【摘要】 针对沉箱式水下防护设施服役期间落物撞击识别困难的问题，基于有限元法与机器学习算法，提出了一种结合信号特征提取的撞击位置与应力识别方法。首先，利用有限元软件建立防护设施数值模型，模拟落物撞击过程并提取应力响应信号；其次，提取应力信号的峰峰值、峭度值、裕度脉冲指数和信号作为典型特征值，构建机器学习数据集；然后，选用神经网络、线性回归和支持向量回归（support vector regression，简称SVR）等算法进行训练，并重点分析了不同特征值对识别精度的影响。结果表明：选用应力积分作为特征值时的识别误差最小；其中撞击位置的平均识别误差在1.2 m以内，撞击应力的识别误差在3.4%以内；应力积分是沉箱式水下防护设施撞击识别的最优特征值。更多还原

【Abstract】 To address the inherent difficulties in identifying the impacts of falling objects on caisson-type subsea protection facilities during their service life, an identification framework for impact location and stress based on the finite element method(FEM) and machine learning is proposed in this study. First, a numerical model of the protection facility is developed to simulate the impact process and extract stress response signals. Second, representative time-domain features, namely peak amplitude, peak-to-peak value, kurtosis, clearance factor, impulse factor, and stress integral, are extracted from the stress histories to construct the machine learning datasets. Third, machine learning algorithms, including artificial neural networks(ANN), linear regression(LR), and support vector regression(SVR), are employed to train the predictive models, with particular emphasis on analyzing the influence of different features on identification accuracy. The results indicate that the identification error is minimized when the stress integral is used as the feature value. Specifically, the mean spatial error for impact localization is controlled within 1.2 m, and the relative error for stress prediction is below 3.4%. Consequently, the stress integral is demonstrated to be the most robust optimal feature for impact identification in such subsea protection facilities.更多还原