Research on anomaly detection method combining distance correlation coefficient and autoencode

This study proposes a method of anomaly detection based on a combination of distance correlation coefficient-based feature selection algorithm and autoencoder. In this paper, we use the distance correlation coefficient to analyze the correlation of the original feature set, and divides the feature set into multiple feature subsets according to the correlation between features. The features within each feature subset are filtered by the constructed feature representativeness evaluation indexes to remove redundant features. Then, we built a convolutional denoising autoencoder to enhance the anomaly detection ability of the autoencoder in the time dimension. In the constructed autoencoder, a modular design approach is used to divide the encoder and decoder structures into encoding and decoding units, and the accuracy of fitting the network to the training data can be tuned by adjusting the number of these two units. Finally, the proposed method is validated with a turbofan engine. The results show that the proposed method outperforms other traditional methods in accuracy and has application value.