物理规律监督的RC柱地震破坏模式可解释机器学习方法

A physics-supervised interpretable machine learning approach for seismic failure modes prediction of RC columns

针对传统机器学习模型难以有效兼顾预测准确率与可解释性,且存在模型预测结果与实际物理规律不相符的问题,提出物理规律监督的RC柱地震破坏模式可解释机器学习方法。首先结合RC柱的地震破坏模式演化机制、试验数据规律分析和工程经验,揭示RC柱关键特征参数与地震破坏模式之间的物理规律;然后采用多分类逻辑回归算法提出了RC柱地震破坏模式的可解释模型;进而融合物理规律和等比例K-折交叉验证算法确定了符合物理规律且预测误差最小的最优超参数;最后基于多分类逻辑回归可解释模型和最优超参数,建立物理规律监督的RC柱地震破坏模式多分类逻辑回归预测模型。分析结果表明:该模型可以有效兼顾准确率与可解释性,预测结果不仅满足关键特征参数与地震破坏模式之间的物理规律,而且可以合理反映RC柱地震破坏模式之间的竞争关系,避免了传统逻辑回归模型存在预测结果与实际物理规律不相符的缺陷;与传统的经验判别方法和机器学习模型相比,该模型的准确率可以分别提高12%~35%和4%~9%。

Traditional machine learning models have difficulties in balancing the prediction accuracy and model interpretability,as well as have inconsistent results between model predictions and actual physical laws.In order to overcome these limitations,a physics-supervised interpretable machine learning approach was proposed to predict the seismic failure modes of reinforced concrete(RC)columns.The physical laws between key characteristic parameters and seismic failure modes of RC columns were revealed first based on the seismic failure modes evolution mechanism of RC columns,laws analysis of test data,and engineering experience.Then,an interpretable model for predicting the seismic failure modes of RC columns was proposed by multiclass logistic regression algorithm.Furthermore,the optimal hyperparameter satisfying the physical laws and with the minimum error were obtained by encoding the physical laws and equal proportion K-fold cross-validation algorithm.Finally,a physics-supervised multiclass logistic regression(PMLR)prediction model for failure modes of RC columns was established based on the multiclass logistic regression interpretable model and optimal hyperparameter.The results show that the proposed model has satisfactory accuracy and natural interpretability.The prediction results not only meet the actual physical laws between key characteristic parameters and seismic failure modes of RC columns,but also reflect the competitive relationship between seismic failure modes of RC columns.The proposed model overcomes the limitation that the prediction results are inconsistent with actual physical laws by traditional logistic regression(LR)model.The prediction accuracy of the proposed model is improved by 12%-35%and 4%-9%respectively when compared with traditional empirical classification methods and machine learning models.