当前位置: X-MOL 学术Comput. Ind. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Stable and explainable deep learning damage prediction for prismatic cantilever steel beam
Computers in Industry ( IF 8.2 ) Pub Date : 2020-12-20 , DOI: 10.1016/j.compind.2020.103359
Darian M. Onchis , Gilbert-Rainer Gillich

Deep learning models fulfill the goal of characterizing the condition of beams in an non-invasive manner by accurately classifying accelerometer data. But the high probabilistic accuracy achieved on the validation set, while being a necessary indicator is usually not sufficient in most operational situations. With the occurrence of a damage, the reliable prediction must be also explainable in human terms incorporating the features that generated that particular result. This will enhance the trust and also the possibility of correction for the future functioning conditions. For obtaining the interpretable model, we correlate model agnostic global and local explanations with the use of the LIME and respectively the SHAP algorithm. Since the local explanations might be unstable, we introduce a compound stability-fit compensation index as a quality indicator in order to accept an explanation. This index is computed using both the condition number and the R2 fit indicator. Extensive testing, showed us the benefits of our method to completely and trustfully characterize the location and the depth of damaged beams.



中文翻译:

棱形悬臂钢梁的稳定,可解释的深度学习损伤预测

深度学习模型可通过对加速度计数据进行准确分类来实现以无创方式表征光束条件的目标。但是,在大多数操作情况下,虽然作为必要的指标,但在验证集上实现的高概率准确性通常是不够的。随着损坏的发生,可靠的预测还必须以人类的方式加以解释,并结合产生特定结果的特征。这将增强信任,并为将来的运行状况提供更正的可能性。为了获得可解释的模型,我们将模型不可知的全局和局部解释与LIME和SHAP算法的使用相关联。由于当地的解释可能不稳定,为了接受说明,我们引入了复合稳定性拟合补偿指数作为质量指标。使用条件编号和R 2适配指示器。广泛的测试向我们展示了使用此方法可以完全可靠地表征受损光束的位置和深度的好处。

更新日期:2020-12-21
down
wechat
bug