【摘要】 针对桥梁结构健康监测中振动信号的缺失及复杂信号动态特性难以准确捕捉导致信号预测精度不足的问题，提出一种融合麻雀搜索算法(sparrow search algorithm,SSA)与互补集合经验模态分解(complete ensemble empirical mode decomposition,CEEMD)-双向长短期记忆神经网络(bi-directional long short-term memory,BiLSTM)的桥梁信号预测方法。所提出的混合方法将缺失数据插补任务转化为时间序列预测任务，首先利用CEEMD对桥梁监测信号进行多尺度分解，提取固有模态函数(IMFs)以增强信号的特征表达能力；采用Bi-LSTM神经网络训练该数据集，并引入SSA对CEEMD-BiLSTM模型的关键超参数进行优化，以提升模型对非线性和时序特性的学习能力，从而实现对原始实测信号数据的子序列进行预测。基于实际桥梁长期监测数据，开展信号预测实验，分别与传统单一预测模型、BiLSTM模型及未引入优化的CEEMD-BiLSTM模型进行对比。研究结果表明，融合SSA和CEEMD-BiLSTM方法能够显著提高信号预测的准确性和稳定性，相较于其他单一与组合预测模型，平均预测误差明显降低。更多还原

【Abstract】 To overcome the limitations of incomplete vibration signals in bridge health monitoring and the difficulty of accurately capturing the dynamic characteristics of complex signals, a bridge signal prediction method that integrates the sparrow search algorithm(SSA) with complementary ensemble empirical mode decomposition(CEEMD) and a bidirectional long short-term memory(BiLSTM) network was proposed. The task of filling in missing data was transformed into a time series prediction task by the proposed hybrid method. It begins by using CEEMD to perform multi-scale decomposition on bridge monitoring signals, extracting intrinsic mode functions(IMFs) to enhance the signal’s feature representation. The data set was then trained using a Bi-LSTM neural network, and SSA was introduced to optimize the key hyperparameters of the CEEMD-BiLSTM model, thereby enhancing the model’s ability to learn nonlinear and temporal characteristics, thus enabling the prediction of subsequences from the original measured signal data. Using long-term monitoring data from an actual bridge, a series of signal prediction experiments were conducted, and the proposed method was compared with conventional models, including standalone BiLSTM and the non-optimized CEEMD-BiLSTM. The results show that the SSAoptimized CEEMD-BiLSTM approach significantly enhances prediction accuracy and stability while reducing the average prediction error compared to other models.更多还原