Data-driven early fault detection framework for biomass auger reactors

Effective early detection of mechanical faults in biomass reactors is crucial for preventing unscheduled downtime, material loss, and safety hazards. However, fault prediction in this domain remains highly challenging due to the nonlinear, dynamic, and stochastic nature of auger-based feeding systems, which are further influenced by heterogeneous feedstock properties (e.g., particle size, moisture content) and fluctuating operational conditions. This thesis presents a data-driven early fault detection framework tailored to biomass auger reactors, aiming to support proactive maintenance by identifying potential faults before they escalate. The framework establishes an end-to-end workflow, beginning with advanced data preprocessing that transforms raw auger motor torque signals into structured, labeled time-series samples suitable for machine learning (ML) models. A range of baseline models is evaluated, and a hybrid deep learning architecture is proposed, combining bidirectional long short-term memory (BiLSTM) networks with systematically engineered time-series features. This hybrid model, referred to as TSF-BiLSTM, is designed to enhance predictive performance, particularly in scenarios with limited fault data availability. The approach is validated using real-world data from field-deployed biomass reactors processing diverse feedstock types. Compared to conventional BiLSTM, LSTM, and classical ML classifiers, the proposed method demonstrates improved early fault detection capability. Statistical analyses confirm its generalizability across varying operational conditions and feedstock characteristics. Real-time visualizations further demonstrate the model’s potential for integration into online process monitoring systems, supporting more sustainable and reliable biomass energy operations.