Causal augmented ConvNet: A temporal memory dilated convolution model for long-sequence time series prediction

A number of deep learning models have been proposed to capture the inherent information in multivariate time series signals. However, most of the existing models are suboptimal, especially for long-sequence time series prediction tasks. This work presents a causal augmented convolution network (CaConvNet) and its application for long-sequence time series prediction. First, the model utilizes dilated convolution with enlarged receptive fields to enhance global feature extraction in time series. Secondly, to effectively capture the long-term dependency and to further extract multiscale features that represent different operating conditions, the model is augmented with a long-short term memory network. Thirdly, the CaConvNet is further optimized with a dynamic hyperparameter search algorithm to reduce uncertainties and the cost of manual hyperparameter selection. Finally, the model is extensively evaluated on a predictive maintenance task using the turbofan aircraft engine run-to-failure prognostic benchmark dataset (C-MAPSS). The performance of the proposed CaConvNet is also compared with four conventional deep learning models and seven different state-of-the-art predictive models. The evaluation metrics show that the proposed CaConvNet outperforms other models in most of the prognostic tasks. Moreover, a comprehensive ablation study is performed to provide insights into the contribution of each sub-structure of the CaConvNet model to the observed performance. The results of the ablation study as well as the performance improvement of CaConvNet are discussed in this paper.