A data-efficient active learning architecture for anomaly detection in industrial time series data

What the paper says

Abstract Anomaly detection is becoming increasingly important and has found its way into manufacturing applications. The potential is seen in use cases such as maintenance cost reduction, machine fault reduction, or increased overall production based on industrial time series data. However, obstacles arise in practice. Supervised algorithms lack limited and expensive labeled training data, and unsupervised algorithms do not have the capabilities for evaluation and tracking. We propose a data-efficient architecture for anomaly detection using energy consumption time series data to address these limitations. To do so, we design an active learning model that optimizes an unsupervised model by integrating budgeted expert feedback. Our solution builds on an autoencoder to leverage latent space representations for an additional supervised feedforward network trained with expert knowledge labels to distinguish between normal data and anomalies. Four different strategies for querying the still-unlabeled data are compared so that the expert’s resources are used efficiently. We validate our concept in an industrial robotic screwdriving application based on energy data for condition monitoring. Findings for the application tested indicate that anomaly detection performance can be significantly increased by 59% for the F1 score with active learning compared to unsupervised models. Furthermore, models trained only on energy consumption data exhibit the same performance as models trained on difficult-to-obtain mechanical process data, thus confirming the practicality of our proposed approach and data efficiency for the use of easily accessible energy data in manufacturing applications. While our approach enables an active learning model to be added to an existing unsupervised model, it allows for straightforward benchmarking and extension to other manufacturing applications.