Predictive maintanance with graph based feature selection & machine learning in electrical busses

This thesis employs a graph-based feature selection method to analyse the relationships between the characteristics of a CAN (Controller Area Network) Bus data set pertaining to electric buses. Furthermore, it investigates the prediction performance of targeted alarms utilising artificial intelligence techniques. The study employed data obtained from CAN Bus systems of diverse vehicles over an extended period. The raw data packages underwent extensive pre-processing techniques to create data sets suitable for the structural needs of machine learning models. In this process, a hybrid graph-based feature selection tool was developed using a combination of statistical filtering methods, including Pearson correlation analysis, Cramer's V statistical table and ANOVA F-test, and optimisation-based community detection algorithms, such as InfoMap, Leiden, Louvain and Fast Greedy. The developed tool was employed to identify feature subsets, evaluate the relational position of the features and their relationship with the targeted alarms, select distinctive features and perform dimension reduction. In the selection of machine learning models, a number of different approaches were considered, including neural network architectures and traditional classifiers. Amongst these, the Support Vector Machine (SVM), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and Recurrent Neural Network (RNN) models were applied, and their performance was then evaluated. In order to mitigate the impact of class imbalance on model performance, SMOTEEN (Synthetic Minority Oversampling Technique-Edited Nearest Neighbours) and binary search-based time interval down-sampling methods were employed. In order to identify the optimal distinctive classifier space for each model, a grid search and random search-based hyperparameter optimisation (hyperparameter tuning) was conducted. The performance of the developed models was evaluated based on the metrics employed for model validation. The predictions of the models were explained by examining the weights (importance) of the features and utilising the LIME (Local Interpretable Model-Agnostic Explanations) tool to identify the features that triggered the predictions, thereby gaining insights into the functionality (design) of the system. The present study has investigated the prediction performance of specific alarms in vehicles utilising CAN Bus data with machine learning techniques. Additionally, a graph-based feature selection tool has been developed, enabling the visual interpretation of relationships between features and facilitating the enhancement of corrective, preventive, predictive and proactive maintenance (design improvement) approaches. It is concluded that the techniques employed throughout the machine learning life cycle, as conducted in the study, can provide valuable insights and contribute to the advancement of maintenance strategies through machine learning, particularly in contexts where data quality and access to expert opinion may present challenges, in alignment with the principles of Industry 4.0. Keywords: Predictive maintenance (PdM), Machine learning (ML), Graph based feature selection, Explainable artificial intelligence (XAI).