CNN Fault Arc Signal Detection Using 3D Feature Matrices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional fault arc detection methods struggle to adapt to diverse load environments and fail to effectively detect series arc faults due to abnormal current patterns below protection thresholds.
Innovation Solution
A method utilizing a convolutional neural network (CNN) for fault arc signal detection, which involves filtering sampled current signals with band-pass filters, extracting time-frequency eigenvectors, constructing a 3D matrix, and training a 2D CNN model for online arc detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional threshold-based arc detection method is used, then the detection method is simple to implement, but it cannot adapt to different load environments and has low detection accuracy
Solution Approach 1:
The patent transforms the detection approach from using fixed thresholds to using learned parameters from a convolutional neural network. The CNN model is trained on diverse load data to automatically adapt detection parameters to different load environments, eliminating the need for manual threshold adjustment while improving adaptability.
Solution Approach 2:
The patent replaces the conventional mechanical threshold-comparison detection mechanism with an intelligent convolutional neural network system. The CNN automatically learns detection patterns from training data and performs adaptive detection without requiring manual parameter setting, thereby improving adaptability to different loads.
2Measurement precision
If conventional overcurrent protection device is used, then the device structure is simple, but it cannot effectively detect series arc faults with current below protection threshold
Solution Approach 1:
The patent replaces the simple overcurrent protection device with a convolutional neural network-based detection system. The CNN analyzes current waveforms and identifies arc fault patterns even when current remains below traditional protection thresholds, significantly improving detection accuracy for series arc faults.
Solution Approach 2:
The patent moves from one-dimensional current threshold comparison to two-dimensional waveform analysis using convolutional neural networks. By analyzing the temporal and amplitude characteristics of current waveforms across multiple dimensions, the system can detect arc faults that conventional single-threshold methods miss.
3Adaptability or versatility
If fixed threshold is set for arc detection, then the detection rule is simple, but it requires different thresholds for different loads reducing versatility
Solution Approach 1:
The patent implements a self-adaptive detection system where the convolutional neural network automatically adjusts detection parameters based on the loaded data characteristics. The system performs self-service by eliminating the need for manual threshold configuration for different loads, as the CNN automatically learns appropriate detection criteria during training.
Solution Approach 2:
The patent creates a universal detection model that can handle multiple load types with a single trained system. The convolutional neural network is trained on diverse load data during the offline phase, enabling it to universally detect arc faults across different load environments without requiring separate threshold configurations for each load type.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
A fault arc signal detection method using a convolutional neural network, comprising: enabling a sampling signal subjected to analog-digital conversion to respectively pass through three different band-pass filters; respectively extracting a time-domain feature and a frequency-domain feature from a half wave output of each filter; constructing a two-dimensional feature matrix by means of extracted time-frequency feature vectors from the output of each filter, and stacking the feature matrices corresponding the outputs of the three filters to construct a three-dimensional matrix for each half wave; and processing a multi-channel feature matrix by using a multi-channel two-dimensional convolutional neural network, and determining, according to the output result of the neural network, whether the half wave is an arc. The detection method based on the convolutional neural network has higher accuracy and reliability in recognizing a fault arc half wave, can implement targeted training for different load conditions, and is self-adaptive.