Semisupervised Soft Sensor via Variable Topology Optimization-Aided Discrepancy-Aware Convolutional Autoencoder for Sintering Process of Ternary Cathode Materials

The residual lithium content in ternary cathode materials (TCMs) serves as a critical quality indicator, necessitating the development of real-time soft sensors for effective process control. However, traditional soft-sensor methods encounter substantial challenges arising from the high dimensionality of sintering variables, the spatiotemporal coupling induced by multiphysics phenomena, and the mismatch in sampling frequency between process variables and residual lithium measurements. To this end, a semi-supervised soft sensor modeling method via variables topology optimization aided discrepancy-aware convolutional autoencoder (VTO-DACAE) for residual lithium content of TCMs is proposed. Firstly, the representative features are selected from high-dimensional sintering variables using expert experience combined with mechanistic knowledge. Subsequently, multi-correlation analysis is performed on the selected sintering variables, followed by topological optimization and reordering to generate multi-channel correlation feature maps, which characterize the complex spatiotemporal relationships inherent in the sintering variables. On this basis, a discrepancy-aware channel fusion module is designed, in which adaptive residual connections preserve the original topological features while a discrepancy-aware factor is introduced to enhance the discriminative learning of multi-channel features dynamically. Furthermore, a semi-supervised learning strategy is developed to leverage unlabeled data fully during the production process. Ultimately, the application of the proposed method to the actual sintering process of TCMs validates its effectiveness and superiority.