A simple yet efficient unsupervised anomaly detection network for industrial machine
Danxuan Wu, Qijun Chen · 2023
Anomaly detection, aiming to detect items, events, or observations that do not match expected patterns or other items in a data set, is a crucial problem in many fields such as network intrusion detection, fraud detection and so on. Due to the rarity of labeled anomalous datasets, the research aims at detecting anomalies without anomalous samples and only using normal ones which called unsupervised learning. While traditional machine learning devoted to find a hard marginal between abnormal and normal ones, some unavoidable disturbance along with the signal can cause deviation. Moreover, the deep learning research of this subject focus on construct an autoencoder network only using normal data to memory the patterns of the normality. However, considering the generalization ability of neural network, it may occur that abnormal ones can reconstruct well thus leading the mistake of predicting. In this passage, to overcome the above two problem, we design a noise generator to simulate some reasonable noises, utilize and combine the neural network and memory bank to strengthen the ability of the network to memorize the patterns of the normal instances. We experiment and compare our algorithm on the MIMII sound anomaly dataset and the corresponding results shows the effectiveness and efficiency of our approach. It reaches 91.37 % in Accuracy and 93.76 % in F1 score. To our knowledge, it is the state of the art work up to now.