Exploiting the generative adversarial framework for one-class multi-dimensional fault detection

Abstract

Fault detection is an important and demanding problem in industry. However, in many applications, abnormal data are difficult to be collected and are not available during training. In this paper, we propose a one class fault detection scheme for multi-dimensional problems, based on the unsupervised training of a Generative Adversarial Network (GAN). The generator tries to learn the manifold of normal behavior of the process, while the final decision of fault occurrence is taken from the discriminator. The network architectures of the discriminator and the generator and the hyper-parameters, that are used during training, are crucial for the stability of the GAN model. To ensure system convergence during training and to enhance the accuracy of the unsupervised classifier, a model selection algorithm is used. The latter one evaluates each trained model on a validation set of the normal training dataset, based on a proposed performance metric. Also, a method for evaluating the generating ability of each trained generative model is introduced, based on the reconstruction cost of an Autoencoder (AE). The proposed evaluation method of generative models, is used on the search algorithm, which selects the final classifier model. Finally, it is shown by experiments that the proposed system outperforms One Class SVM and Isolation Forest algorithm, two state of the art anomaly detection methods, on two testing cases.