Constraint Guided Autoencoders to Enforce a Predefined Threshold on Anomaly Scores: An Application in Machine Condition Monitoring

Abstract

Anomaly detection (AD) is an important task in a broad range of domains. A popular choice for AD are Deep Support Vector Data Description models. When learning such models, normal data is mapped close to and anomalous data is mapped far from a center, in some latent space, enabling the construction of a sphere to separate both types of data. Empirically it was observed: (i) that the center and radius of such sphere largely depends on the training data and model initialization which leads to difficulties when selecting a threshold, and (ii) that the center and radius of this sphere strongly impacts the model AD performance on unseen data. In this work, a more robust AD solution is proposed that (i) defines a sphere with a fixed radius and margin in some latent space and (ii) enforces the encoder, which maps the input to a latent space, to encode the normal data in a small sphere and the anomalous data outside a larger sphere, with the same center. Experimental results indicate that the proposed algorithm attains higher performance compared to alternatives and that the difference in size of the two spheres has a minor impact on the performance.