A Grid Fault Diagnosis Framework Based on Adaptive Integrated Decomposition and Cross-Modal Attention Fusion

Research Background

With the continuous expansion and increasing complexity of modern power systems, the stable operation of the grid faces growing challenges. Grid faults can occur due to natural disasters, equipment failures, and local grid structural weaknesses, among other factors. These faults not only affect the normal operations of electricity users but can also lead to widespread blackouts, causing significant losses. According to the U.S. Energy Information Administration, the United States experiences over 500 grid fault incidents annually, affecting power supply to millions of users. In China, annual power loss due to grid faults exceeds billions of yuan. Hence, quickly and accurately detecting and diagnosing grid fault types has become a key topic in power system research.

Research Source

This paper, titled “a grid fault diagnosis framework based on adaptive integrated decomposition and cross-modal attention fusion,” was written by Jiangxun Liu, Zhu Duan, and Hui Liu, primarily affiliated with the Institute of Artificial Intelligence and Robotics and the School of Traffic & Transportation Engineering at Central South University. The article was published in the journal Neural Networks, with the acceptance date of May 19, 2024.

Research Process and Details

Research Process

1. Data Preprocessing:

   • Current and voltage signals need to be normalized due to different magnitude issues.
   • Five advanced decomposition algorithms are used to decompose the original signals, resulting in subsequences containing different temporal and frequency domain features.
   • A random forest model is employed to reduce the dimensionality of the high-dimensional data, identifying key features to reduce redundancy.
   • Based on the Comprehensive Information Entropy Value (CIEV), appropriate weights are assigned to the main subsequences, integrating the decomposed data to generate model input data.

2. Multimodal Feature Learning:

   • Features from numerical mode, image mode, and graph mode are extracted using Deep Residual Convolutional Neural Networks (DRCNN) and Heterogeneous Graph Transformers (HGT).

3. Cross-Modal Attention Fusion Mechanism (CMAF):

   • Implicit features are weighted accordingly and undergo cross-modal attention-weighted feature fusion.
   • The final fault category prediction is output through a Softmax layer.

Research Details

1. Signal Decomposition:

   • Five algorithms, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Wavelet Packet Decomposition (WPD), Variational Mode Decomposition (VMD), Empirical Wavelet Transform (EWT), and Local Mean Decomposition (LMD), are used to process the data, resulting in pattern components rich in feature distribution.

2. Feature Selection:

   • The random forest model combines multiple sampling and multiple decision trees to screen high-dimensional feature data, retaining representative IMF components and reducing feature dimensionality.

3. Comprehensive Information Entropy Value Weighting:

   • CIEV indicators are calculated for each subsequence to measure the feature information value, forming a weighted covariance matrix that coordinates signal’s fuzzy entropy and mutual approximate entropy, enhancing the feature integration effect.

4. Two-Dimensional Visualization:

   • Numerical data is visualized in two dimensions using the Gramian Angle Field (GAF) method, generating image mode data.

5. Cross-Modal Feature Fusion:

   • The three modal features are pooled through Avgpool and Maxpool, and then processed through a Multilayer Perceptron (MLP) to get new hidden states, undergoing cross-modal attention fusion. The results are finally input into the classifier to obtain the prediction.

Research Results

The method proposed in this paper achieved high diagnostic accuracy in the verification on the Three-Phase Transmission Line (TTL) dataset and the VSB Power Line dataset, reaching 99.4% and 99.0% respectively. Detailed experimental data for the TTL dataset classification is as follows:

Modal	Model Accuracy	Precision	Recall	F1-score	AUC
Single Modal (Numerical)	0.966	0.969	0.957	0.965	0.980
Single Modal (Image)	0.983	0.982	0.983	0.983	0.990
Single Modal (Graph)	0.985	0.985	0.986	0.986	0.992
Multimodal (Numerical + Image)	0.982	0.981	0.984	0.982	0.990
Multimodal (Numerical + Graph)	0.985	0.982	0.988	0.985	0.991
Multimodal (Three Modal)	0.994	0.994	0.995	0.994	0.997

Research Conclusions and Significance

The proposed grid fault diagnosis framework based on adaptive integrated decomposition and cross-modal attention fusion demonstrates significant application value in feature extraction and pattern recognition of grid data. By integrating multiple decomposition algorithms and modal feature fusion, this method can extract key information from complex grid data, improving accuracy and robustness of diagnosis. For practical engineering applications, this method has important reference value in handling large-scale, complex grid fault diagnosis.

Research Highlights

• Proposed and designed the Comprehensive Information Entropy Value (CIEV) indicator to effectively utilize the time-frequency characteristics of data, improving feature extraction efficiency.
• Enhanced generalization performance by combining the results of multiple excellent decomposition methods using an adaptive integrated decomposition algorithm.
• Designed a cross-modal attention fusion mechanism to integrate different modal features through weight allocation, improving the accuracy and stability of the diagnosis model.

Other Valuable Information

The proposed method in this paper is not only suitable for existing grid fault diagnosis but also has potential application value for fault detection and classification in other large complex systems. Moreover, the combination of deep learning and information entropy theory has far-reaching implications for improving fault detection efficiency and reducing manual intervention steps.