A Machine Learning Method to Early Detect Catastrophic Failures of Marine Diesel Engines: A Case Study

Catastrophic failures totally destroying or compromising components of marine engines are sudden and often unpredictable. Due to their abruptness, they are highly dangerous for navigation, crew, and passengers. Then, a big effort is required to maintain the ships healthy. Despite this, most of the works on maintenance focus on predicting the decay law of components, namely the gradual degradation of the performance. No care was paid to sudden, quick, and anomalous events leading to catastrophic failures. We propose here a preliminary study to address the lack of catastrophic failure early detection methods: we evaluate the second derivative, not just the rate, of the difference between the real sensor measurements and expected values of engine variables. Expectations are obtained by a Decision Tree, which implements a specific Machine Learning algorithm. The second derivative of the computed error gives more information on catastrophic failures than the error, because, more than the rate, it can be a symptom that something is suddenly happening inside the component. Also, information is obtained before the sensor measurements reach the thresholds that produce warning alarms. In this way, the ship operators can be notified in advance to shut down the engine. Simulation shows the effectiveness of the method. In fact, the second derivative, with its peaks and high frequency, makes it possible to recognize in advance some abrupt failure is occurring in the engine. In this way, an operator could shut down the engine in suitable time to avoid a big damage.