Evaluation of Artificial Neural Networks for Predictive Maintenance

Sammanfattning: This thesis explores Artificial Neural Networks (ANNs) for predictive time series classification for Predictive Maintenance (PdM). Time slicing and time shifting are methods used, to enable the models to find features over time, and to predict into the future, respectively. Architectures of increasing complexity are explored for Feed Forward Neural Networks (FFNNs), Convolutional Neural Networks (CNNs) & Long Short-Term Memory (LSTM) networks, of which the best performing are compared. CNNs & LSTM are found to perform better than FFNNs since they are designed to handle sequences of data. This research shows that a model with high accuracy might in fact be a bad model for PdM, especially when the data set is imbalanced. Additional metrics such as Confusion Matrices and Receiver Operating Characteristic (ROC)-curves are needed to evaluate models. This thesis shows that consistent, representative and a lot of data of good quality is needed for a well performing ANN. ANNs for PdM reduces the required domain knowledge, and perform well for common/frequent classes, but less so for the less frequent classes.