Improved Outlier Detection for Failure Forecasting using Anomaly Score Threshold Optimization and Ensemble Methods

This paper examines the usefulness of optimizing the anomaly score threshold determining the predicted class of the Isolation Forest binary classifier for outlier detection, as well as the benefits of combining classification models with different preprocessing arguments in an ensemble for an overall prediction. The purpose of anomaly detection in this work is the timely forecasting of failures of specific type for predictive maintenance based on sensorial data related to the production process. The implementation was carried out on an anode production equipment in the aluminium industry. The experimental results show that, especially for larger forecasting horizons, the introduced approach significantly increases the classification evaluation performance compared to our previous work, with an improvement in Matthews Correlation Coefficient of up to $\mathbf{0. 0 6 2}$.