A re-optimized deep auto-encoder for gas turbine unsupervised anomaly detection

The use of hidden features or reconstruction errors extracted by deep auto-encoder (DAE) is becoming popular to discriminate anomalies from normal. Nevertheless, the fact that the existing methods only involving one of these two aspects loss the useful information from the other one motivates this investigation of method to combine reconstruction errors with hidden features. More importantly, anomalies are not removed in the training set when optimizing the traditional DAE, which weakens the discrimination of the reconstruction error. Aiming at these two problem, a new deep learning method, the so-called Re-optimized Deep Auto-Encoder (R-DAE), is developed to improve the detective performance for gas turbine unsupervised anomaly detection. First, the developed R-DAE takes the hidden features and the induced reconstruction errors as the final features. Second, to make the reconstruction error more discriminative, a sample selection mechanism is designed to attempt to remove anomalies from original unannotated training set, which makes the R-DAE almost unaffected by abnormal samples during training. Third, to effectively process time series, isolation forest is used to detect the obtained final features. Experiments on the real-life operation data of a gas turbine sample fleet validate the excellent detective performance of the proposed method.