An Attention-Based Deep Sequential GRU Model for Sensor Drift Compensation

Abstract

Sensor accuracy is vital for the reliability of sensing applications. However, sensor drift is a common problem that leads to inaccurate measurement readings. Owing to aging and environmental variation, chemical gas sensors in particular are quite susceptible to drift with time. Existing solutions may not address the temporal complex aspect of drift, which a sequential deep learning approach could capture. This article proposes a novel deep sequential model named Concatenated GRU & Dense layer with Attention (CGDA) for drift compensation in low-cost gas sensors. Concatenation of a stacked GRU (Gated Recurrent Unit) block and a dense layer is integrated with an attention network, that accurately predicts the hourly drift sequence for an entire day. The stacked GRU extracts useful temporal features layer by layer capturing the time dependencies at a low computational expense, while the dense layer helps in retention of handcrafted feature knowledge, and the attention mechanism facilitates adequate weight assignment and elaborate information mapping. The CGDA model achieves a significant mean accuracy over 93%, outperforming several state-of-the-art shallow and deep learning models besides its ablated variants. It can greatly enhance the reliability of sensors in real-world applications.