MCMDA: A continual learning and frugal AI based quality inspection mechanism for edge computing platforms

Industrial-edge empowered Deep Learning (DL) solutions facilitate indigenous Predictive Quality Inspection (PQI) of Mass Production (MP) processes and products. However, the existing DL-based PQI systems fail in Mass Customized and Personalized Product (MCPP) setups. DL models mandate maintenance to ensure sustainable solutions and challenges faced are (i) unavailability of previous product data, (ii) limited data availability of new MCPPs, and (iii) resource-efficiency and near-real-time execution of these mechanisms at the industrial edge. To address aforementioned issues, this paper proposes a Memory Aware Synapses (MAS) and frugal Artificial Intelligence (AI) solution named MAS-Cloning over MixUp-based Data Augmentation (MCMDA). The MAS scheme, a regularization-based continual learning scheme, ensures the learning of new MCPPs while preserving the past information under unavailability of previous product data. The limited data availability of new MCPPs’ is resolved via Frugal-AI solutions such as knowledge transfer in better weight initialization of new output heads (weight cloning) and MixUp-based data augmentation for less enchanted predictive/regression tasks. A mechanism to generate optimal synthetic data using MixUp-based data augmentation is incorporated for supreme DL-model performance. Compared to state-of-the-art schemes, MCMDA enhances model performance by 71.3%, reduces storage necessity by 32.92%, and minimizes training cost by 8.71%–57.71% for real-world injection molding use-case.