Cellular fault prediction of graphical representation based on spatio-temporal graph convolutional networks

With the development of the 5th generation (5G) networks and the massive amount of data, the operation and maintenance (O&M) mindset needs to change from "post-maintenance"to "pre-prevention". In the wireless network, the fault has propagation and causality, and a fault in the current timestamp is also one of the factors that affect whether a fault occurs in the subsequent timestamp. Existing fault prediction methods study the time dependencies and inter-measure dependencies of key performance indicators (KPIs) while ignoring the important causal dependencies of faults in cellular networks. Therefore, it is still challenging to study the fault causal dependency in cellular networks. To tackle the above problems, we propose a novel framework for wireless cell fault prediction to deal with the problem of causal dependency. First, we build an undirected graph based on KPIs and fault-related knowledge collected from the base station through graphical representation. Second, we introduce the graphical model to learn the dependency relationship containing different KPIs at synchronous timestamps and the causal relationship between fault codes at asynchronous timestamps. Finally, we employed the attention mechanism of the graphical model to further strengthen the correlation between parameters during the training process. We conduct extensive prediction experiments on fault events (whether a fault occurs) and fault code (which type of fault occurs) tasks on fault datasets of real wireless cells. Experimental results show that our framing method is state-of-the-art and achieves higher accuracy than traditional fault prediction methods.