Infrared Thermography-Based Thermal Diagnostics of Flow-Boiling Minichannels with Machine-Learning-Assisted Profile-Quality Assessment

Infrared thermography provides non-contact access to spatially resolved wall-temperature information, but large flow-boiling archives require traceable separation of physically meaningful variations from measurement artefacts. This study develops a physics-guided workflow for quality assessment of longitudinal infrared-temperature profiles extracted from the central line of an externally observed heated foil in rectangular minichannels. The frozen dataset comprised 143,821 complete profiles, 35,666,447 spatial points, 447 source files, and 20 grid families. Leakage-safe source-file-level partitioning was used. Controlled evaluation covered eight local synthetic anomaly morphologies at three intensities and a separate global-offset challenge. A Random Forest using 64 engineered profile descriptors was compared with a five-channel masked residual one-dimensional convolutional neural network operating directly on complete profiles. On the locked test set, the Random Forest achieved an area under the precision–recall curve of 0.915, sensitivity of 0.731, and false-positive rate of 0.082, compared with 0.871, 0.452, and 0.063 for the neural model. Both models deteriorated under external grid-family transfer and remained near chance for uniform temperature offsets. The Random Forest is retained as the preferred supervised benchmark; geometry and grid-family transfer remain the principal unresolved limitation.