A Deep Stacked Dual-Guided Autoencoders for Supervised Industrial Soft Sensing with Dynamic Adaptation Mechanisms

As industrial production processes become more complex and data-driven, traditional soft sensing models are facing multiple challenges, including difficulties in handling nonlinear systems and problems with time-varying characteristics. To address these challenges, this paper proposes a supervised dual-guided stacked autoencoder with dynamic adaptation mechanisms (JIT-SDGAE). In this model, a new pre-training objective function is proposed for supervising the pre-training process of the network, which allows the extraction of quality-related features, and it reconstructs the same original input data for each dual-guided autoencoder (DGAE) aiming to better characterize complex data patterns and learn quality-related features. And the introduction of just-in-time learning in the model enables the model to continuously learn from newly received data and quickly adapt to the dynamic changes in the production process, effectively solving the time-varying problem that is difficult to cope with by traditional methods. Finally, the efficiency of the JIT-SDGAE model was verified on the CO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> absorption column process.