MODEL FOR AUTOMATED DETECTION OF SPECTRAL ANOMALIES IN THE DEGRADATION OF MARINE DIESEL ENGINES USING A DENOISING AUTOENCODER

Abstract

The aim of this article is to automate the process of detecting anomalies associated with the degradation of marine diesel engines using a denoising autoencoder (DAE). The methods for anomaly detection are focused on identifying deviations from typical engine operation. From a Prognostics and Health Management (PHM) system perspective, such deviations can serve as indicators of impending failures. Detecting faults is the initial and critically important stage in data-driven PHM systems. The developed model in this article for automated spectral anomaly detection in the degradation of marine diesel engines, irrespective of the fault type, is based on a denoising autoencoder (DAE). The DAE is trained on preprocessed data from normal engine operation. DAE training aims to find optimal encoder and decoder parameters that minimize a loss function. An important aspect is that the DAE is trained on noisy data, aiding in extracting robust and informative features. Subsequently, the trained DAE model is used to calculate the speed and acceleration of anomaly assessment at each time step in degradation data in the event of a fault. Generic and dynamic threshold values are simultaneously established. These calculations and threshold values dynamically change over time, enabling real-time fault detection. The proposed model has demonstrated high efficiency in detecting three types of faults in marine diesel engines: air filter clogging, turbocharger malfunction, and frequency-controlled fan faults. The model can be used for online fault detection, facilitating timely corrective actions. It holds promise for practical use in PHM systems for marine diesel engines. However, further development tasks include automating data labeling for both fault classification and Remaining Useful Life (RUL) predictions, as well as analyzing the root cause of faults and isolating them. Solving these tasks will enhance the model's effectiveness and accuracy.