Deep Learning-based Anomaly Detection for Compressors Using Audio Data

By pressurizing natural gas in pipelines, the compression system interlocks upstream gas production and downstream consumer use. Considering the installation cost of $1 to $2 million US dollars for a compressor, the failure of the component can be costly. Therefore, the anomaly detection for the compression system is essential. In this paper, a deep learning-based anomaly detection method is proposed to identify the failure of midstream compressors using audio sensor data. Firstly, short-term Fourier transform (STFT), Melfrequency cepstral coefficients (MFCC), and spectral centroid (SC) features are computed using the input audio signals. Secondly, deep learning-based feature extraction is applied to create high-level features. Finally, a principal component analysis step and a support vector machine are applied to classify normal and anomaly audio signals. The proposed method was evaluated using two datasets with a total of 10196 audio signals collected from a compressor. The experimental results demonstrate that MFCC features are better than STFT features for anomaly detection and the combined deep MFCC features and SC features can achieve the best normal and anomaly signal classification performance, 100% for both datasets, using the proposed method.