Nonlinear dynamic industrial process fault detection based on attention-enhanced LSTM autoencoder

Abstract

In the field of complex industrial process monitoring, capturing multivariate relationships in data is a major challenge for traditional shallow models. To address this, this paper proposes an attention-enhanced long short-term memory autoencoder (AELSTM-AE). AELSTM-AE utilizes the powerful feature extraction capabilities of deep learning methods to learn the intrinsic patterns of process time series data through a downscaling compression and upscaling reconstruction process. At the core of the AELSTM-AE is the information-enhanced attention (IEA) mechanism, which computes the dot-product similarity between corresponding encoding and decoding layer outputs in multiple subspaces. By weightily fusing these outputs, the IEA serves as an information supplement for the decoding layer, significantly reducing reconstruction errors. Trained on normal data, the reconstruction model captures long-term dependency features. In contrast to traditional error metrics, a novel statistical metric Loss is employed, which is the average absolute error across each time-series sample. This metric is consistent with the ability of the reconstruction model to extract long-term dependency patterns, improving the utilization of global information and thus enhancing the robustness of the method. The effectiveness of this method is validated on the Tennessee Eastman process and an actual blast furnace ironmaking process.