A Generalizable and Interpretable Deep Transfer Learning Framework for Industrial Condition Monitoring under Noise-Induced Domain Shift

Abstract

Industrial condition monitoring systems increasingly operate under heterogeneous deployment conditions where the statistical distribution of monitoring data differs from that observed during model development. Such heterogeneous domain shifts, arising from both cross-process variability and realistic industrial noise, significantly degrade the performance and reliability of conventional deep learning models. Addressing this challenge constitutes the primary objective of this work.
This paper proposes a generalizable and interpretable monitoring framework based on an adaptive Transfer Learning (TL) strategy combined with a hybrid CNN-Kolmogorov-Arnold Network (KAN) architecture. First, the impact of realistic industrial noise (Gaussian, Colored, Impulsive, and Mixed) is statistically characterized using both univariate and multivariate tests, demonstrating significant domain shifts between noise-free and noisy conditions. Subsequently, a pre-trained model is adapted using original (noise-free) data from the target domain, while validation and online evaluation are performed under noisy conditions to emulate realistic deployment scenarios.
The proposed CNN-KAN model combines convolutional layers for robust temporal feature extraction with KAN-based functional representations that enhance interpretability and structured learning. Additionally, an adaptive transfer mechanism dynamically selects the optimal fine-tuning configuration, effectively balancing stability and plasticity during domain adaptation. Experimental evaluations were conducted on two heterogeneous industrial processes (DAMADICS and GPN) involving both fault diagnosis and cyberattack detection scenarios. Results demonstrate that the proposed framework achieves high classification performance under nominal conditions while maintaining stable degradation behavior under severe noisy environ-ments. Furthermore, quantitative interpretability analyses based on latent feature importance, class-wise contribution maps, and gradient-based feature attribution confirm that the CNN-KAN architecture provides consistent and structured explanations aligned with the monitoring process. Comparative analyses against existing transfer learning and domain adaptation approaches fur-ther demonstrate the effectiveness of the proposed framework for industrial condition monitoring under complex domain-shifting conditions.
Overall, the obtained results demonstrate that the proposed adaptive transfer learning framework provides an effective and interpretable solution for industrial condition monitoring under heterogeneous domain-shifting conditions.