An anomaly detection method for gas turbines based on single-condition training with zero-fault sample

Enhancing anomaly detection performance is essential for effective gas turbine condition monitoring and health maintenance. However, in industrial applications, gas turbine operating conditions frequently change, and fault data are scarce or even unavailable. Therefore, identifying anomalies in unknown conditions with training based only on normal data is challenging. Inspired by human communication, where listeners can identify a specific speaker in a crowd regardless of speech rate or intensity, this paper develops a semi-supervised automatic anomaly detection method for gas turbines based on Mel frequency mapping, called Mel Frequency Mapping Anomaly Detection (MFMAD). This method uses training data composed entirely of normal signals (semi-supervised) under single operating conditions to identify abnormal vibration behaviors in other operating conditions of gas turbines. Based on this concept, we developed the following key technologies: (1) Utilizing Mel frequency mapping technology to convert vibration signals from linear Hertz (Hz) frequency to nonlinear Mel frequency, and the fault characteristics under different working conditions are mapped to a unified space. (2) Through Convolutional autoencoder (CAE) semi-supervised learning, only Mel spectrograms of normal vibration signals are used to learn the normal spectral structure in the training stage. In the testing phase, the Structural Similarity Index (SSIM) between the original signal and the reconstructed signal is used as a discriminative indicator to identify abnormal signals. To verify the effectiveness of this method in anomaly detection, the state-of-the-art Area Under the Receiver Operating Characteristic (AUROC) metric is used to evaluate anomaly detection performance. The method achieved remarkable results on two laboratory datasets, with AUROCs of 0.997 and 0.983, respectively. Additionally, on the gas turbine real testbed dataset, the AUROC reached 0.868. This research provides a new solution for early fault warning and maintenance of gas turbines.