A multi-agent decentralized system for prediction of course-outs for autonomous trucks

The expansion of the Autonomous Haulage System (AHS) for large-scale trucks in remote areas remains limited by fixed rule closed systems. The lack of capacity to learn from collective data challenges the predictability of operational events in dynamic scenarios, such as slippage events during winter conditions. Techniques such as Machine Learning and ANN face difficulties due to environmental variability and data heterogeneity, reducing their reliability for predicting these events. This doctoral research develops a comprehensive model divided into three approaches capable of describing and predicting slippage events for Autonomous Haulage Trucks operating under winter conditions. The first evaluates a basic swarm intelligence and machine learning combo, ensuring feature importance and accuracy. Subsequently, two more complex approaches are proposed, involving the development of two fitness functions within a swarm intelligence framework. These fitness functions are designed to be minimized by balancing performance with consistency challenges, while also addressing data heterogeneity. The model seeks to minimize the error associated with these fitness functions by employing various capabilities, including swarm intelligence combined with machine learning, ANN, and federated learning. Additionally, the model incorporates solutions for issues related to parameter weighting during the final aggregation of locally trained models in federated learning, as well as identifying the most appropriate datasets for different groups of trucks. The proposed approaches will be compared against traditional methods of Machine Learning, Artificial Neural Network, and Explainable Artificial Intelligence techniques such as LIME, and SHAP. The evaluation of the approaches will focus not only on their predictive abilities but also on their usefulness in feature importance, aiding interpretability and understanding of course-out occurrences for autonomous trucks during winter conditions. The predictive abilities will be validated using real data collected from a large-scale autonomous fleet operating in extreme winter conditions in Canada, covering operational and environmental data gathered over winter seasons to support algorithm development and validation. The results demonstrate a strong ability to predict truck slippages with proposed methods and identify key variables affecting autonomous truck course-out events. This highlights the value of applying the proposed models to operational data to improve autonomous truck performance during severe winter conditions.