Feature-fused stacked autoencoders for robust thermal tomographic image classification

Abstract Thermal tomographic imaging is an emerging modality for analyzing internal heat-flow behaviour and identifying localized heating anomalies in complex engineering systems. Classifying such images remains challenging due to their high dimensionality, susceptibility to measurement noise and inadequate availability of annotated datasets. This study develops two lightweight deep feature-learning frameworks based on Stacked Autoencoders (SAE) and a proposed adaptive weighted Multi-Scale Feature-Fused SAE (MSFF-SAE) for robust classification of thermal tomographic images. The SAE captures hierarchical abstractions through progressive compression and reconstruction of input data, supported by sparsity regularization, L 2 penalty, a fine-tuned supervised classifier and stochastic noise injection during training. The MSFF-SAE further enhances class separability by fusing multi-level latent features from different encoder depths with the bottleneck representation to preserve both local and global thermal cues. A synthetically generated dataset of 2,070 thermal tomographic images derived from a controlled hot air-flow chamber and COMSOL simulations is used for evaluation. Under a 5-fold cross-validation protocol, adaptive weighted MSFF-SAE achieves a classification accuracy of 97.78 ± 0.93%, outperforming the baseline SAE (96.96 ± 1.45%) with reduced inter-fold variability, indicating stable generalization. Both models exhibit low standard deviation (less than 2%) and macro-averaged ROC-AUC values exceeding 0.99 across folds. The noise-robustness analysis also confirms MSFF-SAE’s superior measurement noise-resilience under increasing Gaussian perturbations. These findings highlight the potential of feature-fused stacked autoencoders as efficient and noise-resilient learners for automated assessment of heating non-uniformities of thermal tomographic images.