Attention-based Dependability Prediction for Industrial Wireless Communication Systems

Wireless communication systems are ever-increasing important for industrial applications, supported by organizations such as the German Electro and Digital Industry Association (ZVEI), 5G Alliance for Connected Industries and Automation (5G ACIA), and 3rd Generation Partnership Project (3GPP). Industrial wireless communication systems (IWCSs) have high requirements for dependability, where dependability prediction can support to assess and improve the IWCSs. With the fast development of machine learning techniques, several Long Short-term Memory (LSTM) models have been proposed and indicate effectiveness for the dependability prediction task. However, these models ignore the truth that wireless devices are always resource-constrained and relationships between logical links can increase the prediction accuracy. Therefore, we propose the attention-based dependability prediction model which includes a sequence-to-sequence model and attention mechanism. We vary the attention mechanism with several Transformer variants to reduce time complexity and conducted experiments on a realistic measured data set. We compared the execution time and prediction performance of these models. Results indicate that the Sinkhorn-based model can meet the real-time requirement and has the best performance, and the Performer-based model has the lowest execution time, which can be applied for harsh real-time industrial applications.