基于SAConvFormer算法的焊接故障诊断在非平衡数据集上的应用  

Application of SAConvFormer Algorithm in Welding Fault Diagnosis on Imbalanced Datasets

作　　者：付惠斌 李晨 陈翀[3] Fu Huibin;Li Chen;Chen Chong(Sany Group Co.,Ltd.,Changsha 410100,China;Zhejiang Zhongtian Zhihui Installation Engineering Co.,Ltd.,Hangzhou,310015,China;Guangdong Provincial Key Laboratory of Cyber-Physical System,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]三一集团有限公司,长沙410100 [2]浙江中天智汇安装有限责任公司,杭州310015 [3]广东工业大学广东省信息物理融合重点实验室,广州510006

出　　处：《机电工程技术》2024年第7期18-22,共5页Mechanical & Electrical Engineering Technology

基　　金：国家自然基金资助项目(62302103)。

摘　　要：预测性维护在制造业中扮演着重要角色,而焊接设备的有效维护更是关键,旨在减少企业支出并实现无人车间的目标。然而,焊接领域的研究尚处于初级阶段,而深度学习在此领域的应用也相对较少。针对这一问题,提出一种基于空间注意力卷积Transformer(SAConvFormer)的故障诊断模型,以解决焊接设备故障预测的挑战。通过收集焊接过程中的数据,该算法通过空间注意力机制增强卷积神经网络,从而更准确地预测焊接过程中的各种故障类型。实验结果显示,SAConvFormer模型在预测正常类型故障方面的召回率达到了95%,误差仅为2%。对焊接偏差类型的故障,模型的召回率稳定在80%左右,而不完全熔合类型的故障预测准确性相对较低,但仍保持在70%以上。与传统算法相比,SAConvFormer模型在召回率上表现优异。这一研究成果不仅在技术上取得了进展,也为焊接设备的故障诊断提供了一种新的有效方法,具有重要的理论和实践意义。Predictive maintenance plays a crucial role in the manufacturing industry,with effective maintenance of welding equipment being particularly pivotal for reducing corporate expenditures and achieving the goal of unmanned workshops.However,research in the welding domain remains in its early stages,and the application of deep learning in this field is relatively limited.Addressing this issue,a fault diagnosis model is proposed based on the Spatial Attention Convolutional Transformer(SAConvFormer)to address the challenge of predicting welding equipment faults.By collecting data from the welding process,this algorithm enhances convolutional neural networks through a spatial attention mechanism,thereby more accurately predicting various fault types in welding processes.Experimental results demonstrate that the SAConvFormer model achieves a recall rate of 95% with an error margin of only 2% for predicting normal fault types.For welding deviation fault types,the model maintains a stable recall rate of approximately 80%,while the prediction accuracy for incomplete fusion fault types is relatively lower but still exceeds 70%.Compared to traditional algorithms,the SAConvFormer model exhibits excellent performance in terms of recall rate.This research not only represents technological advancement but also provides a novel and effective approach for fault diagnosis in welding equipment,with significant theoretical and practical implications.

关 键 词：预测性维护 焊接设备 故障诊断 空间注意力机制 深度学习 

分 类 号：TG43[金属学及工艺—焊接]