Significant advances in utilizing deep learning for anomaly detection have been made in recent years. However, these methods largely assume the existence of a normal training set (i.e., uncontaminated by anomalies), or even a completely labeled training set. In many complex engineering systems, such as particle accelerators, labels are sparse and expensive; in order to perform anomaly detection in these cases, we must drop these assumptions and utilize a completely unsupervised method. Moreover, only identifying the anomaly is insufficient: operators of these complex systems need additional localization information to identify the root cause of the anomaly and make an informed response. In this paper, we introduce the Resilient Variational Autoencoder (ResVAE), a deep generative model that is designed for anomaly detection, is resilient to anomalies in the training data, and yields feature-level anomaly attribution. During training, the ResVAE learns the anomaly probability for each sample as a whole and for each individual feature, and uses those probabilities to ignore anomalous examples in the training data. We apply our method to detecting anomalies in the accelerator status at the SLAC Linac Coherent Light Source (LCLS). Using shot-to-shot data from the beam position monitoring system, we identify and characterize several types of anomalies apparent in the accelerator, including many instances of known failures modes (e.g., beam loss) that are missed by current detection methods.
|Consider for long presentation||No|