Abstract:As a crucial indicator of the grinding process, the soft measurement of grinding fineness enables real-time estimation, reducing costs while improving production efficiency. However, in practical production, the accuracy and efficiency of soft measurement are significantly reduced due to the time-delay disturbances present in various stages of the grinding process. Therefore, handling these time delays is particularly important. Traditional time-delay identification algorithms rely heavily on accumulated expert knowledge and face challenges such as long computation times and the disorderliness of time-delay data, making it difficult to apply and adjust in real-time when dealing with the massive time-series data generated in grinding processes.This paper employs four zeroth order optimization algorithms—Tree-Structured Parzen Estimator (Tpe), Covariance Matrix Adaptation Evolution Strategy (Cma-Es), Genetic Algorithm (GA), and Quasi-Monte Carlo Sampling (QMC)—to optimize time delays. Pearson correlation analysis is performed on key variables in the grinding process to obtain an ordered time-delay optimization result, after which the time-series data from different sensors are aligned. Subsequently, an eXtreme Gradient Boosting (XGBoost) model is used to predict grinding fineness. The results show that all four optimization algorithms achieve coefficients of determination (R2) above 0.8, demonstrating higher accuracy and efficiency compared to predictions without time-delay alignment. Experimental validation with real grinding process data confirms that zeroth order optimization algorithms significantly reduce the time required for traditional time-delay optimization while effectively resolving the disorderliness of time-delay sequences, making them highly practical and meaningful.