Direct current distribution board health state assessment method based on multiple electric parameters
By using a graph attention network with multiple electrical parameters and a variational inference module to assess the health status of the DC power distribution board, the problem of insufficient robustness under small sample conditions is solved, and high-precision fault diagnosis and status identification are achieved, ensuring the stable operation of the ship's power system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-04-07
AI Technical Summary
Existing methods for assessing the health of DC switchboards are not robust enough under small sample conditions, have difficulty effectively utilizing unlabeled data, and have poor multimodal data fusion performance, making them unable to adapt to the fluctuations in electrical parameters under complex ship environments.
A health status assessment method with multiple electrical parameters is adopted. Multimodal fusion of unlabeled monitoring data is performed through graph attention network and variational inference module to construct spatiotemporal graph structure. Variational lower bound loss is used to measure the similarity of potential distributions. The method is trained with a small amount of labeled data and the classification module is optimized using cross-entropy loss function.
It improves the accuracy and stability of DC power distribution board health status assessment, can automatically identify abnormal conditions and analyze the causes, extend equipment service life, reduce maintenance costs, and ensure ship operation safety.
Smart Images
Figure CN121808545A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of ship power systems, in particular to a DC distribution board health state evaluation method based on multiple electrical parameters. BACKGROUND
[0002] As a new type of power distribution device in ship power systems, the DC distribution board has the advantages of fast response speed, high distribution efficiency, small cabin space occupation, strong running stability, etc., and can effectively support the power demand of the ship. Due to the complexity and variability of the ship operating environment, such as ocean vibration, salt spray corrosion and load fluctuation, the DC distribution board is bound to fail, affecting the safe and stable operation of the ship power system, and even endangering the safety of the ship. Therefore, a high-precision and robust DC distribution board fault diagnosis method is crucial in the field of ships, which can effectively avoid equipment downtime, reduce maintenance costs, and ensure the safety of ship operation.
[0003] At present, there are mainly two methods for domestic and foreign DC distribution board health evaluation: the first method is based on traditional signal processing, which obtains electrical parameters such as voltage and current through sensors, and manually extracts features to establish a mapping between features and health states; the second method is based on deep learning, which designs a neural network to automatically mine abnormal patterns from massive data for end-to-end evaluation. The existing evaluation methods require a large amount of labeled data, but due to the high reliability of the DC distribution board, it is difficult to collect enough fault samples.
[0004] In existing related patent technologies, such as patent CN118445699A, a complex equipment small sample health state intelligent evaluation method based on graph metric learning is disclosed, which realizes evaluation through steps such as graph data modeling, graph vector representation, and label fusion, but does not introduce a variational inference module, which cannot effectively learn the latent feature distribution using unlabeled data; it processes unknown label samples using the uniform distribution assumption, lacks the mining of the inherent distribution of data, and lacks robustness in small sample scenarios; at the same time, this method does not design a special fusion strategy for the multiple electrical parameter characteristics of the DC distribution board, and has poor adaptability to electrical parameter fluctuations in complex ship environments.
[0005] In summary, the existing technology generally has the technical problems of poor small sample adaptability, poor multi-modal data fusion effect, insufficient use of unlabeled data, and insufficient adaptation to complex ship environments, and there is an urgent need for a DC distribution board health state evaluation method that can achieve high precision and high stability evaluation under small sample conditions. SUMMARY
[0006] The purpose of the present application is to provide an intelligent evaluation method that can handle small sample problems with high precision, thereby solving the technical problem that existing DC distribution board health evaluation methods cannot accurately identify abnormal states under small samples. The present application is suitable for DC distribution boards, and the technology can automatically monitor whether the DC distribution board appears abnormal during operation, analyze the abnormal reason, identify the health state type, and provide guidance and suggestions for maintenance personnel.
[0007] To achieve the above-mentioned purpose, the technical solution of the present application is: a DC distribution board health state evaluation method based on multiple electrical parameters, comprising the following steps: S1, collecting multiple electrical parameter unlabeled monitoring data of the DC distribution board, performing multi-modal fusion on the unlabeled monitoring data, and constructing a space-time graph structure; S2, inputting the space-time graph structure into a graph attention network for feature extraction to obtain high-dimensional space-time dependent features; S3, inputting the high-dimensional space-time dependent features into a variational inference module, and utilizing a variational lower bound loss to measure the similarity of the latent distribution; S4, taking the variational loss obtained in step S3 as the optimization objective of the graph attention variational network, updating the parameters of the graph attention network and the variational inference module, and saving the parameters of the graph attention network after training; S5, collecting a small amount of labeled monitoring data of the DC distribution board, extracting features through the graph attention network, and predicting by using a classification module, and utilizing a cross-entropy loss function to measure the error between the predicted label and the true label; S6, freezing the parameters of the graph attention network, taking the prediction error in step S5 as the optimization objective, and updating the parameters of the classification module; S7, utilizing the trained graph attention network and classification module to evaluate the health state of the DC distribution board monitoring data to be tested, and outputting the evaluation result.
[0008] Further, in step S1, the multi-modal fusion of the monitoring data includes voltage, current and power, and the sampling points of each sample are 1024; different parameters are integrated into a graph structure through a multi-modal fusion method. The fusion method is a cross-modal attention and space-time graph construction.
[0009] Further, the cross-modal attention is used to weight and fuse different electrical parameters, and the calculation formula is represented as:
[0010] In the formula, is the fused feature, is the attention weight, is the feature of the i-th mode, and exp is the exponential function.
[0011] Further, in step S1, the space-time graph structure is constructed in the following manner: the fused features are regarded as nodes, the time and space dependencies are regarded as edges, and the graph is represented as G=(V,E), wherein V is the node set and E is the edge set.
[0012] Further, in step S2, the backbone of the graph attention network is a Graph Attention Network (GAT), and the network adaptively aggregates neighbor node information through an attention mechanism, including an input graph G=(V, E), node features H, updated features H' output after attention calculation, wherein: The attention coefficient calculation process is as follows:
[0013] In the formula, is the attention coefficient of node i to j, is a linear transformation matrix, is an attention vector, and LeakyReLU is an activation function; The node update is:
[0014] In the formula, is a neighbor set of node i, is an activation function.
[0015] Further, in step S3, the features are input into the variational inference module, and the similarity of the latent distribution is measured by using a variational lower bound loss, including: in the variational inference module, it is assumed that the latent variable follows a Gaussian distribution, the posterior distribution q(z|x) is approximated by an encoder, and the KL divergence is minimized by maximizing the evidence lower bound (ELBO); wherein, the variational lower bound loss calculation formula is represented as:
[0016] In the formula, is a variational lower bound, is an expectation function, is a KL divergence, is a prior distribution.
[0017] Further, in step S4, the variational loss obtained in S3 is used as the optimization target of the graph attention variational network to update the parameters of the graph attention network and the variational inference module, including: taking the variational loss as the optimization object, using the gradient descent algorithm to train the graph attention variational network, updating the parameters of the graph attention network and the variational inference module, and when the variational loss is minimized, the trained network is obtained.
[0018] Further, the optimization target of the network is represented as:
[0019] In the formula, represents the parameters of the graph attention network and the variational inference module, represents the optimal parameters of the network, and the parameters can be updated by the gradient descent method and back propagation Update, the update step can be represented as:
[0020] In the formula, Indicates the learning rate, and the update method adopts a cosine decay strategy. The calculation process can be represented as:
[0021] In the formula, is the minimum learning rate, is the initial learning rate, which is set to 0.001, t is the current iteration number, and T is the half cycle length.
[0022] Further, in step S5, the labeled monitoring data can be represented as , wherein, and represent the sample and the label respectively, and represent the sample space and the label space respectively, represents the number of samples, and under the small sample task, the number of samples is less than 30. The classification module adopts a fully connected layer softmax function for classification:
[0023] In the formula, is the predicted label, is the output of the graph attention network. According to the output of the softmax; The cross-entropy loss function is used to measure the error, and the calculation process is represented as:
[0024] In the formula, K is the number of health states, is the real label one-hot encoding. represents the classification loss error.
[0025] Further, in step S6, the prediction error in S5 is used as an optimization target to update the parameters of the classification module, including: taking the error between the predicted value and the real value as the optimization object, using the gradient descent algorithm to train the network, keeping the parameters of the graph attention network unchanged, updating the parameters of the classification module, and when the error between the predicted value and the real value is the smallest, the trained classification network is obtained, wherein the gradient descent algorithm is used to update the parameters of the classification module, and the network optimization target is represented as: The parameters of the classification module can be updated by the gradient descent method and back propagation. Update, the update step can be represented as:
[0026] In the formula, This represents the parameters of the classification module. This represents the optimal parameters of the network.
[0027] The beneficial effects of this invention are: The present invention provides a DC power distribution board health status assessment method based on multiple electrical parameters, which includes a graph attention network, a variational inference module, and a classification module. This method can identify the health status of DC power distribution boards, improve their reliability, and extend their service life. Attached Figure Description
[0028] Figure 1 This is a flowchart of the DC power distribution board health status assessment method based on multiple electrical parameters according to the present invention; Figure 2 This is a graph showing the changes in training loss and validation loss during the training process of this invention. Figure 3 This is a confusion matrix diagram representing the evaluation results of this invention. Detailed Implementation
[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0030] Example: In accordance with the invention description of the DC distribution board health status assessment technology based on multiple electrical parameters, a DC distribution board health status assessment method is designed and developed. Figure 1 This describes the specific processing flow of the method, which mainly includes the following steps: (S1) Collect unlabeled monitoring data of multiple electrical parameters of the DC distribution board. This unlabeled monitoring data does not contain health status labels. Perform multimodal fusion on the data to construct a spatiotemporal graph structure.
[0031] Specifically, unlabeled monitoring data can be represented as .in, Indicates a sample, Representing the sample space, This indicates the number of samples. For small-sample evaluation tasks, the goal is to learn general knowledge from a large amount of unlabeled data and fine-tune it using a small amount of labeled data from downstream tasks. A sliding window is used to segment the data, with each sample having 1024 sampling points. Cross-modal attention is used to fuse different parameters to construct a graph structure.
[0032] (S2) Input the graph structure into the graph attention network to extract features and obtain high-dimensional spatiotemporal dependent features.
[0033] Specifically, the backbone of this graph attention network is GAT (Graph Attention Network). The network adaptively aggregates neighbor node information through an attention mechanism. Specifically, the input graph G=(V,E), node features H, and the output updated features H' after attention calculation.
[0034] The attention coefficient is calculated as follows:
[0035] In the formula, Let be the attention coefficient between node i and j. It is a linear transformation matrix. Let be the attention vector, and LeakyReLU be the activation function.
[0036] The node is updated to:
[0037] In the formula, Let i be the set of neighbors of node i. This is the activation function.
[0038] (S3) Input the features into the variational inference module and use the variational lower bound loss to measure the similarity of the latent distributions; Specifically, this variational inference module is based on a variational autoencoder (VAE), where the encoder q(z|x) learns the mean. and variance Reparameterization ,in The decoder p(x|z) reconstructs the input.
[0039] (S4) The variational loss obtained in S3 is used as the optimization objective of the graph attention variational network. The Adam optimizer is used for optimization, and the gradient descent algorithm is used to update the parameters of the graph attention network and the variational inference module. When the variational loss is minimized, the training ends and the parameters of the graph attention network are saved.
[0040] Specifically, the optimization objective of the network can be expressed as:
[0041] In the formula, This represents the parameters of the graph attention network and the variational inference module. The optimal parameters of the network can be obtained through gradient descent and backpropagation. Update, the update steps can be represented as:
[0042] In the formula, The learning rate is represented by a cosine decay strategy, and the calculation process can be expressed as follows:
[0043] In the formula, To minimize the learning rate, The initial learning rate is set to 0.001, t is the current iteration number, and T is the half-cycle length.
[0044] (S5) Collect a small amount of tagged monitoring data of DC power distribution boards, extract features through graph attention network, and use classification module for prediction. Measure the error between predicted label and real label using cross-entropy loss function.
[0045] Specifically, label monitoring data can be represented as .in, and These represent the sample and the label, respectively. and Represent the sample space and label space, respectively. This indicates the number of samples. In small sample tasks, the number of samples is typically less than 30. The classification module uses the softmax function of a fully connected layer for classification.
[0046]
[0047] In the formula, For the predicted labels, The output of the graph attention network is shown. Based on the softmax output, the error is measured using the cross-entropy loss function. The calculation process can be expressed as follows:
[0048] In the formula, K represents the number of healthy states. One-hot encoding for the actual label. This represents the classification loss error.
[0049] (S6) Freeze the parameters of the graph attention network and use the prediction error in S5 as the optimization target to update the parameters of the classification module; Specifically, the network optimization objective can be expressed as:
[0050] In the formula, This represents the parameters of the classification module. The optimal parameters of the network can be obtained through gradient descent and backpropagation. Update, the update steps can be represented as:
[0051] (S7) Use the trained graph attention network and classification module to assess the health status of the DC power distribution board under test using monitoring data.
[0052] Based on the above process, the loss curve during model training and the confusion matrix of the final prediction result are obtained as follows: Figure 2 and Figure 3 As shown in the diagram, the loss curves reveal that both training and validation losses exhibit a stable decreasing trend with increasing training epochs, eventually converging to around 0.8. The decreasing trends of the training and validation losses are largely consistent, and their final values are close, indicating that the model does not exhibit significant overfitting. The curve shape suggests that the model stabilizes after approximately 40 epochs, indicating that the current training epoch setting is reasonable. However, the validation loss still fluctuates slightly in the later stages; it is recommended to further adjust the learning rate or add a regularization term to optimize convergence stability. The confusion matrix shows that the model can effectively distinguish several fault modes in the microgrid.
Claims
1. A method for assessing the health status of a DC switchboard based on multiple electrical parameters, characterized in that, Includes the following steps: S1. Collect unlabeled monitoring data of multiple electrical parameters of DC distribution board, perform multimodal fusion on the unlabeled monitoring data, and construct a spatiotemporal graph structure; S2. Input the spatiotemporal graph structure into the graph attention network for feature extraction to obtain high-dimensional spatiotemporal dependent features; S3. Input the high-dimensional spatiotemporal dependent features into the variational inference module and use the variational lower bound loss to measure the similarity of the potential distributions; S4. Use the variational loss obtained in step S3 as the optimization objective of the graph attention variational network, update the parameters of the graph attention network and the variational inference module, and save the parameters of the graph attention network after training; S5. Collect a small amount of labeled monitoring data of DC power distribution boards, extract features through the graph attention network, and use the classification module for prediction. Use the cross-entropy loss function to measure the error between the predicted label and the true label; S6. Freeze the parameters of the graph attention network, use the prediction error in step S5 as the optimization objective, and update the parameters of the classification module; S7. Using the trained graph attention network and classification module, perform a health status assessment on the monitoring data of the DC power distribution board to be tested, and output the assessment results.
2. The method for assessing the health status of a DC distribution board based on multiple electrical parameters according to claim 1, characterized in that, In step S1, multimodal fusion of monitoring data, including voltage, current, and power, is performed, with 1024 sampling points for each sample. Different parameters are integrated into a graph structure using a multimodal fusion method. The fusion method is cross-modal attention and spatiotemporal graph construction.
3. The method for assessing the health status of a DC distribution board based on multiple electrical parameters according to claim 2, characterized in that, Cross-modal attention is used to weight and fuse different electrical parameters. The calculation formula is as follows: In the formula, The characteristics after fusion For attention weights, Let be the characteristic of the i-th mode, and exp be the exponential function.
4. The method for assessing the health status of a DC distribution board based on multiple electrical parameters according to claim 1, characterized in that, In step S1, the spatiotemporal graph structure is constructed as follows: the fused features are regarded as nodes, and the temporal and spatial dependencies are regarded as edges. The graph is represented as G=(V,E), where V is the set of nodes and E is the set of edges.
5. The method for assessing the health status of a DC distribution board based on multiple electrical parameters according to claim 1, characterized in that, In step S2, the backbone of the graph attention network is the Graph Attention Network (GAT). The network adaptively aggregates neighbor node information through an attention mechanism, including the input graph G=(V,E), node features H, and outputs updated features H' after attention calculation, where: The attention coefficient is calculated as follows: In the formula, Let be the attention coefficient between node i and j. It is a linear transformation matrix. Let LeakyReLU be the attention vector and LeakyReLU be the activation function. The node is updated to: In the formula, Let i be the set of neighbors of node i. This is the activation function.
6. The method for assessing the health status of a DC distribution board based on multiple electrical parameters according to claim 1, characterized in that, In step S3, the features are input into the variational inference module, and the variational lower bound loss is used to measure the similarity of the latent distributions. This includes: in the variational inference module, assuming that the latent variables follow a Gaussian distribution, the encoder approximates the posterior distribution q(z|x), and the evidence lower bound (ELBO) is maximized to minimize the KL divergence; wherein, the variational lower bound loss is calculated as follows: In the formula, As the variational lower bound, Let be the expected function. Let KL divergence be the KL divergence. This is the prior distribution.
7. The method for assessing the health status of a DC distribution board based on multiple electrical parameters according to claim 1, characterized in that, In step S4, updating the parameters of the graph attention network and the variational inference module by using the variational loss obtained in S3 as the optimization objective of the graph attention variational network includes: using the variational loss as the optimization object, training the graph attention variational network using the gradient descent algorithm, updating the parameters of the graph attention network and the variational inference module, and obtaining the trained network when the variational loss is minimized.
8. The method for assessing the health status of a DC distribution board based on multiple electrical parameters according to claim 6, characterized in that, The optimization objective of the network is expressed as: In the formula, This represents the parameters of the graph attention network and the variational inference module. The optimal parameters of the network can be obtained through gradient descent and backpropagation. Update, the update steps can be represented as: In the formula, The learning rate is represented by a cosine decay strategy, and the calculation process can be expressed as follows: In the formula, To minimize the learning rate, The initial learning rate is set to 0.001, t is the current iteration number, and T is the half-cycle length.
9. The method for assessing the health status of a DC distribution board based on multiple electrical parameters according to claim 1, characterized in that, In step S5, the tag monitoring data can be represented as ,in, and These represent the sample and the label, respectively. and Represent the sample space and label space, respectively. This indicates the number of samples. In small sample tasks, where the number of samples is less than 30, the classification module uses the softmax function of a fully connected layer for classification. In the formula, For the predicted label, This is the output of the graph attention network. Based on the output of the softmax function; The error is measured using the cross-entropy loss function, and the calculation process is expressed as follows: In the formula, K represents the number of healthy states. One-hot encoding for the actual label. This represents the classification loss error.
10. The method for assessing the health status of a DC distribution board based on multiple electrical parameters according to claim 1, characterized in that, In step S6, updating the parameters of the classification module using the prediction error from S5 as the optimization objective includes: using the error between the predicted value and the true value as the optimization object, training the network using the gradient descent algorithm, keeping the parameters of the graph attention network unchanged, and updating the parameters of the classification module. When the error between the predicted value and the true value is minimized, the trained classification network is obtained. The network optimization objective, expressed as follows: Gradient descent and backpropagation can be used to... Update, the update steps can be represented as: In the formula, This represents the parameters of the classification module. This represents the optimal parameters of the network.