Unsupervised anomaly detection for industrial data using Generative Adversarial Networks

Identifying an appropriate fault threshold is a challenging task in the context of anomaly detection and requires a significant amount of expert knowledge. The use of modern technology, such as artificial intelligence (AI), can facilitate the identification of an optimal threshold for a given data set during the training phase. At the beginning of a drive's operational life, only normal samples are available for training purposes. Furthermore, the data available for training can be used to distinguish between normal and abnormal conditions, with performance depending on the degree of fault. This is because the threshold depends on the samples that are most similar to the normal state. The proposed methodology uses a generative adversarial network (GAN) for anomaly detection. During the training phase, the GAN was trained using samples from the normal state. In the subsequent application phase, only the discriminator was used to detect faulty states. It is not necessary to provide a sophisticated interpretation of an anomaly score. The method was evaluated in two case studies. The first case study is the detection of bearing faults and the second is the detection of commutator surface faults using synthetic data. Both studies achieved near error-free prediction using both the Case Western Reverse University (CWRU) bearing data set and an own bearing data set. In addition, this work provides some heuristics for the training process, so that no faulty data were needed to evaluate the training performance.