A Unified Approach for Ensemble Function and Threshold Optimization in Anomaly-Based Failure Forecasting

This paper introduces a novel approach to anomaly-based failure forecasting that jointly optimizes both the ensemble function and the anomaly threshold used for decision making. Unlike conventional methods that apply fixed or classifier-defined thresholds, the proposed framework simultaneously tunes the threshold of the failure probability or anomaly score and the parameters of an ensemble function that integrates multiple machine learning models—specifically, Random Forest and Isolation Forest classifiers trained under diverse preprocessing configurations. The distinctive contribution of this work lies in introducing a weighted mean ensemble function, whose coefficients are co-optimized with the anomaly threshold using a global optimization algorithm, enabling adaptive, data-driven decision boundaries. The method is designed for predictive maintenance applications and validated using sensor data from three industrial domains: aluminum anode production, plastic injection molding, and automotive manufacturing. The experimental results demonstrate that the proposed combined optimization significantly enhances forecasting reliability, improving the Matthews Correlation Coefficient by up to 6.5 percentage units compared to previous approaches. Beyond its empirical gains, this work establishes a scalable and computationally efficient framework for integrating threshold and ensemble optimization in real-world, cross-industry predictive maintenance systems.