An Efficient Configuration Based on Deep Learning Models for Predictive Maintenance of Solar-Powered IoT Devices

Predictive maintenance is always a field that receives most concerns in industrial applications. Hence, with the recent development of deep learning, the field has become one of the hot topics for using different machine learning and deep learning methods to solve many challenging cases. However, with IoT devices powered by solar panels, its performance and lifespan are not only affected by the main tasks but also gain the impact from external factors. Hence, the data collected from these IoT devices are often turbulent, unlike most of devices that are in-house established with stable power supply. Because of the mentioned problem, the predictive maintenance for solar-powered IoT devices can be considered a difficult challenge. Therefore, conventional models have show its limitations due to unstable, rare case data for training. In this study, we have carried out different configurations of depth and wide network with several novel deep learning models from Recurrent Neural Network (RNN), Long-Short Term Memory (LSTM) to Gated Recurrent Unit (GRU) for analysis and evaluation. Eventually, based on our study of model configuration, in order to make model converge faster, achieve higher performance while having lower consumption in computation cost and memory for small device integration, we propose the combination of GRU and LSTM that can achieve up to 72.7% in AUC score compared to 55.7%, 64.6%, 57.9% of two-layer RNN, two-layer LSTM and two-layer GRU respectively. The proposed model can adapt better to prediction of whenever solar-powered IoT devices are failed in the next cycle.