Abstract 

To address the challenges of static feature indicators and high algorithmic complexity in lithium-ion battery state of health prediction, this study proposes a dynamic prediction framework integrating Gaussian Mixture Models with Dirichlet Process online clustering. A multidimensional dynamic feature extraction mechanism is constructed to mine highly representative indirect health indicators from voltage, temperature, and other time-series signals during charge/discharge cycles. To reduce computational complexity, Dirichlet Process clustering is employed for real-time data partitioning, coupled with Welford's algorithm for recursive model parameter updating. Experimental validation on NASA battery datasets tests demonstrates that the proposed method achieves a 1.14% reduction in mean absolute error compared to traditional Gaussian Process Regression, while maintaining high robustness under noise interference. During online prediction, the single reasoning time consumption is only 0.068 ms, which is 83.6% more efficient than the parameter estimation method based on the expectation maximization algorithm, meeting the millisecond-level real-time requirements of Battery Management Systems. This research provides a real-time, adaptive SOH monitoring solution for BMS operating under complex conditions.

相关链接：

https://ieeexplore.ieee.org/document/11347204