Intelligent evaluation method for health state of complex equipment based on graph metric learning

Abstract

A kind of complex equipment small sample health state intelligent evaluation method based on graph metric learning, comprising the following steps: 1) according to prior knowledge, the component data of equipment, sensor monitoring data are fused into graph data model;2) the structural features and node features contained in small sample graph data are embedded into low-dimensional vector representation space, to obtain the graph vector representation of sample;3) the graph vector representation and small sample label information are fused, to obtain the updated graph representation vector;4) the graph vector representation is mapped to fully connected graph, and the metric matrix and feature matrix are iteratively updated using graph metric learning, to transfer the knowledge on the labeled graph sample to the unlabeled graph sample;5) the prediction label distribution of unlabeled graph sample, i.e. health state evaluation result, is obtained using Softmax layer.The present application fully excavates the potential correlation features between data, and significantly improves the accuracy of equipment health state evaluation under the condition of sample sparsity.

Description

Technical Field

[0001] This invention relates to the field of health status assessment of complex equipment, and to an intelligent assessment method for the health status of complex equipment with small samples based on graph metric learning. Background Technology

[0002] The scientific and effective monitoring of the health status of complex equipment and its core components (such as aircraft engines, elevator traction machines, and bearings) is of significant scientific importance and practical value for ensuring the safe operation of complex equipment and reducing maintenance costs. Health status assessment of complex equipment is one of the key technologies of Diagnostic, Predictive, and Health Management (DPHM), aiming to utilize the large amount of monitoring data collected by sensors and domain knowledge to conduct qualitative or quantitative assessments.

[0003] Existing intelligent assessment methods for the health status of complex equipment with small samples are mainly divided into data generation-based methods and meta-learning-based methods. Data generation-based methods for intelligent assessment of the health status of complex equipment with small samples primarily include synthetic minority class oversampling, generative adversarial networks (GANs), and variational autoencoders (VAEs), which assess health status by expanding the amount of data and improving its quality. Synthetic minority class oversampling can discover similar samples within the minority class and generate new minority class samples through random linear interpolation, expanding and balancing the sample size. The original and synthetic samples are then used together as the training set for the classification model. Wei et al. (Engineering Applications of Artificial Intelligence, 2020) proposed a novel imbalanced fault diagnosis framework that uses Cluster-MWMOTE and MFO-optimized LS-SVM to diagnose faults in finite and complex bearing data, further improving the model's adaptability to sample imbalance. Generative adversarial networks (GANs) are networks based on the zero-sum game concept. Through continuous evolutionary interaction between the discriminator and generator, they generate high-quality auxiliary samples with real sample features, improving the accuracy of the classification model. Dixit et al. (IEEE Transactions on Instrumentation and Measurement, 2021) combined meta-learning with Generative Adversarial Networks (CACGANs) for auxiliary classifiers. Meta-learning was used to extract network parameters, and CACGAN was used to generate auxiliary samples, improving the accuracy of health assessment for small samples. Variational autoencoders are also a commonly used deep generative model. By learning the latent variables of the input data through autoencoders and processing them with a normal distribution before resampling, auxiliary data similar to but different from the original data can be generated, which to some extent solves the problems of difficult training and measurement of training progress in GANs. Wang et al. (AppliedSoft Computing, 2020) proposed a fault diagnosis method based on VAEs and GANs. Fault data is generated through VAEs, and network parameters are continuously optimized using GANs, improving classification accuracy in imbalanced data.

[0004] Meta-learning-based methods model small samples by learning key experiences and knowledge from previous tasks to quickly adapt to small sample models for new tasks. These methods mainly include fine-tuning-based meta-learning and metric-based meta-learning. Fine-tuning-based meta-learning first learns initial parameters from a large number of known tasks, and then quickly learns sample features from a small amount of data in new tasks. Hu et al. (IEEE Transactions on Industrial Informatics, 2021) proposed a meta-learning task ranking model to gradually improve the adaptability of knowledge to tasks, and finally learns the optimal initialization parameters through an improved model-independent meta-learning model, demonstrating the effectiveness of the method on small sample power fault data and bearing data. Metric-based meta-learning maps small sample feature information to the same feature space, making similar samples as close as possible and dissimilar samples as far apart as possible. It determines the label of unknown samples by learning the similarity or distance between unknown and known samples, achieving unlabeled sample classification. Wang et al. (Mechanical Systems and Signal Processing, 2021) proposed a feature space metric-based meta-learning model, which, by combining supervised learning and metric learning, utilizes the attribute information of individual samples and the similarity of sample groups to improve the accuracy of small sample fault diagnosis.

[0005] Data generation-based methods heavily rely on the quality of both the original and generated samples; generating a large number of invalid samples can easily lead to blurred boundaries in health assessments. Meta-learning-based methods do not consider the relationship between monitoring data and component data, and without prior knowledge, they struggle to accurately extract data features. This invention proposes a small-sample health assessment method for complex equipment based on graph metric learning. This method integrates existing detection data and prior knowledge, and uses graph metric learning to assess the health status of complex equipment. Summary of the Invention

[0006] To deeply extract the intrinsic information and structural features contained in small sample data and address the problem that deep learning-based health status assessment methods for complex equipment often rely on large amounts of labeled data, this invention proposes a small sample health status assessment method for complex equipment based on graph metric learning. This method deeply integrates prior knowledge, sensor monitoring data, and component data, and uses a graph metric learning network to transfer the feature information of labeled graph vector representations to the graph vector representations to be classified. Through knowledge transfer, the classification result of the vector representations to be classified is obtained.

[0007] To solve the technical problem, the present invention adopts the following technical solution:

[0008] A method for intelligent assessment of the health status of complex equipment in small samples based on graph metric learning includes the following steps:

[0009] 1) Data fusion of health status of complex equipment, i.e., graph data modeling;

[0010] 2) Obtain the graph vector representation of the samples by multi-set pooling of the health status graph of complex equipment;

[0011] 3) Small sample label information fusion: The graph vector representation obtained in the first step is fused with the label information to obtain a new graph representation vector containing label information;

[0012] 4) Map the graph vector representation to a fully connected graph, and use graph metric learning to iteratively update the metric matrix and feature matrix, gradually transferring from labeled graph samples to unlabeled graph samples;

[0013] 5) The predicted label distribution of the unlabeled map samples is obtained by using the Softmax layer, which is the health status assessment result.

[0014] Furthermore, in step 1), the complex equipment health status data fusion process is as follows:

[0015] Based on prior knowledge within the system, component data and sensor monitoring data of complex equipment are fused together to perform graph data modeling. Using raw data as nodes and dependencies between data as edges, various types of relationships between prior knowledge, sensor monitoring data, and component data are modeled in a natural way, effectively integrating potential information affecting the health status of complex equipment.

[0016] Definition 1: Graph data: Graph data is represented as G = (V, E), where V = (v1, v2, ..., v n ), Let the node set and edge set be represented respectively, and the feature matrix and adjacency matrix be respectively. and A∈{0,1} N×N ,in It is node v i The characteristic is that (v) i ,v j When A ∈ ij =1.

[0017] Furthermore, in step 2), the multi-set pooling process of the complex equipment health status graph is as follows:

[0018] Using graph-structured data as input, the similarity and correlation of node features are obtained through the graph multi-head attention mechanism, forming clusterable node clusters, i.e., multi-sets. The multi-sets are compressed, and the nodes in the set are aggregated into a new supernode. The aggregated supernodes constitute a new graph. The supernodes learn the internal relationships of the supernodes through a self-attention module, resulting in a new supernode vector representation.

[0019] The graph multi-pooling process is represented as follows:

[0020] Pooling(H,A)=GMPool2(SelfAtt(GMPool1(H,A)),A′) (2)

[0021] Wherein, GMPool1 is the first-level graph pooling layer; SelfAtt is the self-attention layer, which can accurately obtain the association relationship between super nodes; GMPool2 is the second-level graph pooling layer, and its output is a graph-level feature representation; H is the feature vector of the node; A is the adjacency matrix; A′ is the adjacency matrix of the super node after the first graph pooling. Step 2) includes the following sub-steps:

[0022] 2.1) Input the graph structure data into the graph convolutional network layer to learn its node feature information and topological structure information, and obtain the vector representation of the node;

[0023] 2.2) After the first-level graph pooling layer, it takes the node vector representation as input, divides the nodes into multiple sets according to the principles of similarity and relevance, and then pools the features of the nodes in these multiple sets to obtain a new vector representation of the supernode.

[0024] 2.3) After passing through a self-attention layer, the intrinsic relationships between supernodes are further learned, and the feature representations of new supernodes are updated; the self-attention function between nodes is:

[0025] SelfAtt(H)=LN(Z+rFF(Z)); Z=LN(H+MH(H,H,H)) (1)

[0026] Where FF is the feedforward layer, LN is the layer normalization function, and MH is the multi-head attention function.

[0027] 2.4) Input the feature representation of the new supernode into the second-level graph pooling layer block, and further aggregate the feature representation of the supernode to generate the feature representation of the entire graph. This feature representation is the graph vector representation of the graph structure data.

[0028] Furthermore, in step 3), the process of fusing small sample label information is as follows:

[0029] The graph-level vector representation obtained in step 1) is fused with the label information of the small sample data to obtain a new graph representation vector containing label information, which is then input into the downstream graph metric learning layer; for a known label l i Image sample x i ∈T, the new graph vector representation is calculated using the following formula:

[0030]

[0031] Where, f(x) i) is the graph vector representation of the i-th sample, h(l) i ) is a tag l i one-hot vector, It is a new graph vector representation;

[0032] For the sample to be classified, since its label is unknown, h(l) in formula (3) will be used. i Replace ) with an average function that is uniformly distributed over the number of categories K. Considering the uncertainty of unknown labels using average probability, the calculation formula is as follows:

[0033]

[0034] By fusing graph representation learning and label information, graph samples can be mapped into graph-level vector representations, which are then input into the downstream graph metric learning layer. Knowledge transfer is performed through labeled graph samples to learn the category of the sample to be classified.

[0035] Step 4) includes the following sub-steps:

[0036] 4.1) Map the vector representations of labeled and unlabeled graphs to a fully connected graph, using the graph vector representation as the initial node features and the similarity measurement results between nodes as edges;

[0037] 4.2) Obtain the graph metric matrix based on the metric function. The graph metric matrix is ​​an adjacency matrix with metric information, where the values ​​of the matrix elements represent vectors that represent the similarity between nodes.

[0038] 4.3) Perform softmax processing on each row of the graph metric matrix to ensure that the sum of the weights of each node and all other nodes is 1;

[0039] 4.4) A graph convolutional neural network is used to iteratively aggregate and update the feature information of the vector-represented nodes.

[0040] The graph metric learning layer aims to calculate the similarity between graph vector representations. It iteratively trains and updates the graph vector representations and metric matrix using a metric function, bringing similar vector representations closer together and dissimilar vector representations further apart. In this model, a fully connected graph is constructed using graph vector representations as nodes and the similarity metric results between graph vector representations as edges. A graph convolutional network is used to aggregate and update the vector representations, and the graph metric matrix is ​​iteratively updated. Feature information from labeled graph vector representations is transferred to the graph vector representation to be classified, and the classification result of the vector representation to be classified is obtained through knowledge transfer.

[0041] A graph metric matrix is ​​an adjacency matrix that contains metric information. The values ​​of its elements represent vectors, and the vectors represent the similarity between nodes. The formula for calculating this similarity is:

[0042]

[0043] Where k represents the number of layers in the graph metric learning process. Let vector represent the feature representation of node i and vector represent the feature representation of node j, and function It is a parameterized symmetric function used to measure the similarity between nodes. To ensure the symmetry and reasonableness of the distance attribute, this function satisfies the following two conditions:

[0044]

[0045] function Using a multilayer perceptron, the calculation formula is as follows:

[0046]

[0047] in, Let represent the similarity between two vectors representing nodes i and j. A Multilayer Perceptron (MLP) is a trainable distance function used to learn the absolute distance between the two vectors representing nodes. This function increases its expressive power through multiple hidden layers and can be combined with the activation function of a neural network to further fit and obtain a more accurate metric relationship. Furthermore, each row of the graph metric matrix undergoes softmax processing to ensure that the sum of the weights of each node and all other nodes is 1.

[0048] Graph vector representation uses a graph convolutional neural network to iteratively aggregate and update the feature information of the nodes in the vector representation. The calculation formula is as follows:

[0049]

[0050] in, Let A represent the vector representation of a node in layer k+1, where l represents the vector length, ρ represents the pointwise Leaky ReLU nonlinear activation, and θ represents a learnable weight matrix. In the metric matrix A, the connection weight between a node and itself is 1, while the connection weight with other nodes is 0.

[0051] Step 5) includes the following sub-steps:

[0052] 5.1) Based on the graph metric learning results, the feature vectors of the graph samples to be classified are... The probabilities y mapped to K categories through the softmax layer k :

[0053]

[0054] in, It is the vector representation of the graph sample to be classified. The softmax() function maps it to the range [0, 1] based on the implicit features of the graph vector representation and outputs the probability of the health status category.

[0055] 5.2) Combining the probability y calculated above k The cross-entropy loss function is used to calculate the loss between the predicted and actual results, and the Adam optimizer is used to minimize the loss function. The calculation formula is as follows:

[0056] loss(φ(T;Θ),Y)=-∑ k y k logP(Y * =y k |T) (9)

[0057] Where φ(T; Θ) represents the few-shot learning task, y k This is the health status prediction result of the image to be classified, Y. * It represents the true health status of the image to be classified.

[0058] 5.3) Continuously iterate the above training process until the vector representation of the health status of complex equipment converges, obtaining the final small-sample complex equipment health status assessment model. Input the complex equipment graph data to be assessed into the model to obtain the classification results of the complex equipment health status.

[0059] After completing all the above steps, the health status assessment of complex equipment is finished.

[0060] The beneficial effects of this invention are as follows: Addressing the problem that traditional few-sample learning models struggle to utilize prior knowledge and extract correlation information between component data and monitoring data, this paper proposes a graph metric learning-based intelligent assessment method for the health status of complex equipment in small samples. This method utilizes graph representation learning to embed graph data into a low-dimensional vector representation space, and then uses graph metric learning to analyze the similarity relationships between small-sample graph data in the vector space. Furthermore, it achieves knowledge transfer of health features through graph metric learning. Combining graph neural networks and metric learning significantly improves the accuracy of health status assessment, providing technical support for predictive maintenance. Attached Figure Description

[0061] Figure 1 This is a diagram of the architecture of a few-sample graph metric learning model.

[0062] Figure 2 This is a data model construction diagram of the health status of an aero-engine, where 1 is the fan, 2 is the combustion chamber, 3 is the high-pressure rotor, 4 is the low-pressure turbine, 5 is the low-pressure compressor, 6 is the high-pressure compressor, 7 is the low-pressure rotor, 8 is the high-pressure turbine, and 9 is the nozzle. Detailed Implementation

[0063] Combining a specific example of an aero-engine graph metric learning network, and referring to... Figure 1 The invention will be further illustrated below.

[0064] Reference Figure 1 and Figure 2 A method for intelligent assessment of the health status of complex equipment in small samples based on graph metric learning includes the following steps:

[0065] 1) Fusion of aircraft engine health status data

[0066] Based on prior knowledge related to aero-engines, component data and sensor monitoring data of aero-engines are fused into graph-structured data to facilitate subsequent use of graph models to explore potential relationships between the data. Figure 2 A health status diagram data model for aero-engines was constructed, taking aero-engine components as an example.

[0067] 2) Graph vector representation of aero-engine data

[0068] The pre-constructed graph sample dataset T is used as the model input, containing labeled graph samples belonging to different health state categories (K=4 is the number of health state categories, x1, x2, x3, and x4 are labeled graph samples for each health state category) and a graph sample to be classified (t=1, x...). q (For the graph sample to be classified), the feature information of the input graph sample data is embedded into a low-dimensional vector representation using graph convolution learning to obtain the node vector representation of the sample. According to formula (1), the similarity and correlation of node features are obtained using the graph multi-head attention mechanism to form a clusterable node cluster, i.e., a multi-set. The multi-set is compressed and the nodes in the set are aggregated into a new supernode. The aggregated supernode constitutes a new graph. The supernode learns the internal correlation of the supernode through a self-attention module to obtain the graph vector representation of the aero-engine data sample.

[0069] 3) Tag information fusion

[0070] The graph vector representation of the aero-engine data and the label information are fused together. According to formula (3), a new graph representation vector containing label information is obtained. For the sample to be classified, as described in formula (4), since its label is unknown, the average probability label is used as the initial value. Through graph representation learning and label information fusion, the graph sample can be mapped into a graph-level vector representation. It is then input into the downstream graph metric learning layer. Knowledge transfer is performed through the labeled graph sample to learn the category of the sample to be classified.

[0071] 4) Graph Metric Learning

[0072] The vector representations of labeled and unlabeled graphs are mapped to a fully connected graph. Graph vector representations are used as nodes, and the similarity measurement results between graph vector representations are used as edges. A graph convolutional network is used to aggregate and update the vector representations, and a metric algorithm is used to iteratively update the graph metric matrix. According to the requirements of formula (6), the metric function is set as formula (7), and the absolute value function is used to measure the distance between two nodes. Finally, a new graph metric matrix, i.e., the adjacency matrix, is calculated using formula (5). Then, softmax processing is performed on each row of the graph metric matrix to ensure that the sum of the weights between each node and all other nodes is 1. A graph convolutional neural network is used to iteratively aggregate and update the feature information of the vector representation nodes. The metric representation update and vector representation update are repeated twice to transfer the feature information of the labeled graph vector representation to the graph vector representation to be classified.

[0073] 5) Health status assessment

[0074] First, based on the graph metric learning results, the feature vectors of the graph samples to be classified are... The probability of the aircraft engine health status is mapped to the K categories through the softmax layer of formula (9), and then the loss function of formula (10) is used to reduce the loss between the probability calculated in formula (9) and the correct health status label, so that the calculated classification result continuously approaches the correct label, thereby making the aircraft engine health status representation vector more accurate.

[0075] Based on this, the training process is continuously iterated to reduce the loss, so that the embedding of the aircraft engine health status converges, and the final aircraft engine health status classification result is obtained.

[0076] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for intelligent assessment of the health status of complex equipment in small samples based on graph metric learning, characterized in that... The method includes the following steps: 1) Data fusion of health status of complex equipment, i.e., graph data modeling; 2) Obtaining graph vector representations of samples through multi-set pooling of complex equipment health status graphs: Using graph structure data as input, the similarity and correlation of node features are obtained using a graph multi-head attention mechanism to form clusterable node clusters, i.e., multi-sets; compression is performed on the multi-sets, and the nodes in the sets are aggregated into new supernodes, which constitute a new graph; the supernodes are then processed by a self-attention module to learn the inherent relationships between supernodes, resulting in a new supernode vector representation; including the following sub-steps: 2.1) Input the graph structure data into the graph convolutional network layer to learn its node feature information and topological structure information, and obtain the vector representation of the node; 2.2) After the first-level graph pooling layer, it takes the node vector representation as input, divides the nodes into multiple sets according to the principles of similarity and relevance, and then pools the features of the nodes in these multiple sets to obtain a new vector representation of the supernode. 2.3) After passing through a self-attention layer, the intrinsic relationships between supernodes are further learned, and the feature representations of new supernodes are updated; the self-attention function between nodes is: （1）； in, H Let be the feature vector of the node, FF be the feedforward layer, LN be the layer normalization function, and MH be the multi-head attention function; 2.4) The feature representation of the new supernode is input into the second-level graph pooling layer block, and the feature representation of the supernode is further aggregated to generate the feature representation of the entire graph. This feature representation is the graph vector representation of the graph structure data. 3) Small sample label information fusion: The graph vector representation obtained in the first step is fused with the label information to obtain a new graph representation vector containing label information; 4) Map the graph vector representation to a fully connected graph, and use graph metric learning to iteratively update the metric matrix and feature matrix, gradually transferring from labeled graph samples to unlabeled graph samples; 5) The predicted label distribution of the unlabeled map samples is obtained by using the Softmax layer, which is the health status assessment result.

2. The intelligent assessment method for the health status of complex equipment in small samples based on graph metric learning according to claim 1, characterized in that, The process of step 1) is as follows: Based on prior knowledge in the system, component data and sensor monitoring data of complex equipment are fused together, i.e., graph data modeling is performed. The original data is used as nodes and the dependencies between data are used as edges. Various types of relationships between prior knowledge, sensor monitoring data and component data are modeled in a natural way, effectively integrating potential information that affects the health status of complex equipment.

3. The intelligent assessment method for the health status of complex equipment in small samples based on graph metric learning according to claim 1, characterized in that, In step 2), the calculation formula using graph pooling and self-attention networks is as follows: （2）； Among them, GMPool1 is the first-level graph pooling layer; SelfAtt is the self-attention layer, which can accurately obtain the relationship between super nodes; GMPool2 is the second-level graph pooling layer, whose output is a graph-level feature representation. A It is an adjacency matrix; It is the adjacency matrix of the supernodes after the first graph pooling; It is a vector representation of graph-structured data.

4. The intelligent assessment method for the health status of complex equipment in small samples based on graph metric learning according to claim 3, characterized in that, The process of step 3) is as follows: The graph-level vector representation obtained in step 1) is fused with the label information of the small sample data to obtain a new graph representation vector containing label information, so as to be input into the downstream graph metric learning layer; To the known label Image sample ∈ T The new graph vector representation calculation formula is: （3）； in, It is the first i The graph vector representation of each sample, It is a tag one-hot vector, It is a new graph vector representation; For the sample to be classified, since its label is unknown, the formula (3) will be used. Replace with a number of categories K Average function of uniform distribution Taking into account the uncertainty of unknown labels with average probability, the calculation formula is as follows: （4）。 5. The intelligent assessment method for the health status of complex equipment in small samples based on graph metric learning according to claim 4, characterized in that, Step 4) includes the following sub-steps: 4.1) Map the vector representations of labeled and unlabeled graphs to a fully connected graph, using the graph vector representation as the initial node features and the similarity measurement results between nodes as edges; 4.2) Obtain the graph metric matrix based on the metric function. The graph metric matrix is ​​an adjacency matrix with metric information, where the values ​​of the matrix elements represent vectors that represent the similarity between nodes. 4.3) Perform softmax processing on each row of the graph metric matrix to ensure that the sum of the weights of each node and all other nodes is 1; 4.4) A graph convolutional neural network is used to iteratively aggregate and update the feature information of the vector-represented nodes.

6. The intelligent assessment method for the health status of complex equipment in small samples based on graph metric learning according to claim 5, characterized in that, In step 4), the formula for calculating the graph metric matrix is: （5）； in, k This indicates the number of layers in the graph metric learning process. Nodes are represented by vectors i The sum vector represents the node j Feature representation, function It is a parameterized symmetric function used to measure the similarity between nodes. To ensure the symmetry and reasonableness of the distance attribute, this function satisfies the following two conditions: （6）； function Using a multilayer perceptron, the calculation formula is as follows: （7）； in, Two vectors represent nodes. i, j The similarity is obtained by using a multilayer perceptron (MLP) as a trainable distance function to learn the absolute distance between two vectors representing nodes. This function increases its expressive power through multiple hidden layers and can be combined with the activation function of a neural network to further fit and obtain a more accurate metric relationship.

7. The intelligent assessment method for the health status of complex equipment in small samples based on graph metric learning according to claim 6, characterized in that, In step (4), the graph vector representation uses a graph convolutional neural network to iteratively aggregate and update the feature information of the vector representation nodes. The calculation formula is as follows: （8）； in, express k+ Vector representation of a level 1 node. l Indicates the length of the vector. ρ This represents the pointwise Leaky ReLU nonlinear activation process. Represents a learnable weight matrix; metric matrix A A node's connection weight to itself is 1, while its connection weight to other nodes is 1. .

8. The intelligent assessment method for the health status of complex equipment in small samples based on graph metric learning according to claim 7, characterized in that, Step 5) includes the following sub-steps: 5.1) Based on the graph metric learning results, the feature vectors of the graph samples to be classified are... Mapped to via the softmax layer K Probability in the category : （9）； in, It is the vector representation of the image sample to be classified. The () function maps the implicit features represented by the graph vector to the range [0, 1] and outputs the probability of the health status category; 5.2) Combining the probabilities calculated above The cross-entropy loss function is used to calculate the loss between the predicted and actual results, and the Adam optimizer is used to minimize the loss function. The calculation formula is as follows: (10)； in, Represents a few-shot learning task. These are the health status prediction results for the image to be classified. It represents the true health status of the image to be classified; 5.3) Continuously iterate the above training process until the vector representation of equipment health status converges, and obtain the final small sample equipment health status assessment model. Input the equipment graph data to be assessed into the model to obtain the equipment health status classification result.

Images