Sensitivity Analysis-Based Explainable Diffusion Transformers for Anomaly Detection in Consumer Electronics Manufacturing

Consumer electronics play a crucial role in the artificial intelligence internet of things (AIoT), with anomaly detection (AD) being particularly critical for the consumer product manufacturing industry. However, existing AD methods suffer from limitations such as poor detection accuracy and lack of explainability, hindering their widespread adoption in industrial manufacturing. To address these issues, we propose SADiTAD, a Sensitivity Analysis-based Diffusion Transformer for Anomaly Detection. This model comprises a diffusion transformer (DiT)-based reconstruction enhancement sub-network and a vision transformer (ViT)-based detection sub-network. In the DiT sub-network, we introduce a structural similarity index measure (SSIM)-guided one-step denoising method to expedite the denoising process. Additionally, to enhance the model's explainability, we develop a sensitivity analysis-based ViT (SA-ViT) model, which evaluates the sensitivity of input embeddings to various image regions to determine if the fault region is being accurately identified during anomaly detection. The proposed SADiTAD model has been evaluated on public datasets MVTec AD and VisA, demonstrating superior performance over existing state-of-the-art anomaly detection methods and providing enhanced explainability.