Work

A Novel Attention-Based Hybrid Deep Learning Framework for Interpretable Subclass Identification in Paint Manufacturing

Year: 2026

Type: article

Abstract: Industrial paint manufacturing processes involve complex interactions among heterogeneous sensor measurements, making fine-grained subclass identification challenging. Existing studies in fault diagnosis and attention-based learning primarily focus on homogeneous data representations, image-transformed signals, or single-variable optimization, limiting their ability to model cross-domain dependencies in real-world industrial systems. Moreover, severe class imbalance in manufacturing datasets further reduces the reliability of conventional deep learning approaches. To address these limitations, this study proposes an attention-based hybrid model (ABHM) that directly operates on multisensor data via sensorwise tokenization and dual multihead self-attention. Rather than introducing a fundamentally new attention mechanism, the contribution lies in a domain-specific architectural adaptation that explicitly captures both intradomain temporal dependencies and interdomain physical interactions across heterogeneous process variables without requiring signal-to-image transformation. An imbalance-aware training strategy is incorporated into the end-to-end learning framework via synthetic minority oversampling technique-based oversampling to enhance fine-grained subclass discrimination. Experimental evaluations on real-world paint manufacturing data demonstrate that ABHM consistently outperforms state-of-the-art deep and attention-driven baselines in classification accuracy, robustness to class imbalance, and provides enhanced diagnostic transparency through attention-based feature attribution. These findings indicate that the proposed framework effectively bridges existing research gaps in industrial multisensor subclass identification and provides a scalable solution for intelligent manufacturing systems.

Source: IEEE Transactions on Industrial Informatics

DOI: 10.1109/tii.2026.3730745

Authors: Aziz Kubilay Ovacikli, Mert Yagcioglu

Institution: Istanbul Arel University

Language: English

Topic: Fault Detection and Control Systems

Subfield: Control and Systems Engineering

Field: Engineering

Domain: Physical Sciences

SDG: Industry, innovation and infrastructure

Open Access status: hybrid

Estimated APC paid: $2,645