Arc Fault Detection Algorithm Based On Machine Learning

Arc fault is one of the important reason of electrical fire. Arc fault detection device can timely detect the arc fault and disconnect the circuit as soon as possible, so as to avoid the happening of the fire. Arc fault detection technology, therefore, becomes a research hotspot in the field of the current fire prevention technology.This thesis proposes an overall framework for arc fault detection algorithm based on machine learning, including the arc fault feature extraction, data preprocessing and arc fault classifier.When the algorithm is training, the arc fault features are extracted from the arc fault sample data and preprocessed, then the arc fault classifier is trained with the preprocessed feature data and classifier model is set up. Arc fault detection algorithm could work after trainning. It can extract features data from the actually tested circuit, do preprocessing, and use the trained classifier model to classify from arc fault form the inputing preprocessed data, so as to detect arc fault.Based on the analysis of arc fault data in time domain, frequency domain and hierarchical structure, this thesis determines the arc fault features needed, such as energy, sub-band energy ratio,short-term average power, spectral centroid, bandwidth, zero-crossing rate, pulse number, variance and so on.Combining with the characteristics of arc fault detection, this thesis designs the data preprocessing method for the arc fault detection algorithm, including data cleaning based on Condensed Nearest Neighbor, data normalization based on feature-based standardization and data dimensionality reduction based on PCA(Principle Component Analysis).In order to train out an arc fault classifier with strong generalization ability, this thesis proposes a combination classifier consisting of Logistic regression, SVM, random forest classifier, and designs and implements these three classifiers and the second layer classifier.Tests shows that our arc fault detection algorithm can obtain satisfactory results on both the correct identification rate and misoperation for arc fault detection.