Anomaly detection analysis based on temporal convolutional network model

Abstract

Temporal Convolutional Network (TCN) is a deep learning network based on temporal recursion, which is suitable for processing time series data with information correlation in the temporal dimension.TCNs are commonly applied in a variety of domains, including but not limited to anomaly detection and speech recognition. The DTAAD model is an improved model of TCN based on a two-layer temporal convolutional network and the Transformer module child, however, the DTAAD model has limitations in extracting multi-scale features, so in order to find a better way to identify irregular patterns in time series data., this article chose to go for improving the DTAAD model to achieve a better way to deal with complex time series data . Therefore this article chose to introduce a multi-scale temporal convolutional network to accomplish this improvement. The core of this improvement is to construct multiple layers using various sized convolutional kernels to collect temporal features at scales of different kinds. I also introduced a residual attention mechanism to enhance the representation of features through assigning variable weights to temporal and channel-wise components. This article conducted experiments using three different datasets (MBA, SWaT, SMAP) and the improved model successfully improves the performance in test metrics. This suggests that the improvement has led to an increase in the manifestation of the model in terms of anomaly detection.