creative work. Fig. 1 shows a general flow diagram of an adaptive,extra-distributed intelligent diagnostic procedure based on graph causality for high-quality mechanical devices in a first embodiment of the present invention; 20 Fig. 2 is a diagram of the experimental results of the extra-distributed generalization performance of GCI-ODG onto two data sets of an adaptive, extra-distributed intelligent diagnostic procedure based on graph causality for high-quality mechanical devices in a second embodiment of the present invention; 25 Fig. 3 showed a diagram of the ablation trial study and performance evaluation of a GCI-ODG model of an adaptive, extra-distributed intelligent diagnostic procedure based on graph causality for high-quality mechanical devices in a second embodiment of the present invention; Fig. 4 showed a diagram of the performance analysis of hyperparameter Z and GCI-ODG of a 30 GCI-ODG model of an adaptive,extra-distributed intelligent diagnostic procedure on the BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 9 Basis of graphene causality for high-quality mechanical devices in a second embodiment of the present invention. DETAILED DESCRIPTION To clarify the purpose, technical solutions, and advantages of the present utility model, the technical solutions in the embodiments of the present utility model are described clearly and completely below in conjunction with the attached drawings in the embodiments of the present utility model. Naturally, the described embodiments are only a part of the embodiments of the present utility model, but not all embodiments. Based on the embodiments of the present utility model, all other embodiments that a person skilled in the art can produce without creative work fall within the scope of protection of the present utility model. Embodiment 115 As shown in Fig. 1,is an embodiment of the present invention, which provides an adaptive, extra-distributed intelligent diagnostic method based on graphene causality for high-quality mechanical devices, comprising the following: S1: Collecting vibration acceleration signals under typical fault conditions20 by means of a vibration test experiment of typical faults; Furthermore, the vibration acceleration signals under typical fault conditions comprise the following: four single fault conditions, including an inner ring fault IF, an outer ring fault OF, a rolling element fault BF, and a normal condition NA; and two compound fault conditions, including an outer ring and a rolling element fault, and an inner ring and a rolling element fault, for a total of six fault conditions. The vibration signals in each fault condition are tested using five different sizes of cutouts. These cutouts were produced by wire cutting with the following specific sizes: 0.4 × 1 mm (width 0.4 mm, depth 1 mm), 2 × 2 mm, 2,8×3mm, 3, 4×4 and 4×4mm. To simulate environmental changes under real operating conditions, the vibration signals of the defective bearings were recorded during the experiment at five different speeds (500 rpm, 700 rpm, 900 rpm, 1000 rpm and 1100 rpm). S2: Creating a multilayer graph structure using Euclidean spacing and cosine similarity, connecting nodes using K-neighbors and analyzing 5 time series signals as input for a graphical neural network; Furthermore, in the field of intelligent diagnostics, particularly in the detection and classification of the operating states of high-quality mechanical devices, the data under different operating conditions are complex and highly dimensional. To meet this challenge, the present invention develops a graphical 10-neural network (GNN) based on a multi-layered graph structure. Through a structured representation of time-series data, the method aims toto integrate global information and simultaneously capture local features, thereby improving the generalizability of the model. The creation of a multi-layered graph structure using Euclidean spacing and cosine similarity involves setting 15 of the time series data collected by the mechanical device at different domain speeds and loss quantity states (=1,2,⋯,) as () , where =1,2,⋯, operating condition states in Kunter dern-th domain, where the time series of each operating condition state is denoted as a set of labels () ∈ {1,2,⋯,}, where the time window length is defined as 20, and where each time series () is subdivided into non-overlapping subseries, where L represents the original time series length, so that In the node generation phase, each subseries is mapped to a node, (), where the node set is represented as =, ,, ,⋯, , and where multidimensional time domain features are extracted to generate the feature vector h, ().which maps the 25 local dynamic behavioral properties under a single operating condition, where within each domain layer local connection relationships are created by similarity measures between the nodes, where the nodes, () and, () BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 11 are given, where the similarity is calculated by calculating the Euclidean distance, which is represented as follows: (, () ,, () )=‖, () −, () ‖ where each node is connected to its nearest neighbor node by K-neighbors to form a layer-internal graph structure () = () ,ℰ ();5 where after creating the single-layer graph () the layer-spanning Connections are defined to capture a global behavioral model under different operating conditions, where for any two nodes, () ∈ and, () ∈ from different layers, the similarity of the nodes across layers is measured by the cosine similarity, which is represented as follows: 10 cos_sim, () ,, () = , () ⋅, () ‖, () ‖‖, () ‖ where,If the cosine similarity is greater than a preset threshold, i.e., a cross-layer edge connection is introduced between, () and, (), the features of the operating conditions in different domain states are integrated, a graph structure with multilayer dependencies is created, and an adjacency matrix and a feature matrix, created by the layer-internal and cross-layer connections, together form the graph structure data = (,) that are fed into the GNN model. The result data of this graph are fed into the GNN model for the classification and detection of the operating states of mechanical devices under different operating conditions. The graph structure effectively captures the local features of each operating state through node connections within a single layer and simultaneously uses a cross-layer connection mechanism to achieve deep integration of different To obtain status information. This designer enabled the model toDuring the training process, the complex data properties in environments with multiple layers and multiple conditions are fully learned, which significantly improves the generalizability to unknown conditions. S3: Introducing a "backdoor" fitting strategy and an approximate intervention method based on labels of pseudo-environments to eliminate disturbance from environmental factors; furthermore, the introduction of the "do" operation enables the intervention of the environment variable M. The "do" operation eliminates the influence of the environment variable M on the graph feature, so that the model only captures the stable causal relationship |() between and the label variable Y and is no longer affected by noise and instability.which are caused by changes in the environment. In contrast to traditional conditional probability (|), the "do" operation essentially frees the model from its dependence on non-causal correlations in extra-distributed samples by intervening in the variables, thereby improving the robustness of the model in situations with distributional changes. However, it should be noted that while it would be ideal to compute (|()) directly through physical intervention, this approach is often difficult to implement in practice due to cost and resource constraints for experiments. Therefore, a "backdoor" fitting strategy (BDA) based on observational data is introduced. The "backdoor" adaptation strategy includes an intervention for an environment variable M by introducing a "do" operation; the "do" operation eliminates the influence of the environment variable M on the graph feature.so that the model only captures the stable causal relationship |() between and the identifier variable Y20 and is no longer affected by the noise and instability caused by environmental changes. In contrast to traditional conditional probability (|), the "do" operation essentially frees the model from its dependence on non-causal correlations in extra-distributed samples by intervening in the variables, thereby improving the robustness of the model in situations with 25 distribution changes. However, it should be noted that while it would be ideal to compute (|()) directly through physical intervention, this approach is often difficult to implement in practice due to cost and resource constraints for experiments. Therefore, a “backdoor” adjustment strategy is introduced based on the observed data 30 BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 13 to approximate the causal intervention effect by solving the following formula: (|())=()[(|,)] where () represents the previous distribution of the environment variable M; where a pseudo-environment estimator is set to (|) such that the node's pseudo-environment variable mv is derived on the basis of the node's self-graph feature, where the derived pseudo-environment variable mv and the self-graph feature are fed into the GNN predictor Qg to perform the joint optimization, which is represented as follows: (|do()) ≥ (|)[(|,)] − ((|) ∥()) 10 where ℒ is the monitoring training loss and ℒ is the regularization loss. This optimization goal enables the model to capture environment-independent causality patterns by decoupling the dependence of environment information on graph features, thus maintaining strong generalization capabilities for extra-distributed tasks and effectively handling the variation in data distribution across 15 different environments. It should be noted thatthat the method of the present invention does not rely on an explicit environment label in the data and also does not require any a priori knowledge of the physical meaning of unobserved environments, while the pseudo-environment need not exactly reproduce what actually follows. Therefore, the 20 pseudo-environments in the present model are modeled as potential variables represented by embedding vectors of the node v. Although the pseudo-environments do not directly correspond to the real world, the present invention expects that these representations will have sufficient information capacity to assist the model in the effective extraction of key patterns from the observed data, 25 thereby capturing more robust causal relationships and improving the generalizability for extra-distributed tasks. The distributional variance of graph-structured data is often closely related to the complex connection patterns between the Node-connected, and structural features in self-graphs can contain stable patterns,which are crucial for generalization. In order to further expand the learning capability of the model, the pseudo-environment representation is therefore generalized to an embedded representation in all layers of the graphical neural network (GNN). In particular, the pseudo-environment () serves as a potential variable in each layer, where the aggregation representation of the features of node v serves for derivation, with a definition in the form of a Z-dimensional numerical vector being made.5 The pseudo-environment estimator comprises a pseudo-environment estimator (|) that derives a pseudo-environment representation during the aggregation process of the features of each layer in the graphical neural network, and the pseudo-environment () serves as a potential variable in each layer, where the aggregation representation of the features of node v serves for derivation, with a definition in the form of a 10 Z-dimensional numerical vector being made, where, in order to achieve an exact derivation of the pseudo-environment,the category distribution () is used as the sampling basis and the model-related probability () based on the node embedding () on the current layer is represented as follows: () = () ()15 where () represents the learnable weight matrix of the l-en layer, where the sampling process can be discretized while maintaining continuity by introducing Gumbel noise, which is represented as follows: () = ( () +) / ∑ ∑( () ) / Gumbel where is the random noise obtained from the Gumbel distribution by sampling, and is a hyperparameter. When converging towards 0, the sampling results converge to discrete values, while at larger values ​​the sampling behaves like a smoothed continuum. By appropriately adjusting the values, we can therefore achieve a smooth continuous optimization in practice and simultaneously ensure the validity of the model in the derivation of the pseudo-environment. It should also be noted thatthat in the extra-distributed generalized OOD task, BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 15 model fitting and robustness are of crucial importance to cope with changes in complex environments. To this end, an Adaptive Expert Ensemble (GNN) predictor is proposed, which is capable of encoding the input self-subgraph Gvin as a dependency of the derived pseudo-environment labels. This pseudo-environment label is mediated by means of a graph-structure-based environment estimator P. In order to capture environment information layer by layer at a finer level, the model introduces a layer-wise adaptive update mechanism that controls the propagation of multiple expert units instantiated by two different models. In particular, for a 10 Expert integration architecture based on graph convolutional networks, the layer-wise feature update is represented as follows: () =, () 1 ∈ () (,) () +(,) () and are the grades of the nodes vundu, (,) and (,) are the linear transformation matrices of the neighbor node information and the self-node information of the 15 expert branch of the 15 layer. An activation function for nonlinear transformations is introduced to ensure that complex interactions between node features can be adequately captured. The framework can be viewed as a causal representation of GCN that achieves adaptive propagation by dynamically selecting between the Z-folding filters of the 20 guiding model. To further improve the robustness and flexibility of the model, an adaptive expert model based on the attention mechanism is created to establish the pairwise interaction relationship between modeled nodes. The sum of the update model is represented as follows: 25 () =, () (,) ∈ () (,) () + (,) () (,) represents the attentional weight between nodes. Therefore, the model generates a high-dimensional embedded representation of node v through adaptive message propagation in the L layer and models the final prediction of node v on this basis through a fully connected layer. Throughout the architecture, the environment information of the nodes is captured and represented by the hidden vector ml-1, which is derived from the layer-wise adaptive expert integration model, thereby dynamically controlling feature transfer and the learning process. The design not only ensures the effective integration of local neighborhood node information but also context-adaptively adjusts the weights and feature paths of each layer through the derived pseudo-environment.This significantly improves the model's generalizability for handling different data distributions. Furthermore, the derived pseudo-environments provide rich contextual information for information propagation in each layer of the graphical neural network, enabling the model to adapt the propagation mechanism according to the different environmental features and thus exhibit greater generalizability and robustness with complex structured data distributions. This architecture can deliver excellent performance by capturing stable relationships in graph structures, particularly for node-level prediction tasks. At the same time, the model is able to maintain a high degree of prediction accuracy and stability in the face of various distributional deviations.which provides the theoretical basis and technical support for handling complex scenarios in practical applications. S4: Generalizing data through a multi-layered representation of the pseudo-environment in combination with a GCN or GAT encoder. Furthermore, generalizing data through a multi-layered representation of the pseudo-environment in combination with a GCN or GAT encoder includes combining GCN or GAT as the main structure of the encoder, using feature extraction capability in a non-Euclidean space, and improving model expression for high-dimensional topological data. Exemplary embodiment 2 As shown in Fig. 2 to 3, is an exemplary embodiment of the present invention, which 30 BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 17 provides an adaptive, extra-distributed intelligent diagnostic method based on graphene causality for high-quality mechanical devices to verify the advantageous effects of the present invention,Scientific proof was provided through economic benefit calculations and simulation experiments.5 First, two types of experimental scenarios were designed to thoroughly investigate the adaptability and robustness of the model. The first type of experiment aims to split the ID and OOD data by selecting data with different rotational speeds but the same error size (2×2mm).which we refer to as BUCEA-R. The second type of experiment focuses on the analysis of the effects of different error sizes at a fixed rotational speed (900 rpm) and is referred to as BUCEA-S. Table 1Specific design of the experimental program Samefaultsize(2x2mm).BUCEA-RSamespeed(900r).BUCEA-S TaskIDDataOODDataTaskIIIDDataOODData TaskI-1 500r,700r,900r, 1100r 1100rTaskII-1 0.4×1,2×2, 2.8×3,3.4×4 4×4 TaskI-2 500r,700r,900r, 1100r 1000rTaskII-2 0.4×1,2×2, 2.8×3,3.4×4 3.4×4 TaskI-3 500r,700r ,1000r, 1100r 900rTaskII-3 0.4×1,2×2, 3.4×4 2.8×3 TaskI-4 500r,900r,1000r, 700rTaskII-4 0.4×1,2×2, 2.8×3,3.4×4 2×2 BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 18 1100r TaskI-5 700r,900r,1000r, 1100r 500rTaskII-5 2×2,2.8×3, 3.4×4, 4×4 0.4×1 To effectively simulate the complexity of data distribution drift in the real-world operation of a mechanical device,Mixed datasets with varying rotation rates and error sizes are used in the experimental design. In particular, the model training phase of each dataset contains data for four different states, all of which are ID (In-Distribution) data. In the testing phase, the dataset contains one type of health state data that did not occur in the training phase, i.e., OOD (Out-Of-Distribution) data, which is used to evaluate the model's ability to recognize and detect OOD data. To fully validate the reliability and robustness of the model, five different health datasets were also used as OOD samples for multiple training and testing rounds. The advantageous performance of the graphical causal intervention-based OOD detection methods is further analyzed and investigated by conducting experiments with different Reference methods for OOD detection were compared. During the selection and division of the data set, the vibration signals are sliced ​​by a sliding window with a length of 1024.To extract the feature information from each sample, ensuring no overlap between adjacent time windows, the total signal length for each condition is 102400, thus dividing the data for each condition into 100 nodes. To improve the generalizability of the model, all ID node samples are randomly interrupted. Subsequently, the ID data nodes are randomly split into a training set, a validation set, and a test set with a partition ratio of 60%, 10%, and 30%, respectively. Finally, all experiments use a uniform learning strategy and a uniform experimental environment, in which the learning rate is set to 0.001 and the number of training rounds (epochs) is set to 300. All models are implemented on the basis of the BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 19 PyTorch framework.to ensure the consistency and comparability of the experiments. Using the BUCEA-R dataset, the effect of rotational speed variation on model performance for the same defect case (2×2mm) is investigated within the scope of the present invention, and the experimental results are shown in Fig. 2(a). The experimental results are shown in Fig. 2(a). It is shown that the classification accuracy of the GCN or GAT model as a backbone model for the OOD data is significantly better than that of the scenario with ID data at low rotational speed and OOD data at high rotational speed when the ID data contain high rotational speeds and the OOD data contain low rotational speeds. This phenomenon shows that the samples with high rotational speeds generally contain stronger and more significant vibrational signal features, and the model is able to extract clearer defect modes and capture high-frequency signal features during the training process. As a result, the model is Even with OOD data at low speeds during the test phase, it is able to identify potential error characteristics.which leads to higher classification accuracy.15 However, if the ID data consist mainly of low-speed samples, the model cannot fully capture the various failure features during the training phase because the vibration signals of the low-speed samples are weaker and lack universality. In this case, the complex failure modes and high-frequency vibration features contained in the high-speed OOD data20 exceed the feature distributions learned during model training, making it difficult for the model to adapt to changes in distribution, resulting in a significant degradation in performance. This result suggests that high-speed samples have an important advantage during model training,and underlines the importance of considering the strength of 25 vibration signal features in both generalization tasks outside the distribution. In the BUCEA-S experiments, the effect of variations in damage size on model generalization was also observed. When the ID data contain larger-scale damage patterns, the model's detection performance is significantly improved for OOD data with small damage, as shown in the experimental results in Fig. 2(b). Large-scale damage is generally associated with more distinct defect features, allowing the model to learn unambiguous defect patterns during the training phase and maintain high detection accuracy during the testing phase. These two sets of experimental results suggest that the intensity of the vibrational signals and the severity of the defect features in the data play a crucial role for the 5 Generalizability of the model plays a role. Strong vibrational signals and significant error modes can affect the model's ability toTo capture and improve critical characteristics and thereby increase the detection performance of extra-distributed samples. A systematic comparison of several classical and state-of-the-art out-of-distribution generalization methods is performed. The experiments comprise four vibrational signal datasets (CWRU, JNU, BUCEA-R and BUCEA-S), with the two common graph neural network (GNN) architectures, GCN and GAT, selected as the encoder backbone; the experimental results are listed in Table 2. The comparison methods were chosen to include classical empirical risk minimization (ERM) as a baseline, which uses the standard supervised loss function without adjustment for distribution drift; in addition, three common independent sample out-of-distribution generalization models were selected: IRM (invariant risk minimization), DANN (Domain-Adversarial Neural Network) and Mixup. To correspond to the characteristics of the data of the graph structure, SRGN is also used.to participate in 20 comparative experiments. To enable a fair comparison, all these methods were tested with GCN and GAT as the encoder backbone. Table 2: Recognition Accuracy of Different Algorithms with OOD Data in Four Datasets BackboneMethodCWRUJNUBUCEA-RBUCEA-S GCN ERM85.6277.3478.6376.81 DANN90.0783.4684.4283.93 Mixup91.4785.6187.6386.15 IRM87.7280.3881.2279.6 SRGNN90.2486.5285.3885.06 BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 21 Proposed96.0390.5490.7890.55 GAT ERM87.5180.8879.8278.39 DANN92.4585.4588.9483.72 Mixup92.0386.3988.2787.42 IRM88.4582.0383.1680.58 SRGNN91.5887.6487.3587.29 Proposed96.8791.3692.3291.73 The experimental results show,that GCI-ODG (graph-based causal derivation for solving the extra-distributed generalization problem) achieves the optimal OOD classification accuracy for all test datasets and both encoder architectures, significantly outperforming existing common methods. This excellent performance is mainly due to the application of causal derivation with the estimation strategy for 5 pseudo-environmental labels, which successfully eliminates environment-induced disturbances in the training data and enables the model to robustly capture stable feature relationships across operating conditions. For both highly heterogeneous datasets, BUCEA-R and BUCEA-S, the OOD accuracy of the GCI-ODG method under the GAT architecture is 92.35% and 91.73%, respectively, and is thus significantly 10 times better than that of the SRGNN and Mixup methods. Considering that these Both datasets contain highly heterogeneous error characteristics and significant distributional variations.GCI-ODG effectively captures the most important feature patterns under different operating conditions through adaptive estimation of the pseudo-environmental characterization, thereby achieving efficient modeling and identification of distribution drift data.15 In the vertical analysis of the performance of different coding architectures, it can be determined that the overall performance of all algorithms, when GAT is selected as the backbone, can be attributed to GCN. This result can be attributed to the adaptive attention mechanism of the GAT model. The GCN model uses the graph convolution operation with fixed weights, which is not able to flexibly distinguish between important and noisy features, especially in the two low-speed samples, and is difficult to effectively capture the most important vibrational signals.which leads to its weak detection performance with high-speed OOD data. In contrast, the GAT model prioritizes high-frequency vibration signals and significant error modes by adaptively adjusting the weights between nodes, which significantly improves the adaptability and generalizability of the model in complex environments.5 Embodiment3 As shown in Figures 3 to 4, an embodiment of the present invention provides an adaptive, extra-distributed intelligent diagnostic method based on graph causality for high-quality mechanical devices to verify the advantageous effects of the present invention.Scientific proof is provided by calculating the economic benefits and simulation experiments. The present invention developed an intelligent diagnostic framework based on graph causality interventions (called GCI-ODG) for the problem of generalizing out-of-distribution problems in wind turbines. Within the scope of this invention, a series of ablation experiments is designed and carried out to thoroughly evaluate the critical contribution of the modules in the proposed methodology; the results of the experiments are shown in Fig. 3. The results of the experiments are shown in Fig. 1. The ablation experiments consisted of three model configurations: First, the regulated loss term was removed and only the standard supervised loss was used for training, called GCI-ODG-A. Secondly, the pseudo-environment representation () of each layer is replaced by a single, globally shared representation. In this case, the model degenerates into a simplified version, GCI-ODG-B.which loses the ability to adaptively adjust the layers.25 Finally, the adaptive environment estimation mechanism is removed by replacing the trainable environment estimator with a non-parametric mean pooling operation, i.e., mean pooling is performed on the Z-spreading branches of each layer to form GCI-ODG-C. The experimental results show that the intelligent diagnostic system GCI-ODG significantly outperforms the 30 BE2025 / 5213 Y20P5BE-YUV005BEL02.04.2025 23 individual ablation versions in all test datasets, confirming the scientific validity and effectiveness of the method. In particular, the performance of GCI-ODG-A on the BUCEA dataset is significantly impaired, revealing the important role of the regularization loss term in highly heterogeneous and significant distribution drift scenarios. The regularization term successfully suppresses the influence of environmental biases by introducing independence conditions between pseudo-environmental labels and graph features.which facilitates the model's learning of environment-independent stable causal relationships. The absence of this mechanism makes the model more dependent on unstable features in the training data, leading to a significant performance degradation in the BUCEA dataset with large distributional variation and highlighting the indispensability of regularization constraints for improving generalizability. In contrast, the performance degradation of GCI-ODG-B demonstrates the central value of hierarchical pseudo-environment representations. The globally divided pseudo-environment representation cannot capture the complex variations between features at each layer, resulting in a model that exhibits a significant lack of flexibility when features are aggregated across multiple layers. Layer-adaptive pseudo-environment representations () are, however, able to dynamically capture and adapt changes in environmental features across different layers.which effectively improves the generalizability of the model on very heterogeneous data and exhibits higher adaptability and stability. Replacing the adaptive environment estimator with the mean pooling operation in GCI-ODG-C is simpler in terms of computational complexity, but has the disadvantage that the classification ability of the model is significantly impaired on OOD data. Mean pooling ignores the feature differences between nodes and cannot flexibly adjust the propagation path and weight assignment. In contrast, the adaptive environment estimator is able to dynamically adjust the weights according to the importance of the node features, thereby capturing important error features and high-frequency vibration signals more effectively and thus improving robustness and stability. Y20P5BE-YUV005BEL02.04.2025 24 The model's resilience under complex operating conditions is improved. The synergy effect of the three together forms a complete GCI-ODG framework, which makes it possible toto demonstrate excellent performance in the extra-distributed generalization task. To investigate the influence of the most important hyperparameters of the model, systematic analyses are carried out in the present invention for the number of propagation branches and the Gumbel softmax parameter, with the aim of uncovering the mechanism of their effect on the performance of the extra-distribution generalization. The experiments were carried out on four datasets to test the sensitivity and influence of the hyperparameter selection on the model performance under various environmental changes. The specific experimental results are in