为实现车辆行驶过程中对桥面不平顺的准确识别，提出了一种基于改进粒子群优化（Improved Particle Swarm Optimization，IPSO）反向传播神经网络（Back propagation Neural Networks，BPNN）自适应未知输入离散卡尔曼滤波器的桥面不平顺识别算法。基于车桥耦合模型，以轮胎-桥面接触点处竖向位移为未知输入，以车轮位移、加速度及车身加速度作为观测向量，设计未知输入离散卡尔曼滤波器；通过改进粒子群算法得出各桥面等级下的测量噪声协方差矩阵的最优值，通过BP神经网络分段实时对桥面不平顺进行等级分类，二者结合实时更新卡尔曼滤波器中的测量噪声矩阵，从而自适应识别不同桥面不平顺。在不同行驶速度、不同桥面等级、不同车桥质量比等工况下进行了仿真分析，并设计了通过振动台进行车桥耦合实验的方案，根据振动台提取的车身加速度和车轮加速度等参数，通过二重积分去趋势项得到车轮位移，整理数据作为观测向量设计滤波器。为匹配振动台的四分之一悬架模型，等比例缩尺二自由度车辆模型参数和桥梁模型参数，以保证缩尺后桥梁的挠曲线和竖向位移的相似性。仿真和实验结果表明：基于IPSO-BPNN自适应未知输入离散卡尔曼滤波器能够在多种工况下自适应调节，相较于传统卡尔曼滤波器，识别的桥面不平顺均方根误差、最大绝对误差和相关系数分别提高11.29%、33.52%、2.84%，该算法不仅识别精度高，而且有很强的鲁棒性。

To accurately identify bridge roughness during vehicle travel, this paper proposes an IPSO-BPNN (Improved Particle Swarm Optimization, Backpropagation Neural Networks) adaptive unknown input discrete Kalman filter algorithm. Using a vehicle-bridge coupling model, the vertical displacement at the tire-bridge contact point is treated as the unknown input, while wheel displacement, acceleration, and vehicle body acceleration are used as the observation vector to design the unknown input Kalman filter. An improved particle swarm optimization algorithm is applied to obtain the optimal measurement noise covariance matrix for different bridge roughness levels. A BP neural network classifies bridge roughness levels in real time, and both methods work together to adaptively update the Kalman filter’s measurement noise matrix. Simulations under various driving speeds, bridge roughness levels, and vehicle-bridge mass ratios were conducted, and a shaking table experiment was designed to validate the approach. To match the quarter suspension model of the vibration table, the parameters of the two-degree-of-freedom vehicle model and the bridge model are scaled proportionally to ensure the similarity of the deflection curve and vertical displacement of the bridge after scaling. Results show that the proposed method improves root mean square error, maximum absolute error, and correlation coefficient by 11.29%, 33.52%, and 2.84%, respectively, demonstrating high accuracy and robustness.