Physics as a Falsification Test for Quantum-Limited Thermal Leak Detection

Thermal cameras detect heat leaks by taking the difference between two thermal images. However, this difference image can be caused by many factors other than an actual leak. In this work, physics is used differently to previous work. Instead of using the heat equation for predicting a temperature field, we use the steady-state solution to eliminate potential candidates. For a leak on a thin plate, the heat must decay with a K0(r/L) function, where L = sqrt(kd/2h) is a property of the material. Furthermore, the leak must be where the baseline image was cold. We improve the detection AUC of 0.75 ± 0.03 to 0.96 ± 0.01 by making four analytic tests (without training data). This is seen across 5 independent simulated datasets. We are able to detect leaks 5-11 times smaller (about 17 mW) by our method. Using a photon counting (quantum-limited) detector model, we show that the sensor is not limiting. Even with a factor of 140 difference in detector noise (from un-cooled micro-bolometers to ideal background limited), the detection limit only varies by a factor of 1.2. However, by eliminating any change in workload between the two images, we see the detection limit reduce by over a factor of 22. Using classifiers that used the same features leads to better performances on the training material, but can reduce performance by up to 0.09 (4 physics features) or 0.17 (12 features) when new materials or increased camera scaling is used. The gap is not bridged using a pre-registered physics constrained learner. All results are using either a calibrated simulator, or using real thermal backgrounds with simulated leaks. We have not imaged a real leak of known power.