Evaluation of Multi-Modal Learning for Predicting Coolant Pump Failures in Heavy Duty Vehicles

Abstract

Coolant Pump failures in heavy-duty vehicles can cause severe collateral damage if they are not detected and resolved in time; the engine will overheat quickly, rendering the vehicle inoperable. Nowadays, a vast amount of heterogeneous sensor data from different sources is being collected in the automotive industry. Such multi-modal data include onboard signals reflecting the overall usage of the vehicle, multi-dimensional histograms that capture the relation between physical quantities, and categorical variables that encode the physical configuration of the vehicle. This work evaluates several multi-modal learning approaches leveraging this diverse data to build a prognosis and health management system for coolant pumps in commercial heavy-duty vehicles. Four auto-encoder architectures are examined to extract features from 2D histograms. These trained models are anticipated to capture key characteristics of the healthy system operation and yield large reconstruction errors when applied on faulty, or near end-of-life samples. Such learned representations are then combined with expert-engineered features. Both early and intermediate fusion are evaluated on a real-world coolant pump replacement dataset. Results indicate that the combination of diverse features was the most effective approach, thereby motivating further research on multimodal methods.