Adaptive transfer learning with C-P-T sensor fusion for few-shot anomaly diagnosis in injection molding process

Injection molding (IM) is a pivotal polymer processing technique that produces precision components across industries. Achieving consistent product quality remains a significant challenge due to inherent variations between material batches, frequent changes in operational conditions, and the critical need to detect and diagnose anomalies. These anomalies stem from deviations in cavity pressure, temperature, and material properties. Addressing these challenges, this paper proposes a few-shot anomaly diagnosis framework based on Capacitance-Pressure-Temperature (C-P-T) multi-sensor fusion and adaptive transfer learning, targeting 12 distinct anomaly classes. Based on C-P-T sensor data we design a symmetric dual-branch Convolutional Neural Network (CNN) architecture to extract domain-invariant features and a novel hybrid loss function to optimize few-shot learning. For cross-domain adaptation, a hierarchical transfer strategy freezes the first three convolutional layers of the source model while fine-tuning the last two layers and classifier, enabling rapid deployment from Acrylonitrile Butadiene Styrene (ABS) to Polypropylene (PP) material domains and different machine domains. Industrial validation demonstrates 98.93% accuracy in the source domain with only 25 samples per anomaly class, and 94.2% accuracy in target domains with just 5 samples. This framework provides high-precision intelligent process monitoring for flexible injection molding production.