Leak detection using deep-learning one-class classification methods

In this paper, we explore two deep-learning one-class classification methods derived from the field of image novelty detection. Specifically, we examine two GAN (Generative Adversarial Network) architectures designed primarily for monochromatic images. These methods are engineered to make classification decisions based solely on positively labeled training data, representing the regular or normal class, without incorporating any external information during either training or inference. Based on our experimental results with a standard image dataset, one of these two methods was selected for further application to spectrogram data of leaks. To facilitate this, we created a specialized dataset using an acoustic camera. This dataset comprised three categories: pure leaks, leaks with disturbances, and pure disturbances. After training the chosen model on a subset of pure leak data, it achieved perfect separation between this category and the pure disturbance category. However, the model struggled to positively classify certain elements in the disturbed leak category due to limitations in the dynamic range of the spectrograms.