Page 116 - 《振动工程学报》2025年第8期
P. 116

第 38 卷第 8 期                       振  动  工  程  学  报                                  Vol. 38 No. 8
               2025 年 8 月                      Journal of Vibration Engineering                       Aug. 2025



                         基于神经常微分方程的机械故障诊断方法



                                                   郭     彪 , 李志农       1,2
                                                            1
                                   (1. 南昌航空大学无损检测技术教育部重点实验室,江西 南昌 330063;
                                      2. 汕头大学智能制造技术教育部重点实验室,广东 汕头 515063)


              摘要: 针对传统的深度学习故障诊断方法中存在架构可解释性差以及盲目堆叠层数导致的参数增加和内存消耗等问题,将神
              经常微分方程(neural ordinary differential equations, NODE)引入到机械故障诊断中。搭建面向机械故障诊断的神经常微分方
              程网络架构,利用神经网络参数化隐藏状态的导数代替指定隐藏层的离散序列。通过构建故障数据与故障类型的非线性关
              系,使用常微分方程求解器(ODE solver)完成对不同故障类别的分类任务,形成一种端对端的故障诊断模式。将该方法应用
              到机械故障诊断领域,搭建特定的神经常微分方程网络模型,通过故障数据的输入完成对不同故障类别的分类任务。将该模
              型应用到航空发动机主轴轴承故障诊断中,并与残差网络模型的故障诊断方法进行对比。试验结果表明,在确保准确率不降
              低的情况下,该方法不仅减少了内存消耗,而且将模型参数数量减少了将近五倍。

              关键词: 故障诊断; 神经常微分方程; 动力学系统; 残差网络
              中图分类号: TH133.33    文献标志码: A    DOI:10.16385/j.cnki.issn.1004-4523.202307050

                          Mechanical fault diagnosis method based on neural ordinary

                                                  differential equations


                                                            1
                                                   GUO Biao , LI Zhinong 1,2
                       (1. Key Laboratory of Nondestructive Testing Ministry of Education, Nanchang Hangkong University,
                     Nanchang 330063, China; 2. Key Laboratory of Intelligent Manufacturing Technology (Shantou University),
                                            Ministry of Education, Shantou 515063, China)


              Abstract: Based on the problems of poor interpretability, as well as parameter increase and memory consumption caused by blind
              stacking layers in traditional fault diagnosis method based on deep learning, Neural ordinary differential equation (NODE) is intro‑
              duced into mechanical fault diagnosis, the network structure of NODE for machinery fault diagnosis is constructed. In the construct‑
              ed structure, the derivatives of the parameterized hidden states of the neural network are used to replace the discrete sequences of
              the specified hidden layers. By constructing a nonlinear relationship between fault data and fault types, an ordinary differential equa‑
              tion solver (ODE solver) is used to complete the classification of different fault types, and an end-to-end fault diagnosis model is
              formed. The proposed method is applied to mechanical fault diagnosis to build a specific NODE network model, and the classifica‑
              tion task of different fault categories is accomplished through the input of fault data. The constructed model is applied to the fault di‑
              agnosis of spindle bearing in the aircraft engine, and compared with the fault diagnosis method based on residual network model.
              The experimental results show that the constructed model and residual network model have satisfactory accuracy. However, the
              constructed  model  not  only  reduces  the  memory  consumption,  but  also  reduces  the  number  of  model  parameters  by  almost  five
              times.
              Keywords: fault diagnosis; neural ordinary differential equation; dynamics system; residual network


                  在大数据背景下,机械故障智能诊断的深入研                          用深度神经网络对旋转机械进行故障诊断,并与贝
                                                                                                        [2]
              究和应用迎来了新的机遇,特别是基于深度学习的                            叶斯、支持向量机等算法进行比较。JING 等 利用
                                                        [1]
              机械故障诊断方法取得了很大的进展。LIU 等 利                          深度卷积网络对齿轮箱进行自适应特征提取,并证

                  收稿日期: 2023-07-17; 修订日期: 2023-10-22
                  基金项目: 国家自然科学基金资助项目(52075236);江西省自然科学基金重点项目(20212ACB202005);智能制造技术
                          教育部重点实验室(汕头大学)开放课基金资助项目(STME2024002);广东省普通高校创新团队资助项目
                         (2020KCXTD012)
   111   112   113   114   115   116   117   118   119   120   121