Virtual sensor-based imputed graph attention network for anomaly detection of equipment with incomplete data

For the safe operation of complex equipment, it is essential to implement accurate anomaly detection on the key parts of equipment. However, due to the extreme conditions of the complex equipment, some sensors sometimes fail to collect signals, which makes existing methods unable to take full advantage of these incomplete multi source data. To better detect anomalies by incomplete data, we proposed a virtual sensor-based imputed graph attention network, which generates signals to impute the time of sensor record failure by generative adversarial network (GAN) and extracts the features of complete signals mixed with real signals and generated signals by graph attention network (GAT). Creatively, "virtual sensor ", a sensor representation is introduced into GAN as part of the input to make generated signal have the characteristic of its respective channel information. Through it, the signals of different channels as recorded by real sensors can be obtained. Additionally, the graph structure of multi-source signal is obtained by learning with no prior knowledge. Compared to existing methods with complete data, the proposed method is able to have better performances even without complete data. Furthermore, to demonstrate the ability for missing data imputation, we discuss the generation effect and downstream task performance of the proposed model and its variants in ablation experiments.