Predictive Maintenance for off-road vehicles based on Hidden Markov Models and Autoencoders for trend Anomaly Detection.

Off-road vehicle maintenance is getting increasingly important as the unplanned stops might significantly damage the entire work delaying the process beyond the limits. While preventive maintenance consists of replacing periodically vehicle components, predictive maintenance predicts if and when a failure is going to occur. This thesis investigates unsupervised and supervised methods for predicting vehicle maintenance. In order to achieve this goal, we have used probabilistic methods based on the concept of Markov chains (HMMs), distance (kNN) and tree (Isolation Forest) based algorithms, and finally deep neural networks (Autoencoders). We built a model reaching about 81% of F1 score in the prediction of diagnostic messages occurrences. The model figures the trend out and perceives the closeness to the failure by claiming faulty values with at least 30 hours in advance. The thesis is divided into three parts and it goes as follows. The first one is predictive-oriented data preparation and it involves all the passages from the analysis of two datasets to the building of the final one as a base of the future model. The second part involves the so-called feature selection, where all the present parameters are reviewed and, through a comparison between each variable and the response label, the set of the most significant features is selected by 4 methods. The first method is related to the Mutual Information which compares the distributions of data. It highlights an evident weak dependence among features and response variables. The second one is about Recursive Feature Selection with Support Vector Classifier, which makes a ranking of the variables from the first relevant to the last one. Then, we applied a tree-based method based on Importance of features called Extremely Randomized Trees. The last one is a distance-based method and it considers Dynamic Time Warping distance to create many time series and to get out a score from them. Because of the IoT data structures, just the feature monitoring directly the values that cause the damage survived after the selection process. The last part provides models to predict a maintenance need. The first one is a probabilistic model called Hidden Markov Model that, by creating K hidden states, can be used as a trend anomaly detector to find out when values start to be anomalous. It reaches scores around 60-65%. The second one is called kNN Outlier Detection and it is a distance-based model. It has scores around 70%, showing it limits under noisy data. The next model is called Isolation Forest and it is based on sequential cuts of the data intervals in order to isolate each point. The results are about 78%. Then we will apply Autoencoders Neural Networks, reaching 79-80% of precision and recall with stratified cross-validation. The model is a neural network that for first destroys data by reducing their dimensions and then it reconstructs them to compute the reconstruction error. The final method is a two step model and we named it Autoencoder on Hidden Markov Model, which increases the interpretability of Autoencoders and it makes the model more adaptable. The results are slightly better than the last, achieving 80-81% of precision and recall.