A hybrid-driven soft sensor model with symbolic representation for enhanced self-interpretability
Lingyun Wei, Han Liu, Runyuan Guo, Wenqing Wang, Xueqiong Tian · Measurement Science and Technology · 2025
Abstract Deep learning-based soft sensors often encounter the challenge of black-box problems, which lack interpretability and fail to provide intuitive mapping relationships. In complex industrial processes, physical sensors are susceptible to wear and tear, causing a divergence between test data distributions and training data, thus undermining sensing performance stability. To address these challenges, this paper proposes a hybrid-driven and self-interpretable soft sensor with symbolic representation, called the Kolmogorov–Arnold conditional autoencoder (KACAE). This framework employs symbolic representation to intuitively express the mapping relationships between variables, providing the model’s self-interpretability. It also incorporates domain knowledge to determine specific activation function types and simplify the network structure, achieving an effective balance between performance and computational complexity. Furthermore, an adaptive weighted mechanism is introduced to eliminate correlations between hidden features, enhancing model stability. Finally, the effectiveness, self-interpretability, and stability of the KACAE are validated through two industrial case studies.