Yankai Cao

Ensuring the safe and reliable usage of Lithium-ion batteries (LIBs) necessitates accurate degradation modeling. While data-driven methods offer promising prospects for modeling battery degradation, the intricate structures often make them specific to datasets. Furthermore, the black-box nature of data-driven models complicates the understanding of their decision-making process. In this thesis, we delve into data-driven modeling of battery degradation, focusing on capacity estimation and cycle life prediction, to address the challenges of generalizability and interpretability.To improve the generalizability of the model, we first propose adopting a simple and robust machine learning model, partial least squares regression (PLSR), for joint battery capacity estimation and remaining useful life (RUL) prediction. Experimental results on three battery cells cycled at varied conditions demonstrate superior generalizability of the suggested model over complex and sophisticated methods. Another approach we propose to improve the generalizability of the model is to use transfer learning. This approach presents excellent performance in handling the significant diversity in different types of batteries, as it can transfer the knowledge contained in well-studied batteries to a new battery. The key idea involves training a model in one type of battery with sufficient data. Then, the model can be applicable to a new type of battery by fine-tuning some parameters with limited data. Experimental results confirm that transfer learning can effectively enhance the generalizability of data-driven models in capacity estimation and cycle life prediction across different battery types.To build interpretable models, we advocate the use of decision trees for capacity estimation. We start with a classic regression tree with parallel splits for capacity estimation, but it requires a tree depth of 11 to achieve satisfactory performance. To address this challenge, we adopt optimal regression trees with hyperplane splits and propose a novel algorithm, DE-LR-ORTH, to train such a tree. DE-LR-ORTH initially conducts a one-step optimal hyperplane split for each branch node via differential evolution, followed by logistic regression-based fine-tuning to achieve overall optimality. Additionally, a GPU-accelerated implementation is proposed to significantly reduce the training time. Experimental results reveal a 1.0% capacity estimation error at depth 6 while maintaining high interpretability.

View record

Soft sensor technology is an effective way to measure parameters that are hard to measure in real time. It is of significant importance for the monitoring, control, and optimization of industrial production processes. With the richness of process data and the rapid development of machine learning techniques, data-driven soft sensor technologies are increasingly favored. Although soft sensor models have great potential and value in industrial applications, they still face significant challenges, particularly in the areas of model interpretability and stability. Ensuring interpretability and stability is crucial because it directly impacts the reliability and safety of operations in hazardous industrial environments.This dissertation provides a detailed exploration of soft sensor technologies, focusing on enhancing their interpretability and stability for industrial process monitoring. Chapters 2 and 3 focus on improving the interpretability of soft sensors. Chapter 2 introduces the Extra Trees (ET) algorithm and employs SHapley Additive exPlanations (SHAP) to enhance the interpretability of this inherently accurate but complex model. Chapter 3 explores interpretable feature selection techniques, particularly emphasizing the role of SHAP in the selection of meaningful features from complex industrial data. Subsequently, we utilize the selected interpretable features to establish a simple soft sensor model. In Chapter 4, the main topic shifts to the stability of the soft sensor model; we propose a stable learning algorithm based on the generation of virtual samples to improve stability in the face of industrial disturbances and data scarcity. Chapter 5 delves into the role of causality in soft sensor modeling, demonstrating how mining causal relationships between variables can significantly improve both stability and interpretability. It also emphasizes the importance of incorporating the knowledge of the process to ensure precision in the discovery of causal relationships. Chapter 6 presents two methods for extracting unsupervised and supervised latent causal features. By extracting latent causal features, not only is interpretability retained, but our model also becomes more robust. Finally, we analyze the main contributions and consider how they can be utilized in industrial contexts to improve the efficiency, safety,reliability, and interpretability of soft sensors.

View record