【摘要】 针对水膜厚度传感器的温漂问题及盐度幅相漂移问题，提出了一种基于改进粒子群优化算法(IMPSO)优化反向传播神经网络(BPNN)的路面水膜厚度预测模型。该模型通过在传统粒子群优化算法(PSO)中引入非线性惯性权重和学习因子，有效避免了算法易陷入局部极值的问题。同时，IMPSO算法优化了BPNN的初始权值和阈值，克服了BPNN收敛速度慢、易陷入局部最优解的不足。实验结果表明，与遗传算法优化的BP神经网络(GA-BP)和粒子群优化的BP神经网络(PSO-BP)相比，IMPSO-BP模型在泛化能力上表现更为突出，其均方根误差(RMSE)为0.3043 mm,平均绝对误差(MAE)为0.2256 mm,拟合优度(R~2)为0.9824。在预测误差为±0.5 mm的范围内，预测准确率达93.52%。实验证实了该模型在路面水膜厚度检测中的高精度和强鲁棒性，为其在实际工程中的应用提供了可靠的理论依据和技术支持。更多还原

【Abstract】 In order to address temperature drift and salinity-induced phase-amplitude shifts in water film thickness(WFT) sensors, a pavement WFT prediction model was developed by integrating an improved particle swarm optimization(IMPSO) algorithm with a backpropagation neural network(BPNN). Nonlinear inertia weights and adaptive learning factors into the traditional particle swarm optimization(PSO) framework were introduced in the IMPSO algorithm, and the premature convergence to local optima was effectively prevented. The initial weights and thresholds of the BPNN were optimized, slow convergence and local minima trapping typically encountered with BPNN were overcome. The experimental results show that the IMPSO-BPNN model outperforms the genetic algorithm-optimized BPNN(GA-BP) and PSO-optimized BPNN(PSO-BP) in terms of generalization ability, and a root mean square error(RMSE) of 0.3043 mm, a mean absolute error(MAE) of 0.2256 mm, and a goodness of fit(R~2) of 0.9824 were achieved. With the prediction errors within ±0.5 mm, an accuracy of 93.52% was achieved. The experimental validation confirms the high precision and strong robustness of the model in pavement WFT measurement, providing a reliable theoretical and technical foundation for application in practical engineering.更多还原