Deep Learning-Driven Combustion Anomaly Detection in Diesel Powertrains: A Multi-Sensor Fusion Approach for Real-Time ECM Adaptation

Abstract

Detecting combustion anomalies in diesel-powertrains for real-time adaptation of the engine control model is crucial for improving engine efficiency, emissions, and reliability. A deep learning-based approach that harnesses multi-sensor data fusion holds promise for fulfilling the goals. Improving on existing methods based exclusively on traditional machine-learners, the proposed solution opens new avenues towards faster and more accurate real-time adaptation. A driving factor behind all the developments is the availability of an unexploited multi-sensor dataset capable of detecting combustion anomalies at the dynamic range of a Diesel, where modelling may not guarantee satisfactory results. The Aferred’s dataset allows testing detection solutions that are not limited to pressure wave indicators, enabling their detection based on vibro-acoustic signals, temperature and exhaust gas composition, alone or in combination, leveraging deep learning capabilities. Real-time adaptation of the engine control module relies on a deep-learning-driven state-of-the-art detection strategy. The approach is evaluated for detection time, generalization across operational domains and sensitivity to faults in the employed sensing suite. Precise detection of combustion deviations allows a safer strategy of control parameters’ adjustment over the running cycle of the engine without endangering operability. Implementation of the solution in the system would allow more efficient adaptation to specific scenarios of use of the engine, expanding the range of optimal emissions and fuel consumption.