A Novel Semi-Supervised Learning for Industrial Edge Computing Platforms in Quality Prediction

The manufacturing industry is embracing Deep Learning (DL) and Edge Computing (EC) solutions to escalate productivity and computing efficiency. Proficient quality prediction (quantitative quality) of manufacturing processes and products is the foremost priority of industries, admirably accomplished by DL solutions. Industrial EC platforms host these DL solutions to maintain near-real-time performance and preserve privacy. Besides, the efficacy of such solutions relies on the abundance of labeled data, which is often a misery in manufacturing industries. Semi-Supervised Learning (SSL) and Data Augmentation (DA) paradigms are paramount in mitigating these issues. This work proposes an Ensemble of Self-Training SSL over MixUp-based DA (EnSeMUp) mechanism for Deep Neural Networks (DNNs) to acquire quality prediction under limited labeled data. The ensembled DNNs improve the training performance and confidence in pseudo-labeling the unlabeled data but demand higher storage and computation costs. To adhere to the resource constraints of the EC platform, magnitude-based pruning of ensembled DNNs is conducted. The proposed EnSeMUp mechanism is tested for three real-world manufacturing datasets and achieves a 30.35%–69.45% reduction in Mean Squared Error (MSE) loss compared to limited labeled data, 2.44%–58.34% higher MSE loss than complete labeled data, and 70% of storage and computation saving.