Distillation knowledge-based space-time data prediction on industrial IoT edge devices

In industrial Internet of Things (IIoT), the space–time data prediction algorithm is considered as one of the key technologies for supporting real-time monitoring and intelligent control. However, the complexity of existing algorithms is too high to be deployed on edge devices with limited computational capability. To solve this problem, a novel space–time data prediction algorithm based on knowledge distillation (KD-ST) is proposed to compress teacher network to multi-student networks. Specifically, generative adversarial network (GAN) discrimination and teacher outlier elimination (TOE) are developed to minimize the discrepancy between disparate networks and avoid training errors. Furthermore, a weight transfer strategy is adopted for saving training time. Experiment results demonstrate that compared with the state-of-the-art T-GCN, the proposed Transfer-LSTM improves the real-time response speed by 17.15 times, and the proposed Transfer-1DCNN further improves the real-time response speed by 30.20 times.