A method and system for out-of-distribution fault detection based on energy propagation and graph learning

Abstract

The present invention relates to the technical field of intelligent detection of out-of-distribution faults of construction machinery, and discloses an out-of-distribution fault detection method and system based on energy propagation and graph learning, including: collecting vibration acceleration signals under typical fault states, performing similarity calculations, and obtaining an adjacency matrix consisting of maximum mutual information coefficients as input in a graph neural network. The GraphSage graph convolution method is used to extract features from the adjacency matrix and generate a representation of each node. The energy score of each node is calculated to distinguish between data within and outside the distribution. The out-of-distribution data confidence estimate of each node is strengthened, and out-of-distribution data of different working conditions of rolling bearings are identified and out-of-distribution data are detected. The present invention uses graph neural networks and energy-based models to effectively capture complex fault correlations and improve the accuracy of fault diagnosis. Accurate distinction between data within and outside the distribution is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent detection of external faults of construction machinery, and in particular to a method and system for detecting external faults of construction machinery based on energy propagation and graph learning. Background Art

[0002] The present invention relates to the field of intelligent detection of out-of-distribution faults in construction machinery. It is worth noting that although the focus of this study is on the practice and theory of intelligent fault diagnosis of construction machinery and equipment in the field of performance health management, the defined problem framework is far more than that. It is not only applicable to fault diagnosis, but also includes key issues of PHM such as fault prediction and remaining useful life (RUL) estimation. Although construction machinery and equipment are selected as the main application scenarios of this study due to their importance in the field of construction engineering, the application potential of the problem definition and solutions discussed in this study is not limited to this. In fact, these methods and theories can be easily migrated to other types of complex mechanical systems, such as industrial robots and automated production lines, and even more widely used in any equipment that requires high reliability and performance monitoring. This wide applicability stems from a deep understanding of the nature of the problem and the flexible application of solution strategies, emphasizing the importance of cross-domain applicability and model versatility. Through in-depth research on out-of-distribution data detection and processing methods, it aims to provide a set of reliable, effective and easy-to-extend tools and frameworks for the field of PHM and even for a wider range of industrial applications, so as to promote the application and development of intelligent systems in the real world.

[0003] With the pursuit of sustainable and intelligent buildings, construction machinery and equipment have become an important part of the modern construction industry. The efficient operation of construction machinery and equipment represents an innovative and cost-effective construction method, which coincides with the global efforts to promote green buildings and reduce carbon emissions. However, the operating efficiency of these complex construction machinery and equipment is often affected by various faults, ranging from minor mechanical wear and tear to major system failures. These failures not only lead to a lot of downtime, but also cause significant economic losses, hindering the overall progress and quality of construction projects. Therefore, effective fault diagnosis strategies are essential for the proactive management of construction machinery and equipment, which can ensure that they operate at optimal performance levels and make significant contributions to smart buildings and sustainable building goals.

[0004] The rise of deep learning technology has triggered revolutionary changes in many fields, especially its excellent ability to process complex data and extract key features, which has made it widely adopted. Nevertheless, traditional deep learning still faces challenges in solving the problem of out-of-distribution data (OOD) detection, because most current deep learning methods implicitly assume that the training data and test data follow the same distribution, and the model performance is verified based on their generalization to the test data of the same distribution. At the same time, most current methods are based on the assumption that data samples are generated independently (e.g., image recognition where instances do not interact). Such a premise hinders the adaptability of existing deep learning methods to graph-structured data with interdependencies. Therefore, exploring the relevant applications of graph neural networks (GNNs) represents an important shift in solving the problem of out-of-distribution data in data with intrinsic associations and interconnections. Summary of the invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problems solved by the present invention are:

[0007] To solve the above technical problems, the present invention provides the following technical solutions: a method for out-of-distribution fault detection based on energy propagation and graph learning, comprising: collecting vibration acceleration signals under typical fault conditions, performing similarity calculations, and obtaining an adjacency matrix composed of maximum mutual information coefficients as input in a graph neural network.

[0008] The GraphSage graph convolution method is used to extract features from the adjacency matrix and generate representations for each node.

[0009] The model between the energy function and the probability density is connected to obtain the energy function under the GNN framework, calculate the energy score of each node, set the energy score threshold, and distinguish between data within and outside the distribution.

[0010] Strengthen the confidence estimation of out-of-distribution data for each node, and identify and detect out-of-distribution data under different working conditions of rolling bearings according to the intelligent diagnosis framework of out-of-distribution data based on energy score.

[0011] As a preferred scheme of the distributed external fault detection method based on energy propagation and graph learning described in the present invention, wherein: the vibration acceleration signals under the typical fault state include healthy H, outer ring fault F1, inner ring fault F2, rolling body fault F3, outer ring-rolling body composite fault F4 and outer ring-inner ring-rolling body composite fault F5.

[0012] As a preferred solution of the out-of-distribution fault detection method based on energy propagation and graph learning described in the present invention, wherein: the similarity calculation includes collecting vibration acceleration signals B=[b1, b2, ..., b n ] and C=[c1,c2,...,c n ], discretize and convert into discrete variables.

[0013] Calculate the mutual information between B and C, and select the variable pair with the largest mutual information as a candidate. The mutual information I(B,C) is expressed as:

[0014]

[0015] Where I(B,C) represents the mutual information coefficient, p(b,c) represents the joint probability distribution function of time series B and C, p(b) represents the marginal probability density function of B, p(c) represents the marginal probability density function of C, b i represents the i-th value of time series B, c i represents the i-th value of time series C, i = 1, 2, ..., n. b represents the value of the discretized variable B, and c represents the value of the discretized variable C.

[0016] The maximum mutual information coefficient is used as the similarity evaluation index. The maximum mutual information coefficient value is obtained by normalizing the candidate variables. The statistic is defined as a triple (B, C, MIC), where B and C represent the discretized versions of the input variables, and MIC represents the corresponding maximum mutual information coefficient, which is expressed as:

[0017]

[0018] Among them, MIC(B,C) represents the maximum mutual information coefficient, represents the number of grids divided in the B and C directions, and ξ represents the selection parameter.

[0019] The adjacency matrix formed by the maximum mutual information coefficient has a dimension of N×N, which is used as the input of the graph neural network to convert the one-dimensional time series data into graph data that can be recognized by the graph neural network model, expressed as:

[0020]

[0021] Among them, ρ represents the adjacency matrix, p NN Indicates the maximum mutual information coefficient between two time points in time series B and C at the Nth moment. ij Represents the maximum mutual information coefficient between node i and node j, which is a similarity measure between two nodes. As the weight coefficient of each edge, i = 1, 2, ..., N, j = 1, 2, ..., N.

[0022] As a preferred solution of the out-of-distribution fault detection method based on energy propagation and graph learning described in the present invention, wherein: the GraphSage graph convolution method is used to extract features from the adjacency matrix, and each node representation is generated, including generating node representations by sampling and aggregating features of neighboring nodes, and introducing a weighted average attention mechanism to weight the features of neighboring nodes to highlight important feature information.

[0023] Assumptions represents the node representation of node v at the sth layer, Represents the neighbor set of node v, and the node representation update formula of the s+1th layer is expressed as:

[0024]

[0025] in, represents the node representation of node v in the s+1th layer, σ represents the nonlinear activation function, and W (s) represents the weight matrix of the sth layer, It indicates that node u is in the sth layer, and Aggregate indicates the mean aggregation function, which is expressed as:

[0026]

[0027] The obtained attention weight is normalized by the Softmax function to get α vu , feature aggregation is performed through the weighted average attention mechanism, and the updated node is represented as:

[0028]

[0029] Among them, α vu represents the result after normalization of attention weights, W represents the weight matrix, and h represents the node representation.

[0030] As a preferred solution of the out-of-distribution fault detection method based on energy propagation and graph learning described in the present invention, wherein: the model between the connection energy function and the probability density, the energy function under the GNN framework is obtained, including: in the deep learning model, the K-class classification problem uses the parameter function To achieve, Represents the feature space of the input data x, where each data point x is a D-dimensional vector. Represents the output logits space, each logit vector is a K-dimensional vector corresponding to K categories. θ Represents a parameterized mapping function, parameter function f θ Each data point Mapped to k real-valued logits. After S layers of graph convolution, the GNN model outputs a K-dimensional vector As the logits of each node, it is expressed as:

[0031]

[0032] in, represents the logits obtained after each data point passes through S layers of graph convolution, x i represents the input feature vector of the i-th node, where i represents the index of the node.

[0033] According to the update rule of GNN, represents the function of the predicted logits of the S-order ego-graph centered on instance x, and θ represents the trainable parameters in the GNN model f.

[0034] Input logits into the softmax function to get the probability distribution, so that the model can perform classification tasks. Under the condition of , the conditional probability of the target category y is expressed as:

[0035]

[0036] in, represents the probability that node x is classified as category y, Represents the neighbor nodes and their connection relationships with node x. Represents a parameterized function f θ When inputting feature x and graph structure information The output of , is a K-dimensional vector, each component corresponds to the logit of a category. express The y-th index of , K represents the total number of categories, [y] represents the element with index y, and [k] represents the element with index k.

[0037] Combining the conditional probability of the target category y with the EBM model that defines the energy function and probability density, the energy function induced by the GNN model is expressed as:

[0038]

[0039] in, Represents the calculated energy value.

[0040] The LogSumExp(·) of the logits of the GNN classifier is reused to define the energy function at the data point x, expressed as:

[0041]

[0042] in, represents the energy function, f θ (x) [k] Represents the function f θ On input x and The kth element in the vector calculated below, LogSumExp(·) represents the calculation of the log-likelihood function, Represents a parameterized function f θ When inputting feature x and graph structure information The component corresponding to the kth category in the calculated output logit vector.

[0043] The training loss is expressed as:

[0044]

[0045] in, represents the expected value operator, represents the distribution of the labeled part of the training data. Represents a node with a label, represents the supervised training loss, Represents a parameterized function f θ When inputting feature x and graph structure information In the calculated output logit vector, the y i The weight corresponding to each category, Represents a parameterized function f θ In the input feature x i and graph structure information The component corresponding to the kth category in the calculated output logit vector.

[0046] An energy score threshold is set, and the energy score of each node is calculated according to the energy function under the GNN framework to distinguish between in-distribution and out-of-distribution data.

[0047] If the energy score is lower than the threshold, it is considered as in-distribution data. If it is higher than the threshold, it is considered as out-of-distribution data.

[0048] As a preferred solution of the out-of-distribution fault detection method based on energy propagation and graph learning described in the present invention, wherein: the strengthening of the out-of-distribution data confidence estimation of each node includes defining As a node v in the graph i The initial energy fraction vector, E (0) Represents the node v in the graph i The initial energy fraction vector of The energy function is used to calculate the energy value of the node, which is expressed as follows through the propagation update rule:

[0049]

[0050] Among them, 0<λ<1, which represents the parameter that controls the energy transfer concentration between the node itself and other connected nodes. (s) Represents the energy score vector of the s-th layer node, E (s-1) represents the energy score vector of the s-th layer node, D represents the degree matrix, A represents the adjacency matrix, Represents all node domains, Represents the energy value of the i-th node after the s-th iteration.

[0051] According to the propagation rules of the updated nodes, after S steps of propagation, the final result after energy transfer is obtained in, It represents the final energy value of node i after S-step propagation iterations, and performs out-of-distribution data detection and judgment on the result after the final energy transfer.

[0052] The energy transfer in the graph topology uses the energy values ​​of neighboring nodes to strengthen the out-of-distribution data confidence estimate of each node. For any given data sample point x i , if the average energy score of its one-hop neighbor node is Than its own energy fraction Low, expressed as:

[0053]

[0054] According to the propagation update rule E (s) get:

[0055]

[0056] Among them, E i (s) represents the node energy score vector of node i in layer s, a ij represents the weight of the direct edge connecting node i and node j, E i (s-1) Represents the node energy score vector of node i in the s-1th layer.

[0057] The decision criterion of the out-of-distribution data discriminator is expressed as:

[0058]

[0059] in, represents a binary classification function, which determines whether the input x belongs to the data within the distribution or outside the distribution. ε represents the judgment threshold. Represents the result after final energy transfer.

[0060] As a preferred scheme of the out-of-distribution fault detection method based on energy propagation and graph learning described in the present invention, wherein: the identification of out-of-distribution data and the detection of out-of-distribution data under different working conditions of rolling bearings based on the out-of-distribution data intelligent diagnosis framework based on energy scores include four steps: graph construction, feature encoding, energy score detection and energy propagation update.

[0061] The graph construction includes converting the vibration signals collected by the sensor into graph data using the maximum mutual information coefficient and constructing an adjacency matrix that reflects the nonlinear relationship between faults.

[0062] Feature encoding includes the use of the GraphSage model and the weighted average attention mechanism to extract features from graph data. GraphSage generates node representations by aggregating neighbor node features, and the attention mechanism strengthens the weights of important neighbor features and enhances the representation ability of the model.

[0063] Energy score detection includes introducing out-of-distribution data discrimination based on energy scores, and establishing the energy function of data points by redefining the logits of the classifier. By calculating the energy score to distinguish between in-distribution data and out-of-distribution data, effective out-of-distribution data fault detection is achieved.

[0064] Energy propagation updates include improving the generalization ability of the model in a semi-supervised learning environment, and proposing an energy score propagation mechanism based on a graph structure. By iteratively propagating node energy scores on the graph, using neighbor node information to update energy values, strengthening the energy difference between in-distribution data and out-of-distribution data, and improving the accuracy of out-of-distribution data detection.

[0065] A distributed fault detection system based on energy propagation and graph learning, characterized by: comprising:

[0066] The acquisition module collects the vibration acceleration signals under typical fault conditions, performs similarity calculation, and obtains the adjacency matrix composed of the maximum mutual information coefficient as the input in the graph neural network.

[0067] The feature extraction module uses the GraphSage graph convolution method to extract features from the adjacency matrix and generate a representation of each node.

[0068] The energy score calculation module connects the model between the energy function and the probability density, obtains the energy function under the GNN framework, calculates the energy score of each node, sets the energy score threshold, and distinguishes between data within and outside the distribution.

[0069] The optimization module strengthens the confidence estimation of the out-of-distribution data of each node, and identifies and detects the out-of-distribution data of different working conditions of rolling bearings according to the intelligent diagnosis framework of out-of-distribution data based on energy score.

[0070] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0071] A computer-readable storage medium stores a computer program, which implements the steps of the method described above when executed by a processor.

[0072] Beneficial effects of the invention: This patent proposes an innovative energy-driven graph neural out-of-distribution detection framework, which is designed to cope with the challenges of out-of-distribution data in complex and dynamic environments. The framework uses graph neural networks and energy-based models to effectively capture complex fault correlations and improve the accuracy of fault diagnosis. The framework uses the maximum information coefficient to convert the vibration signals collected by the sensor into graph data, creating an adjacency matrix representing the nonlinear relationship between various fault types. In addition, it includes an energy score-based out-of-distribution data detection module, which redefines the logits of the classifier to establish an energy function, thereby achieving accurate distinction between in-distribution (ID) and out-of-distribution data (OOD). To further enhance the robustness of the model in a semi-supervised environment, an energy score update scheme based on a propagation mechanism iteratively optimizes the energy value within the graph. Experiments conducted on a custom-built mechanical equipment bearing wear monitoring platform verify the superior performance of the framework in detecting and diagnosing various fault conditions. The results show that the framework has high generalization ability and detection accuracy, provides strong technical support for the intelligent diagnosis of construction machinery, and promotes intelligent and sustainable construction practices. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0074] Figure 1 An overall flow chart of an out-of-distribution fault detection method and system based on energy propagation and graph learning provided for the first embodiment of the present invention.

[0075] Figure 2 A diagram showing the influence of different graph construction methods on model performance of an out-of-distribution fault detection method based on energy propagation and graph learning and a system provided in accordance with the second embodiment of the present invention. DETAILED DESCRIPTION

[0076] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0077] Example 1, reference Figure 1 , as an embodiment of the present invention, provides a method for out-of-distribution fault detection based on energy propagation and graph learning, comprising:

[0078] S1: Collect vibration acceleration signals under typical fault conditions, perform similarity calculations, and obtain an adjacency matrix consisting of the maximum mutual information coefficient as input in the graph neural network.

[0079] The vibration acceleration signals under typical fault conditions include healthy H, outer ring fault F1, inner ring fault F2, rolling element fault F3, outer ring-rolling element composite fault F4 and outer ring-inner ring-rolling element composite fault F5.

[0080] Collect vibration acceleration signals B=[b1,b2,...,b n ] and C=[c1,c2,...,c n ], discretize and convert into discrete variables.

[0081] Calculate the mutual information between B and C, and select the variable pair with the largest mutual information as a candidate. The mutual information I(B,C) is expressed as:

[0082]

[0083] Where I(B,C) represents the mutual information coefficient, p(b,c) represents the joint probability distribution function of time series B and C, p(b) represents the marginal probability density function of B, p(c) represents the marginal probability density function of C, b i represents the i-th value of time series B, c i represents the i-th value of time series C, i = 1, 2, ..., n. b represents the value of the discretized variable B, and c represents the value of the discretized variable C.

[0084] The maximum mutual information coefficient is used as the similarity evaluation index. The maximum mutual information coefficient value is obtained by normalizing the candidate variables. The statistic is defined as a triple (B, C, MIC), where B and C are discretized versions of the input variables, and MIC is the corresponding maximum mutual information coefficient, expressed as:

[0085]

[0086] Among them, MIC(B,C) represents the maximum mutual information coefficient, represents the number of grids divided in the B and C directions, and ξ represents the selection parameter.

[0087] It should be noted that the present invention selects ξ=n 0.6 .

[0088] The adjacency matrix formed by the maximum mutual information coefficient has a dimension of N×N, which is used as the input of the graph neural network to convert the one-dimensional time series data into graph data that can be recognized by the graph neural network model, expressed as:

[0089]

[0090] Among them, ρ represents the adjacency matrix, p NN Indicates the maximum mutual information coefficient between two time points in time series B and C at the Nth moment. ij Represents the maximum mutual information coefficient between node i and node j, which is a similarity measure between two nodes. As the weight coefficient of each edge, i = 1, 2, ..., N, j = 1, 2, ..., N.

[0091] S2: Use the GraphSage graph convolution method to extract features from the adjacency matrix and generate representations for each node.

[0092] It should be noted that GraphSage, as a flexible and scalable graph convolution method, generates node representations by sampling and aggregating the features of neighboring nodes. The core idea is to sample a fixed number of neighbors in each layer and aggregate the features of these neighbors to generate node features with stronger representation capabilities. This aggregation method can effectively capture the structural information and features of nodes and their neighbors, and is particularly suitable for large-scale graph data. At the same time, the GraphSage convolution layer can aggregate the information of local neighbors at each layer, which helps to extract and integrate local features layer by layer, so that the final features of the node can reflect its local structure and feature distribution.

[0093] Assumptions represents the node representation of node v at the sth layer, Represents the neighbor set of node v, and the node representation update formula of the s+1th layer is expressed as:

[0094]

[0095] in, represents the node representation of node v in the s+1th layer, σ represents the nonlinear activation function, and W (s) represents the weight matrix of the sth layer, It indicates that node u is in the sth layer, and Aggregate indicates the mean aggregation function, which is expressed as:

[0096]

[0097] Furthermore, in graph neural networks, simple feature aggregation alone may not be able to fully capture the complex relationships between nodes. To this end, we invented the weighted average attention mechanism to weight the features of neighbor nodes to highlight important feature information. The basic idea of ​​the weighted average attention mechanism is to learn the attention weights between each node and its neighbor nodes so that the features of important neighbor nodes occupy a larger proportion in the aggregation process. The obtained attention weights are normalized by the Softmax function to obtain α vu Finally, feature aggregation is performed through the weighted average attention mechanism to obtain the updated node representation:

[0098]

[0099] Among them, α vu represents the result after normalization of attention weights, W represents the weight matrix, and h represents the node representation.

[0100] Furthermore, through the above process, the feature encoder can effectively capture the complex dependencies between nodes and achieve efficient feature extraction and processing in heterogeneous graph data. This method not only enhances the representation ability of node features, but also improves the robustness of the model when processing complex graph structures and heterogeneous graph data. Through the weighted average attention mechanism, the model can dynamically adjust the importance of each node feature, making information transmission more accurate, thereby further improving the accuracy and reliability of fault diagnosis. In addition, combined with the scalability of GraphSage and the flexibility of the attention mechanism, the feature encoder invented by our party provides a new solution for the efficient processing of large-scale graph data and has broad application prospects.

[0101] S3: Connect the model between energy function and probability density to obtain the energy function under the GNN framework, and train the function loss through the negative log-likelihood loss function. According to the energy function under the GNN framework, the energy score of each node is calculated to distinguish between in-distribution and out-of-distribution data.

[0102] In deep learning models, K-class classification problems use parameter functions To achieve, Represents the feature space of the input data x, where each data point x is a D-dimensional vector. Represents the output logits space, each logit vector is a K-dimensional vector corresponding to K categories. θRepresents a parameterized mapping function, parameter function f θ Each data point Mapped to k real-valued logits,. After S layers of graph convolution, the GNN model outputs a K-dimensional vector As the logits of each node, it is expressed as:

[0103]

[0104] in, represents the logits obtained after each data point passes through S layers of graph convolution, x i represents the input feature vector of the i-th node, where i represents the index of the node.

[0105] According to the update rule of GNN, represents the function of the predicted logits of the S-order ego-graph centered on instance x, and θ represents the trainable parameters in the GNN model f.

[0106] Input logits into the softmax function to get the probability distribution, so that the model can perform classification tasks. Under the condition of , the conditional probability of the target category y is expressed as:

[0107]

[0108] in, represents the probability that node x is classified as category y, Represents the neighbor nodes and their connection relationships with node x. Represents a parameterized function f θ When inputting feature x and graph structure information The output of , is a K-dimensional vector, each component corresponds to the logit of a category. express The y-th index of K represents the total number of categories. It represents the element with index y, and [k] represents the element with index k.

[0109] Combining the conditional probability of the target category y with the EBM model that defines the energy function and probability density, the energy function induced by the GNN model is expressed as:

[0110]

[0111] in, Represents the calculated energy value.

[0112] It should be noted that the energy values ​​given above are obtained directly from the GNN classifier predicting logits, that is, there is no need to change the relevant parameters of the GNN.

[0113] The LogSumExp(·) of the logits of the GNN classifier is reused to define the energy function at the data point x, expressed as:

[0114]

[0115] in, represents the energy function, f θ (x) [k] Represents the function f θ On input x and The kth element in the vector calculated below, LogSumExp(·) represents the calculation of the log-likelihood function, Represents a parameterized function f θ When inputting feature x and graph structure information The component corresponding to the kth category in the calculated output logit vector.

[0116] Furthermore, in deep learning, the negative log-likelihood loss function is often used for classification tasks, especially in multi-class classification problems. It measures the difference between the model's prediction of the observed data and the actual label, that is, the difference between the model's predicted probability distribution and the probability distribution of the true label. When training with negative log-likelihood loss, the model will be overconfident in predicting data within the distribution, that is, the energy of data samples within the distribution is reduced. Therefore, for node classification models, GNN models are usually trained by minimizing the negative log-likelihood of labeled training data, that is, supervised training loss, as follows:

[0117]

[0118] in, represents the expected value operator, represents the distribution of the labeled part of the training data.

[0119] Represents a node with a label, represents the supervised training loss, Represents a parameterized function f θ When inputting feature x and graph structure information In the calculated output logit vector, the y i The weight corresponding to each category, Represents a parameterized function f θ In the input feature x i and graph structure information The component corresponding to the kth category in the calculated output logit vector.

[0120] Set the energy score threshold, calculate the energy score of each node according to the energy function under the GNN framework, and distinguish between in-distribution and out-of-distribution data.

[0121] If the energy score is lower than the threshold, it is considered as in-distribution data. If it is higher than the threshold, it is considered as out-of-distribution data.

[0122] It should be noted that in energy-based out-of-distribution detection, the setting of the threshold τ is a key step, which directly affects the model's ability to distinguish between in-distribution data and out-of-distribution data. The threshold setting step includes training the GNN model using training data. During the training process, the model learns how to map input data to output energy scores. The training data usually contains in-distribution data. For the trained model, the energy score of each sample is calculated on the validation set. Based on the distribution of the validation set: the distribution of energy scores is calculated on the validation set. Usually, a dividing point is selected as the threshold so that the energy scores of most in-distribution samples are low and the energy scores of out-of-distribution samples are high. For example, a point can be selected so that 95% of the in-distribution samples have energy scores below the threshold. Once the threshold τ is determined, in the test phase, for each test sample, its energy score is calculated and compared with the threshold. If the energy score is lower than the threshold, the sample is considered to belong to in-distribution data. If it is higher than the threshold, it is considered to belong to out-of-distribution data.

[0123] Furthermore, for example, suppose the distribution of energy scores calculated on the validation set is such that the energy scores of in-distribution data are mostly in the range of [-5, -1]. The energy scores of out-of-distribution data are mostly in the range of [0, 5]. Then a value between -1 and 0 is selected as the threshold, such as τ = -0.5, which can effectively distinguish in-distribution data from out-of-distribution data.

[0124] S4: Strengthen the confidence estimation of out-of-distribution data for each node, and identify and detect out-of-distribution data under different working conditions of rolling bearings according to the intelligent diagnosis framework of out-of-distribution data based on energy score.

[0125] It should be noted that since labeled data is usually scarce in a semi-supervised learning environment, the energy score obtained directly from the GNN model trained using supervised loss on labeled data may not be sufficient to generalize the input data type. Therefore, the key to enhancing the generalization ability of the GNN model lies in how to use the unlabeled data in the training set to help the model better identify the graph topology behind the data. Inspired by the classic non-parametric semi-supervised learning algorithm - label propagation, an energy score update scheme based on the propagation mechanism is proposed, which iteratively propagates the energy scores of each node between interconnected nodes on the observed graph topology.

[0126] definition As a node v in the graph i The initial energy fraction vector, E (0) Represents the node v in the graph i The initial energy fraction vector of The energy function is used to calculate the energy value of the node, which is expressed as follows through the propagation update rule:

[0127]

[0128] Among them, 0<λ<1, which represents the parameter that controls the energy transfer concentration between the node itself and other connected nodes. (s) Represents the energy score vector of the s-th layer node, E (s-1) represents the energy score vector of the s-th layer node, D represents the degree matrix, A represents the adjacency matrix, represents all node domains, E i (s) Represents the energy value of the i-th node after the s-th iteration.

[0129] According to the propagation rules of the updated nodes, after S steps of propagation, the final result after energy transfer is obtained in, It represents the final energy value of node i after S-step propagation iterations, and performs out-of-distribution data detection and judgment on the result after the final energy transfer.

[0130] It should be noted that the basic principle behind the energy score update scheme based on the propagation mechanism is to adapt to the physical mechanism and instance interaction in data generation. Since the input graph can reflect the geometric structure between data samples with a certain degree of similarity on the manifold, it is natural that all node generation and prediction of unknown node labels are conditioned on their neighbors. Therefore, nodes with connected relationships are often sampled from similar distributions. Considering that the energy function at the data point x is an effective basis for judging out-of-distribution data, and the propagation of the energy score in space basically mimics this physical mechanism of data generation, it will enhance the confidence of out-of-distribution data detection. At the same time, the transfer of energy in the graph topology can use the energy values ​​of adjacent nodes to strengthen the out-of-distribution data confidence estimate of each node. For any given data sample point x i , if the average energy score of its one-hop neighbor node is Than its own energy fraction Low, expressed as:

[0131]

[0132] According to the propagation update rule E (s) get:

[0133]

[0134] Among them, E i (s) represents the node energy score vector of node i in layer s, a ij represents the weight of the direct edge connecting node i and node j, E i (s-1) Represents the node energy score vector of node i in the s-1th layer.

[0135] Furthermore, the opposite result can be proved by similar method. Therefore, the energy score update scheme based on the propagation mechanism can push the energy value to most of the surrounded nodes, that is, as the iterative propagation process proceeds, the energy value of the data within the distribution approaches the minimum and the energy value of the data outside the distribution approaches the maximum. This can help the model method amplify the energy difference between the samples within the distribution and the data samples outside the distribution to provide the model's judgment ability.

[0136] The decision criterion of the out-of-distribution data discriminator is expressed as:

[0137]

[0138] in, represents a binary classification function, which determines whether the input x belongs to the data within the distribution or outside the distribution. ε represents the judgment threshold. Represents the result after final energy transfer.

[0139] Samples with high energy values ​​are considered as out-of-distribution data inputs.

[0140] Furthermore, this result shows that the energy score update model based on the propagation mechanism can provide the model with the discriminative ability for out-of-distribution data samples through a simple propagation scheme during inference without any additional training cost.

[0141] According to the intelligent diagnosis framework of out-of-distribution data based on energy score, the recognition and detection of out-of-distribution data of rolling bearings in different working conditions include four steps: graph construction, feature encoding, energy score detection and energy propagation update.

[0142] The graph construction includes converting the vibration signals collected by the sensor into graph data using the maximum mutual information coefficient and constructing an adjacency matrix that reflects the nonlinear relationship between faults.

[0143] Feature encoding includes the use of the GraphSage model and the weighted average attention mechanism to extract features from graph data. GraphSage generates node representations by aggregating neighbor node features, and the attention mechanism strengthens the weights of important neighbor features and enhances the representation ability of the model.

[0144] Energy score detection includes introducing out-of-distribution data discrimination based on energy scores, and establishing the energy function of data points by redefining the logits of the classifier. By calculating the energy score to distinguish between in-distribution data and out-of-distribution data, effective out-of-distribution data fault detection is achieved.

[0145] Energy propagation updates include improving the generalization ability of the model in a semi-supervised learning environment, and proposing an energy score propagation mechanism based on a graph structure. By iteratively propagating node energy scores on the graph, using neighbor node information to update energy values, strengthening the energy difference between in-distribution data and out-of-distribution data, and improving the accuracy of out-of-distribution data detection.

[0146] This embodiment also provides a distributed fault detection system based on energy propagation and graph learning, specifically:

[0147] The acquisition module collects the vibration acceleration signals under typical fault conditions, performs similarity calculation, and obtains the adjacency matrix composed of the maximum mutual information coefficient as the input in the graph neural network.

[0148] The feature extraction module uses the GraphSage graph convolution method to extract features from the adjacency matrix and generate representations for each node.

[0149] The energy score calculation module connects the model between the energy function and the probability density, obtains the energy function under the GNN framework, calculates the energy score of each node, sets the energy score threshold, and distinguishes between data within and outside the distribution.

[0150] The optimization module strengthens the confidence estimation of the out-of-distribution data of each node, and identifies and detects the out-of-distribution data of different working conditions of rolling bearings according to the intelligent diagnosis framework of out-of-distribution data based on energy score.

[0151] The computer device may be a server. The computer device includes a processor, a memory, an input / output interface (I / O for short) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data cluster data of the power monitoring system. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection.

[0152] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0153] Example 2, reference Figure 2 , which is an embodiment of the present invention, provides a distributed out-of-fault detection method and system based on energy propagation and graph learning. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0154] In order to simulate the complex data conditions in the process of intelligent diagnosis and out-of-distribution data detection of actual construction machinery and equipment, a mixed data set containing different health states was used in the testing process of the model. Specifically, the input in the model training process consists of data of 5 different health states, namely, in-distribution data; based on the 5 in-distribution data, a health state data that is invisible in the model training stage, namely, out-of-distribution data, is added to the test set to evaluate the model's detection ability for out-of-distribution data. In addition, in order to verify the reliability and robustness of the model, 6 different health states are used as out-of-distribution data in turn, and the model is trained and tested multiple times. At the same time, by conducting the above training and testing experiments on different out-of-distribution data detection benchmark methods, the superior performance of the out-of-distribution data detection method based on energy score is further analyzed and compared.

[0155] In the selection and division of the data set, the vibration signal data of 6 different operating conditions are truncated through a sliding time window of length 1024 and there is no overlap between the sliding time windows, as the feature information of each node, and there are 500 nodes for each condition. In addition, these node samples are randomly shuffled to ensure that the trained model has good generalization performance. During the training and testing of the model, the training set consists of 2500 node samples of 5 different fault conditions, and the mixed test set includes 500 in-distribution data sample nodes and 100 out-of-distribution data sample nodes from 5 different fault conditions. The specific experimental setting information is shown in Table 1. At the same time, the proposed out-of-distribution data check framework and all benchmark models use consistent learning strategies and environments, with a learning rate and epoch of 0.001 and 200, respectively. All models are implemented based on the PyTorch framework.

[0156] Table 1 Description of each experimental group

[0157]

[0158] First, the performance of the proposed model of our invention under different experimental groups (E1 to E5) is comprehensively analyzed. The accuracy and reliability of intelligent diagnosis of rotating machinery are crucial to ensure the stable operation of construction machinery and equipment. Different types of faults (such as external faults, internal faults, compound faults, etc.) pose different challenges to the proposed model. Therefore, it is very important to evaluate the performance of our invention under a variety of experimental conditions. By analyzing the results of different experimental groups, we can gain a deeper understanding of the model's capabilities and advantages in dealing with diverse faults. Table 2 summarizes the three key evaluation indicators of the five experimental groups: AUROC, AUPR, and FPR95.

[0159] Table 2 Evaluation indicators of EGN-out-of-distribution data in different experimental groups

[0160]

[0161] The data in Table 2 clearly show that EGN-out-of-distribution data has high detection capabilities in various out-of-distribution data types, especially in identifying F3. The data in Table 2 clearly show that EGN-out-of-distribution data has high detection capabilities in various out-of-distribution data types, especially in identifying F3. Specifically, for the E3 group, the AUROC is 94.42%, the AUPR is 95.62%, and the FPR95 is 23.11%. These indicators show that EGN-out-of-distribution data has high accuracy and precision in identifying rolling element faults, while the false alarm rate is relatively low. Thanks to the energy score update scheme of EGN-out-of-distribution data based on the propagation mechanism, it ensures that the energy score of the in-distribution node is significantly lower than that of the out-of-distribution node. Coupled with the energy score-based out-of-distribution data detection module, the AUROC and AUPR values ​​of all experimental groups are over 92%. This shows that EGN-out-of-distribution data can maintain high accuracy and precision in various fault detection. These results are of great significance for intelligent diagnosis of construction machinery; high AUROC values ​​indicate that normal and faulty states can be effectively distinguished, while high AUPR values ​​indicate that the model has high precision and recall in actual fault detection, which is crucial to improving the reliability of equipment operation. In addition, these results also show that the model can not only effectively identify faults, but also do a good job of reducing false alarms.

[0162] Next, a routine experimental verification is conducted, that is, EGN-out-of-distribution data is compared with two types of baseline models for out-of-distribution data detection to verify the superiority of the proposed model. The first type of baseline model is focused on out-of-distribution data detection for processing image output, specifically: MSP, ODIN, Energy. In order to ensure the consistency of variables, the GNN encoder used to process graph structure data replaces the original convolutional neural network backbone. The second type of baseline model is specially designed for processing out-of-distribution data in the field of graph deep learning. Two advanced benchmark models, GPN and GKDE, are selected for comparison.

[0163] First, for the out-of-distribution data detection models that focus on processing image output (MSP, ODIN, Energy), the experimental results show that the AUROC values ​​of these methods are relatively low in each group. The MSP method relies on the probability value of the model output for detection, but due to the complex and changeable vibration signals in construction machinery, it cannot capture enough information. In contrast, thanks to the design of the feature encoder and the energy propagation mechanism, the AUROC values ​​of EGN-out-of-distribution data in each experimental group are significantly higher by more than 20 percentage points, especially in the E3 group, where the AUROC reaches 94.42, while MSP is only 75.28. This shows that EGN-out-of-distribution data is more effective in identifying faults when processing complex vibration signals. Although the ODIN method enhances the detection capability through input perturbations and temperature scaling, its robustness under complex fault modes is still insufficient. The Energy method uses the energy value of the input sample for detection and performs better than MSP and ODIN, while the proposed model still far exceeds this competitor, which further demonstrates the superiority of energy-based out-of-distribution data detection models on graphs. Secondly, for the out-of-distribution data models (GPN and GKDE) in the field of graph deep learning, GKDE estimates the probability density of data distribution for detection, and its AUROC value is relatively stable in all experimental groups, but it performs poorly when processing complex vibration signals. However, EGN-out-of-distribution data still performs better in all experimental groups, especially when dealing with complex rotating machinery faults, with more accuracy and robustness.

[0164] In order to explore the expressiveness of different encoder backbones on the proposed energy score and transfer mechanism modules, three evaluation indicators were set for five different backbones including MLP, JKNet, MixHop, GCN and the proposed encoder in the E1-E5 experimental groups. The experimental results show that different backbones do have a significant impact on the detection performance. In particular, the GNN backbones (JKNet, MixHop, GCN and the proposed method) perform well in processing complex topological features, and their AUROC and AUPR indicators are significantly better than MLP. This shows that the GNN backbone has a stronger ability in encoding and expressing topological features, and can more accurately capture complex vibration signals in construction machinery. At the same time, this also lays a solid foundation for applying graph neural networks to out-of-distribution data detection tasks. Specifically, EGN-out-of-distribution data performs better than other backbones in all experimental groups. This further verifies the effectiveness of our invented energy propagation scheme, which can make full use of the topological information on the graph structure and improve the detection performance and robustness of the model. At the same time, under the MLP backbone, although the detection performance is relatively low, the AUROC and AUPR values ​​are still close to other models using the GNN backbone, indicating that our invention is still competitive even with weaker feature extraction capabilities. This also verifies the effectiveness of the synergy of the modules and the energy propagation mechanism in our invention framework. In contrast, when using JKNet and MixHop as the backbone, the detection performance is significantly improved, indicating that these GNN models have a strong advantage over traditional neural networks in capturing complex fault modes due to their powerful topological feature encoding capabilities and the energy score transfer mechanism between nodes.

[0165] In order to verify the superiority of the maximum mutual information coefficient (MIC) as a similarity metric in the process of converting time series into graph data, a systematic experiment was conducted to compare the effects of three similarity metrics (Pearson coefficient, cosine similarity, and maximum mutual information coefficient) on the composition method. The experimental results are shown in Figure 2. Figure 2 As shown, the impact of different similarity metrics on model performance (AUROC) is demonstrated.

[0166] The results show that the AUROC values ​​of all experimental groups are relatively low when the graph data is constructed by using the Pearson coefficient, a similarity measurement indicator. For example, the AUROC value of the E1 group is only 80.45. This is because the Pearson correlation coefficient is mainly used to measure the linear relationship between two variables. When processing data with a linear relationship, the Pearson coefficient can provide good performance. However, rotating machinery fault data usually exhibits highly nonlinear characteristics, and the Pearson coefficient is difficult to capture these complex patterns and hidden information. In contrast, cosine similarity can capture nonlinear characteristics to a certain extent when measuring the angle relationship between vectors, which makes it advantageous when processing certain types of nonlinear data. For example, the E3 group reached 90.87, but its limitation is that it fails to fully utilize the amplitude information of the data. Finally, the similarity measurement indicator selected by our invention, the maximum mutual information coefficient, as a non-parametric method, can measure any dependency relationship between variables, not just linear or angular relationships, and can capture complex nonlinear relationships and potential patterns in high-dimensional data, which is of great significance for the comprehensive extraction and analysis of features in rotating machinery fault data. The experimental results also clearly show the excellent performance of MIC in each experimental group. This also shows that the prior knowledge of the graph topology can affect the performance of the model. Therefore, it is very important to choose the appropriate topology when performing intelligent diagnosis.

[0167] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for out-of-distribution fault detection based on energy propagation and graph learning, characterized in that: include: The vibration acceleration signals under typical fault conditions are collected, similarity is calculated, and the adjacency matrix consisting of the maximum mutual information coefficient is obtained as the input of the graph neural network. Using the GraphSage graph convolution method, feature extraction is performed on the adjacency matrix to generate a representation of each node; Connect the model between energy function and probability density, obtain the energy function under the GNN framework, calculate the energy score of each node, set the energy score threshold, and distinguish between data within and outside the distribution; Strengthen the confidence estimation of out-of-distribution data for each node, and identify and detect out-of-distribution data under different working conditions of rolling bearings according to the intelligent diagnosis framework of out-of-distribution data based on energy score; The model between the link energy function and the probability density is used to obtain the energy function under the GNN framework. In the deep learning model, the parameter function is used for the K-class classification problem. To achieve, Represents the feature space of the input data x, where each data point x is a D-dimensional vector; Represents the output logits space, each logit vector is a K-dimensional vector, corresponding to K categories; f θ Represents a parameterized mapping function, parameter function f θ Each data point Mapped to k real-valued logits,; After S layers of graph convolution, the GNN model outputs a K-dimensional vector As the logits of each node, it is expressed as: in, represents the logits obtained after each data point passes through S layers of graph convolution, x i represents the input feature vector of the i-th node, where i represents the index of the node; According to the update rule of GNN, represents the function of predicting logits of the S-order ego-graph centered on instance x, and θ represents the trainable parameters in the GNN model f; Input logits into the softmax function to obtain the probability distribution, so that the model can perform classification tasks; given x, Under the condition of , the conditional probability of the target category y is expressed as: in, represents the probability that node x is classified as category y, Represents the neighbor nodes and their connection relationships with node x. Represents a parameterized function f θ When inputting feature x and graph structure information The output of the above example is a K-dimensional vector, where each component corresponds to the logit of a category. express The y-th index of K represents the total number of categories. represents the element with index y, and [k] represents the element with index k; Combining the conditional probability of the target category y with the EBM model that defines the energy function and probability density, the energy function induced by the GNN model is expressed as: in, represents the calculated energy value; The LogSumExp(·) of the logits of the GNN classifier is reused to define the energy function at the data point x, expressed as: in, represents the energy function, f θ (x) [k] Represents the function f θ On input x and The kth element in the vector calculated below, LogSumExp(·) represents the calculation of the log-likelihood function, Represents a parameterized function f θ When inputting feature x and graph structure information The component corresponding to the kth category in the calculated output logit vector; The training loss is expressed as: in, represents the expected value operator, represents the distribution of the labeled part of the training data. Represents a node with a label, represents the supervised training loss, Represents a parameterized function f θ When inputting feature x and graph structure information In the calculated output logit vector, the y i The weight corresponding to each category, Represents a parameterized function f θ In the input feature x i and graph structure information The component corresponding to the kth category in the calculated output logit vector; Set an energy score threshold, calculate the energy score of each node according to the energy function under the GNN framework, and distinguish between in-distribution and out-of-distribution data; If the energy score is lower than the threshold, it belongs to the in-distribution data; if it is higher than the threshold, it belongs to the out-of-distribution data.

2. The out-of-distribution fault detection method based on energy propagation and graph learning according to claim 1, characterized in that: The vibration acceleration signals under the typical fault states include healthy H, outer ring fault F1, inner ring fault F2, rolling element fault F3, outer ring-rolling element composite fault F4 and outer ring-inner ring-rolling element composite fault F5.

3. The out-of-distribution fault detection method based on energy propagation and graph learning as claimed in claim 2, characterized in that: The similarity calculation includes collecting vibration acceleration signals B=[b1, b2, ..., b n ] and C=[c1,c2,...,c n ], discretize and convert into discrete variables; Calculate the mutual information between B and C, and select the variable pair with the largest mutual information as a candidate; the mutual information I(B,C) is expressed as: Where I(B,C) represents the mutual information coefficient, p(b,c) represents the joint probability distribution function of time series B and C, p(b) represents the marginal probability density function of B, p(c) represents the marginal probability density function of C, b i represents the i-th value of time series B, c i represents the i-th value of the time series C, i = 1, 2, ..., n; b represents the value of the discretized variable B, and c represents the value of the discretized variable C; The maximum mutual information coefficient is used as the similarity evaluation index. The maximum mutual information coefficient value is obtained by normalizing the candidate variables. The statistic is defined as a triple (B, C, MIC), where B and C represent the discretized versions of the input variables, and MIC represents the corresponding maximum mutual information coefficient, which is expressed as: Among them, MIC(B,C) represents the maximum mutual information coefficient, represents the number of grids divided in the B and C directions, and ξ represents the selection parameter; The adjacency matrix formed by the maximum mutual information coefficient has a dimension of N×N, which is used as the input of the graph neural network to convert the one-dimensional time series data into graph data that can be recognized by the graph neural network model, expressed as: Among them, ρ represents the adjacency matrix, p NN represents the maximum mutual information coefficient between two time points at the Nth moment of time series B and C; p ij Represents the maximum mutual information coefficient between node i and node j, which is a similarity measure between two nodes. As the weight coefficient of each edge, i = 1, 2, ..., N, j = 1, 2, ..., N.

4. The out-of-distribution fault detection method based on energy propagation and graph learning as claimed in claim 3, characterized in that: The GraphSage graph convolution method is used to extract features from the adjacency matrix, and each node representation is generated, including generating node representations by sampling and aggregating features of neighboring nodes, and introducing a weighted average attention mechanism to weight features of neighboring nodes to highlight important feature information; Assumptions represents the node representation of node v at the sth layer, Represents the neighbor set of node v, and the node representation update formula of the s+1th layer is expressed as: in, represents the node representation of node v in the s+1th layer, σ represents the nonlinear activation function, and W (s) represents the weight matrix of the sth layer, It indicates that node u is in the sth layer, and Aggregate indicates the mean aggregation function, which is expressed as: The obtained attention weight is normalized by the Softmax function to get α vu , feature aggregation is performed through the weighted average attention mechanism, and the updated node is represented as: Among them, α vu represents the result after normalization of attention weights, W represents the weight matrix, and h represents the node representation.

5. The out-of-distribution fault detection method based on energy propagation and graph learning as claimed in claim 1, characterized in that: The strengthening of out-of-distribution data confidence estimation for each node includes defining As a node v in the graph i The initial energy fraction vector, E (0) Represents the node v in the graph i The initial energy fraction vector of The energy function is used to calculate the energy value of the node, which is expressed as follows through the propagation update rule: Among them, 0<λ<1, which represents the parameter that controls the energy transfer concentration between the node itself and other connected nodes; E (s) Represents the energy score vector of the s-th layer node, E (s-1) represents the energy score vector of the s-th layer node, D represents the degree matrix, A represents the adjacency matrix, Represents all node domains, E i (s) represents the energy value of the i-th node after the s-th iteration; According to the propagation rules of the updated nodes, after S steps of propagation, the final result after energy transfer is obtained in, It represents the final energy value of node i after S-step propagation iterations, and performs out-of-distribution data detection and judgment on the result after the final energy transfer; The energy transfer in the graph topology uses the energy values ​​of neighboring nodes to strengthen the out-of-distribution data confidence estimate of each node. For any given data sample point x i , if the average energy score of its one-hop neighbor node is Than its own energy fraction Low, expressed as: According to the propagation update rule E (s) get: Among them, E i (s) represents the node energy score vector of node i in layer s, a ij represents the weight of the direct edge connecting node i and node j, E i (s-1) Represents the node energy score vector of node i in the s-1 layer; The decision criterion of the out-of-distribution data discriminator is expressed as: in, represents a binary classification function, which determines whether the input x belongs to the data within the distribution or outside the distribution. ε represents the judgment threshold. Represents the result after final energy transfer.

6. The out-of-distribution fault detection method based on energy propagation and graph learning as claimed in claim 5, characterized in that: The identification of out-of-distribution data and detection of out-of-distribution data under different working conditions of rolling bearings according to the out-of-distribution data intelligent diagnosis framework based on energy scores includes four steps: graph construction, feature encoding, energy score detection and energy propagation update; The graph construction includes converting the vibration signals collected by the sensors into graph data using the maximum mutual information coefficient and constructing an adjacency matrix reflecting the nonlinear relationship between faults; Feature encoding includes using the GraphSage model and the weighted average attention mechanism to extract features from graph data; GraphSage generates node representations by aggregating neighbor node features, and the attention mechanism strengthens the weights of important neighbor features to enhance the representation ability of the model; Energy score detection includes introducing out-of-distribution data discrimination based on energy scores, establishing the energy function of data points by redefining the logits of the classifier; distinguishing in-distribution data from out-of-distribution data by calculating the energy scores, and realizing effective out-of-distribution data fault detection; The energy propagation update includes improving the generalization ability of the model in a semi-supervised learning environment, and proposing an energy score propagation mechanism based on a graph structure; by iteratively propagating node energy scores on the graph, using neighbor node information to update the energy value, strengthening the energy difference between in-distribution data and out-of-distribution data, and improving the accuracy of out-of-distribution data detection.

7. An out-of-distribution fault detection system based on energy propagation and graph learning using the method according to any one of claims 1 to 6, characterized in that: The acquisition module collects the vibration acceleration signals under typical fault conditions, performs similarity calculation, and obtains the adjacency matrix composed of the maximum mutual information coefficient as the input of the graph neural network; The feature extraction module uses the GraphSage graph convolution method to extract features from the adjacency matrix and generate a representation of each node; The energy score calculation module connects the model between the energy function and the probability density, obtains the energy function under the GNN framework, calculates the energy score of each node, sets the energy score threshold, and distinguishes between data within and outside the distribution; The optimization module strengthens the confidence estimation of the out-of-distribution data of each node, and identifies and detects the out-of-distribution data of different working conditions of rolling bearings according to the intelligent diagnosis framework of out-of-distribution data based on energy score.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images