Self-attention based Difference Long Short-Term Memory Network for Industrial Data-driven Modeling

Abstract

In modern industry, soft sensors provide real-time predictions of quality variables that are difficult to measure directly with physical sensors. However, in industrial processes, changes in material properties, catalyst deactivation, and other factors often lead to shifts in data distribution. Existing soft sensor models often overlook the impact of these distribution changes on performance. To address the issue of performance degradation due to changes in data distribution, this paper proposes a self-attention based Difference Long Short-Term Memory (SA-DLSTM) network for soft sensor modeling. By employing self-attention, industrial raw data is refined to facilitate the extraction of nonlinear features, thereby reducing the difficulty in modeling. A Difference Channel is designed to perform correlation analysis and select significant features from the raw data, followed by extracting the difference information that can reveal changes in the data distribution. The SA-DLSTM soft sensor model is established and validated on two benchmark industrial datasets: Debutanizer Column and Sulfur Recovery Unit. Comparisons with benchmark models, and state-of-the-art models show that SA-DLSTM achieves the best performance across all evaluation metrics, demonstrating the effectiveness of the proposed model.