Optimizing Telemetry Signal Influence for Power Consumption Prediction

Abstract

Automated Guided Vehicles (AGVs) are common elements of contemporary industry. Their continuous operation, and thus detection of anomalies in their operational cycles, is critical for uninterrupted production flow. Prediction of signals, such as momentary power consumption (MPC), is used in most anomaly detection methods. Feature engineering - selection or weighting - can significantly improve prediction quality. In this work, we use a genetic algorithm (GA) to optimize weights for features from AGV telemetry. A 2-layer Long Short-Term Memory (LSTM) network was used to predict MPC. Our primary goal was identifying the most effective weighting strategy for enhancing predictive accuracy. We examined different schemes of population initialization. The performance of each was compared to baseline models. Results show a significant improvement in prediction quality compared to the baseline. Our application of GA optimization in feature engineering contributes to the growing body of knowledge on developing more reliable AGV systems, which can lead to reduced operational costs and enhanced sustainability in various industrial settings.