Modern manufacturing relies heavily on automated visual inspection to guarantee product quality. However, building reliable AI systems for industrial inspection remains challenging: defective samples are rare, production lines are geographically distributed, and AI decisions must be transparent to engineers and operators.
Researchers from UMIL developed new AI approaches for automated defect detection in System-on-Module (SoM) assembly processes. The work focuses on improving anomaly detection systems through explainable AI models and distributed learning strategies suitable for industrial environments.
OneN, a unified and explainable anomaly detection framework
Conventional industrial vision systems typically treat anomaly detection as three independent tasks:
- Detection – determining whether a defect exists
- Localization – identifying where the defect is located
- Classification – determining the defect type
While effective in some settings, this multi-stage pipeline increases system complexity and can reduce robustness when data is limited. To simplify the inspection pipeline, UMIL proposes OneN (One Network), a deep learning architecture capable of performing detection, localization, and classification within a single model.
Figure 1 – Detection, localization, classification
A key advantage of this design is native explainability. Instead of relying on post-hoc interpretation methods, OneN leverages attention maps produced by the Vision Transformer. These maps highlight the image regions that contribute most to the prediction and naturally act as pseudo-segmentation masks for defects. This enables operators to clearly visualize why the model flags a product as defective, increasing trust in the AI system.
Figure 2 – Example of inspected component used in the anomaly detection task
D-ADDA, a distributed anomaly detection architecture using generative models
Modern manufacturing ecosystems are often distributed across multiple factories, suppliers, and locations. In these scenarios, sharing data between sites is often restricted due to privacy concerns or intellectual property protection. To address this challenge, UMIL proposes D-ADDA (Distributed Attention-guided Diffusion-based Anomaly Detection Architecture). The approach uses diffusion generative models to synthesize realistic defects and improve anomaly detection models without requiring centralized datasets.
Figure 3 – D-ADDA distributed architecture
Each local node (production site) deploys a local Data Augmentation Module (DAM) that generates synthetic defective masks using Denoising Diffusion Probabilistic Models (DDPMs). The system produces:
- synthetic defect masks
- corresponding defective images
- augmented training samples
The remote server takes as input normal samples and defective masks to generate synthetic defective images using an attention-guided mechanism. Finally, the local applies cross-attention conditioning, enabling the model to realistically blend generated defects into normal images.
The D-ADDA generative model is designed to remain lightweight (≈48M parameters) and can be trained using only locally available defective samples, eliminating the need for large, centralized datasets. This approach allows each site to train anomaly detection models locally while preserving data privacy.
Figure 4 – D-ADDA image generation
Blog signed by: UMIL team