Autoencoder and Mahalanobis distance-based monitoring indicator estimation for early clinkering detection in boiler

Abstract The performance of coal-fired boilers has a significant impact on the overall yield of thermal power plants. Among the various boiler faults, the clinkering fault diagnosis is one of the most crucial and scarcely addressed topics in literature. Existing clinkering detection methods are boiler-specific and require both healthy and faulty data for training, which is difficult to acquire. To overcome these drawbacks, a generalized method for early clinkering detection is proposed that only requires normal operation data for training. A stacked-denoising-autoencoder is trained such that the reconstruction error departs from the expected value when clinkering occurs. Mahalanobis distance of this error gives the monitoring indicator for clinkering detection, whose threshold is determined as 385.817 using kernel density estimation. The method is validated using real-time boiler data containing clinkering events, which shows that the obtained threshold clearly demarcates between healthy and clinkering conditions with 99.29% accuracy, providing early alert to operators.