SHARE
Research Highlight

Self-Supervised Anomaly Detection via Normalizing Flows

Topic:

Brief: Oak Ridge National Laboratory (ORNL) researchers have developed a novel self-supervised anomaly detection approach based on normalizing flows with an active learning scheme for determining processing parameters in additive manufacturing.  

Accomplishment:  Anomaly detection (AD), separating anomalies from normal data, has many applications across domains, from manufacturing to cybersecurity. AD can be applied to detect anomalous from multimodal data, including images, time series data, tabular, graph and video.  We propose a robust and unified anomaly detection framework for general data and show its superior performance on anomaly detection accuracy and robustness over the state-of-the-art studies.  Our achievements are summarized below:  

  • Our anomaly detection framework built upon neural autoregressive flow (NAF-AD) enables us to explicitly learn likelihoods and thus assign accurate anomaly scores to detect anomalies.
  • We also propose an active learning scheme where new samples with lower likelihoods are merged into the normal training data for training a better detector to approach the decision boundary.
  • Our method outperformed the baseline methods on multiple time-series and tabular benchmarks and is applied to detect anomalies of the meltpool images from additive manufacturing (AM) at the Manufacturing Demonstration Facility (MDF). Our method also allows automatically labeling data, which will improve the working efficiency in AM.

 

(Left) NAF-AD method overview. NAF-AD transforms the data into multiple subspaces and learns a feature space by a neural network. Then we build an accurate density estimation via NAF and assign a higher likelihood to normal in latent space. NAF-AD is free to draw samples and transform them to feature space with explicit likelihoods via invertible mapping. (Right) Additive manufacturing demonstration. Our objective is to determine the correct processing parameters for the deposition of stainless steel.
(Left) NAF-AD method overview. NAF-AD transforms the data into multiple subspaces and learns a feature space by a neural network. Then we build an accurate density estimation via NAF and assign a higher likelihood to normal in latent space. NAF-AD is free to draw samples and transform them to feature space with explicit likelihoods via invertible mapping. (Right) Additive manufacturing demonstration. Our objective is to determine the correct processing parameters for the deposition of stainless steel. We utilized the coaxial laser camera to capture meltpool images to detect normal and anomalies via our proposed NAF-AD method.

Acknowledgement: This research was funded by the AI Initiative, as part of the Laboratory Directed Research and Development Program of Oak Ridge National Laboratory, managed by UT-Battelle, LLC, for the U.S. Department of Energy (DOE).    

Publication resulting from this work: 

  1. Zhang, J., Saleeby, K., Feldhausen, T., Bi, S., Plotkowski, A., and Womble, D. "Self-Supervised Anomaly Detection via Neural Autoregressive Flows with Active Learning."  NeurIPS 2021 Deep Generative Models and Downstream Applications.

Contact: Jiaxin Zhang (zhangj@ornl.gov)

Team: Jiaxin Zhang, Kyle Saleeby, Alex Plotkowski, Pradeep Ramuhalli, David Womble