Deep Learning for Forecasting and Fault Detection via Lightweight LSTM and Graph-Based Architectures

Real-time anomaly detection and fault classification in high-dimensional, non-stationary industrial systems are paramount to ensuring operational safety and preventing catastrophic asset failures. While deep learning methods have advanced multivariate time-series monitoring, standard sequential architectures frequently struggle to balance localized sensitivity with plantwide reliability. This thesis addresses these limitations by introducing a novel, real-time diagnostic framework applied to the benchmark Tennessee Eastman Process (TEP). To overcome the architectural and computational bottlenecks of streaming inference, the proposed pipeline splits operations into two distinct phases. Temporally, Variational Mode Decomposition (VMD) is leveraged strictly as an offline initialization tool during the training phase to extract data-driven spectral anchors from fault-free operational sequences. These nominal anchors are subsequently mapped onto a static, causal bank of 4th-order digital Butterworth filters. This decoupled approach entirely circumvents the computational latencies, boundary end-effects, and causality violations inherent to running iterative optimization algorithms at runtime, enabling efficient sample-by-sample streaming with O(1) temporal complexity. Spatially, a graph-based spatio-temporal architecture—the MANOD Forecaster—is deployed to capture complex topological correlations across N = 52 process variables via a normalized spatial propagation matrix (A˜). The framework is evaluated under an unsupervised classification paradigm using One-Class Support Vector Machine (OCSVM) thresholding and compared against a high-performing Multi-Head LSTM baseline. Frame-level metrics show that the LSTM achieves a misleadingly high Macro F1 -score (0.78) and a perfect fault recall (1.00), which a granular, event-based residual analysis exposes as an artifact of severe over-signaling on nominal states (reducing healthy specificity to 0.48). Crucially, the LSTM remains completely blind to 50% of the actual process anomalies, entirely missing slow-drifting and localized structural shocks. Conversely, the proposed MANOD framework sifts through spatial noise to deliver a flawless 100% Fault Coverage Rate, successfully flagging all 20 plant-wide faults while maintaining an outstanding fault precision of 0.95 and a nominal specificity of 0.94. These findings demonstrate that embedding physical graph topologies alongside data-driven causal filtering provides a robust, mathematically sound, and asset-protective foundation for next-generation industrial diagnostics.