Reliable geological hazard warning depends on accurate sensing data. In order to solve the problems of large noise and long time sequence characteristics of geological monitoring sensor data, we propose a method to identify abnormal data of geological disaster monitoring based on generative adversarial network. Firstly, the RandAugment algorithm is used to enrich the diversity of training data and improve the robustness to noise. Secondly, multi-head self-attention mechanism is used to extract long time series features, and the stability of early warning performance is improved by adversarial training mechanism. Experiments on four real-time series sensor data streams extracted from hidden geological disaster points in Shaanxi Province show that the proposed method has a 5%-10% improvement in AUROC and F1 indexes, compared to widely used machine learning methods.