基于可解释选择性集成框架的离心风机叶片裂纹损伤检测

Crack Damage Detection of Centrifugal Fan Blades Based on Interpretable Ensemble Selection Framework

接触式测量的振动信号是最为常用的旋转设备状态监测手段,但在工业现场离心风机叶片断裂前会产生异常的声响,然而其振动信号却依旧平稳。声信号属于一种非接触测量方法,包含强烈的背景噪声激励,使得其数据驱动网络的可信任程度较低。因此,为有效增强声信号驱动模型的可解释性、评估模型决策的可信度,提出一种可解释的选择性集成框架用于叶片裂纹的检测。首先,采用Alexnet提取的深度特征和可解释的辅助统计特征结合注意力权重机制构建一个多视角的基网络。其次,将声信号时频图像构建二阶张量作为网络的输入,对训练好的每个基网络的深度特征通过Grad-CAM确定激活映射面积并结合仿真结果计算网络的可信任度。接着,对于嵌入式的辅助特征构建一种Diversity Pick LIME(DP-LIME)的可解释模块,结合特征权重分布可视化基网络决策逻辑。最终,根据每个基网络的综合可信任度指标对叶片裂纹检测结果进行选择性决策融合。经离心风机实测数据验证,文中提出的可解释框架具有良好的叶片裂纹检测精度,并可有效提高模型可信度。

Contact-measured vibration signals is the most commonly used means of health monitoring for rotating equipment.However,in the industrial site,the centrifugal fan blade will produce abnormal sounds before breaking,but the vibration signal is still stable.The acoustic signal is a non-contact measurement method and contains strong background noise excitation,making its data-driven network less trustworthy.Therefore,to effectively enhance the interpretability of acoustic signal-driven models and evaluate the credibility of model decisions,an interpretable ensemble selection framework for blade crack detection is proposed.First,a multi-view base network is constructed by adopting Alexnet-extracted depth features and interpretable auxiliary statistical features combined with an attentional weighting mechanism.Secondly,the second-order tensor of the time-frequency image of the acoustic signal is used as input to the network,and the depth features of each trained base network are adopted to determine the activation mapping area by Grad-CAM and to calculate the trustworthiness of the network in conjunction with the simulation results.Next,a Diversity Pick LIME(DP-LIME)interpretable module is constructed for embedded auxiliary features,combined with a feature weight distribution to visualize the decision logic in the base network.Finally,the blade crack detection results are selectively fused for decision-making based on the composite trustworthiness index of each base network.The interpretable framework proposed in this paper has high detection accuracy for blade crack and can effectively improve the credibility of the model,which is verified by the actual measurement data of the centrifugal fan.