Operationalizing Physics-Informed Deep Learning in Fire Safety: A Robust Design and Product Lifecycle Approach

Abstract

Measuring fire heat release rate (HRR) is central to fire safety engineering. In tunnels and facades, however, conventional oxygen-consumption calorimetry is often impractical due to geometric constraints, harsh thermal environments, and sensor degradation caused by smoke and flame. Because surveillance cameras are already widely installed in tunnel infrastructure, this thesis develops VisionKoopmanNet, an image-based deep learning pipeline that estimates HRR directly from optical video and embeds these predictions within a structured operational framework for safety-critical evaluation. The technical pipeline is built in three phases. Phase 1 trains a physics-informed Koopman-LSTM HRR encoder on 756 Fire Dynamics Simulator (FDS) scenarios and reaches R2 = 0.9996 on 1.73 million held-out test samples. Phase 2 trains an EfficientNet-B7 image branch on the NIST Fire Calorimetry Database using an experiment-level stratified-by-peak-HRR train/val/test split (391/84/82 experiments) and a proportional frame-to-HRR index mapping that corrects a hardcoded fps=5.0 assumption that had silently misaligned every window with pre-ignition baseline samples in earlier work. Two complementary Phase 2 results are reported. The image-only HRR estimator, representing the operationally relevant capability for tunnel and facade deployment where calorimetry sensors are unavailable, predicts absolute HRR from camera video alone with no sensor input at inference and achieves held-out test MAE = 251 kW on 50,720 windows drawn from 82 stratified-held-out NIST experiments, with sub-150 kW MAE in the 0–200 kW regime (n = 34,177 windows, 67% of the test set) and documented failure modes on rare megawatt-scale fires (n = 361 at >5 MW). The sensor-fused forecaster, which takes the most recent calorimetry reading as an additive anchor in log-target space alongside image features, achieves test MAE = 8.93 kW but matches a multi-step persistence baseline (8.88 kW) within 0.5% across every per-bin slice tested; empirically, the decoder learns Δ = 0 for every input and the prediction collapses to copying the sensor reading. This residualanchor trap survives every architectural protection investigated (log-target regression, hard spectral-radius projection, end-to-end unfrozen encoder, stratified split, activefire weighted sampler, magnitude-weighted MAE) and is reported as a structural information-theoretic finding rather than a tuning failure. Phase 3 attempts to fuse the two encoders into a shared Koopman latent space; three alignment strategies (mean-squared-error on unpaired data, joint retraining, contrastive InfoNCE) all fail to produce per-sample cross-modal correspondence, with three documented failure modes (no improvement, catastrophic forgetting, and representation collapse). The Phase 3 negative result, traced to a data-level cause (namely that the FDS scenarios and NIST experiments represent different physical fires rather than the same fire observed across two modalities), is reported as a primary scientific contribution. v Alongside technical pipeline development, this thesis answers two research questions on safety-critical operationalization. The first investigates how Critical-to-Quality (CTQ) requirements, Operational Design Domains (ODD), and Stage-Gate lifecycle controls can be designed to ensure rigorous specification, workflow compatibility, and robustness for an offline fire prediction prototype. The second investigates how Failure Mode and Effects Analysis (FMEA), safety wrappers, and human-in-the-loop escalation rules can be structured to govern validation and risk management prior to operational deployment. The primary contributions of this thesis are fourfold: (1) A working camera-based image-only HRR estimator (MAE = 251 kW on held-out NIST testing, sub-150 kW for fires up to 200 kW) suitable for tunnel and facade scenarios where physical calorimetry cannot be installed; (2) A structural negative finding demonstrating the residual-anchor trap in sensor-fused HRR forecasting; (3) An empirical investigation into cross-modal latent alignment without paired multi-modal data, identifying three distinct failure modes in cross-modal learning for fire dynamics; and (4) An endto- end operationalization framework combining CTQ metrics, ODD delimitation, Stage-Gate quality gates, FMEA risk prioritization, and runtime safety wrappers to govern bounded prototype safety. A separate methodological contribution is the documentation of a silent hardcoded frame-rate alignment error that previously inflated Phase 2 metric scores, together with the multi-iteration rebuild that corrected it. The scope is strictly bounded to offline analysis of recorded NIST video sequences; real-time edge deployment, multi-camera fusion, hardware procurement, graphical user interfaces, and formal compliance certification remain out of scope for future work.