A Data-driven Fault Prediction Method for LNG Engine City Buses

Traditional methods of fault detection and diagnosis for vehicles are mainly based on expert knowledge, signal analytics and chemical experiments, which lead to subjectivity, uncertainty and hysteresis somewhat. Motivated by this problem, we propose a novel data-driven method using Rough Set theory and Random Forest algorithm for constructing the predictive model from CAN-BUS data, bus maintenance system, bus daily schedule system and relevant weather data, which is able to reduce fault detection deviation resulting from excessive reliance on expert knowledge. More specifically, we utilize the Rough Set theory for extracting key attributes relevant to LNG bus faults. Then, Random Forest algorithm is applied for model construction of the fault prediction. This method provides the opportunity to predict bus faults during the bus operating. Moreover, the effectiveness of model constructed has been validated using real-world data showing very promising results with precision, recall and F1-score all above 0.8. In addition, key attributes extracted are useful for monitoring the bus fault. This predictive model is able to pre-identify LNG engine city buses with potential fault risks, which is important information for improving bus maintenance processes to avoid extra costs.