A Comparative Analysis of Machine Learning and Deep Learning Techniques for Component Fault Prediction

As the energy sector continues to grow and evolve, the need for advanced maintenance techniques that ensure the reliability and safety of electrical components becomes increasingly important. The adoption of predictive maintenance has become a major concern for companies that require regular component replacement. Enterprises are investing in creating predictive models that anticipate component failure to meet the demands of today's decade. Predictive maintenance is a very useful and powerful tool to ensure component reliability, as it allows companies to monitor the health of their assets and detect potential failures before they occur. Given the large amount of data generated by modern electrical systems, it is very challenging to select an appropriate classification model for predictive maintenance. In this thesis work, we compare two different classification models used for predictive maintenance of electrical components of an Italian energy company: eXtreme Gradient Boosting (XGBoost) and Restricted Boltzmann Machine used as a stand-alone classifier (discrete RBM). The hyperparameters of both models are fine-tuned to obtain the best results. The precision score is used as the evaluation measure since there is a significant imbalance between the two classes in the data. Each method is evaluated in different scenarios based on the number of interventions. To determine the precision of the models, the number of predicted outages must be determined for each scenario, which are true positives (i.e., correctly predicted outages) and false positives (i.e., incorrectly predicted outages). To train and test the classification models, we use a dataset of electrical components from a real-world energy company. The dataset is highly unbalanced, with only a small percentage of instances representing failures. This problem is addressed by using an undersampling technique to mitigate the class unbalancing and improve the performance of the models. Our evaluation results show that both XGBoost and discrete RBM predict potential failures with similar precisions, but XGBoost performs slightly better and requires significantly less computation time than the discrete RBM. Throughout this thesis work we will see why this is a very challenging task and we will discuss the limitations of the models and the data.