Real-time Prediction Method of Remaining Useful Life Based on TinyML

Tiny Machine Learning (TinyML) is a new research area aimed at designing and developing machine learning (ML) techniques for embedded systems and IoT units. Due to the limited resources of embedded system, neural network pruning is widely used to reduce resource occupation. To solve the problem that the Remaining Useful Life (RUL) of the equipment is difficult to calculate accurately and in real time, a pruning method based on L1 norm weight was designed to reduce the memory footprint and computational load of the neural network, and a lightweight two-dimensional convolutional neural network was constructed. Experimental results show that compared with random pruning, this method greatly reduces the influence of neural network parameter reduction on the accuracy of inference results. Meanwhile, a retraining method based on Adam optimization was used to make the RUL curve predicted by the retrained model more close to the real RUL curve. When the weight parameters are reduced by 30%, the model still maintains good prediction accuracy, and can realize the real-time prediction of RUL in the embedded system with limited resources.