Détection d’anomalies robuste basée sur l’apprentissage profond semi-supervisé dans l’industrie du futur

Anomaly detection (AD) in smart manufacturing systems involves identifying abnormal data points or events from a large volume of data from various domains such as engine temperature, bearing vibration, and product images. AD methods using deep learning are gaining attention for their ability to automatically extract key features from large datasets. However, their performance can be limited in practice due to data noise and a lack of knowledge about anomalies. In this thesis, we propose various solutions to enhance the performance of deep learning-based AD methods and explore their applications across different data domains, including images and time series data. First, we demonstrate that when trained through a semi-supervised process (i.e., using only normal data), anomaly detection methods can experience reduced performance if the training data contains noise. We propose a method to separate noise from the training data while training the model simultaneously. We experimented with this approach on time series data, such as temperature and vibration measurements. The results show that this method outperforms current state-of-the-art models, especially for noisy data. For spatially dependent data, such as images, we introduce a novel augmentation technique aimed at addressing logical anomalies, which involve incorrect combinations of normal objects. This technique generates anomalies by modifying the vectors that encapsulate the general information of the image. The results suggest that integrating this augmentation technique can enhance the performance of the AD model without impacting computation time or model complexity.