基于全矢谱-深度置信网络的转子故障诊断方法研究  

Rotor fault diagnosis based on full vector spectrum and deep belief network

作　　者：李泽东 李志农[1] 陶俊勇[2] 许贝 LI Zedong;LI Zhinong;TAO Junyong;XU Bei(Key Laboratory of Nondestructive Testing, Ministry of Education, Nanchang Hangkong University, Nanchang 330063, China;Laboratory of Science and Technology on Integrated Logistics Support, National University of Defense Technology, Changsha 410073, China)

机构地区：[1]南昌航空大学无损检测技术教育部重点实验室,南昌330063 [2]国防科学技术大学装备综合保障技术重点实验室,长沙410073

出　　处：《兵器装备工程学报》2022年第1期48-54,共7页Journal of Ordnance Equipment Engineering

基　　金：国家自然科学基金项目(52075236);装备预研基金项目(6142003190210);南昌航空大学研究生创新专项资金项目(YC2020-056)。

摘　　要：针对传统方法采用单通道信息进行设备故障诊断容易造成误判以及传统故障诊断需要大量专家经验知识的不足,结合全矢谱技术在多通道信息融合中可以全面反映振动信号特征的优势,以及深度学习具有强大的自特征提取能力和较好的模式识别能力,提出了一种基于全矢谱-深度置信网络的转子故障智能诊断方法。对采集到的多通道的机械振动信号利用全矢谱技术进行融合,得到融合信号的主振矢、副振矢和振矢角。将融合后的信号输入到深度置信网络(DBN)中进行训练,利用多个受限玻尔兹曼机无监督预训练的方式层层堆叠进行前向传播,减少模型直接单向训练时的复杂度。然后利用反向传播对模型进行监督优化参数。最后,输出层采用Softmax分类器进行故障模式识别。提出的方法通过转子故障诊断验证,并与全矢谱-DNN和单通道-DBN做比较,提出的方法优于全矢谱-DNN和单通道-DBN方法,能够很好地融合多通道信息,并具有较高的识别率。In the traditional method for equipment fault diagnosis,a misjudgment can be caused by the incompleteness of single-channel signals,and a large amount of professional experience and knowledge were also required.Combined with the advantage of the full vector spectrum(FVS),which can reflect the vibration signal’s features in multi-channel information fusion,and deep belief network(DBN)which has strong self-feature extraction capabilities and better pattern recognition capabilities,an intelligent fault diagnosis method based on full vector spectrum and deep belief network was proposed.In the proposed method,FVS was used to fuse the collected multi-channel mechanical vibration signals to obtain the main vibration vector,auxiliary vibration vector and vibration vector angle of the fusion signal,which is input the DBN for training.Multiple restricted Boltzmann machines were used to perform the unsupervised pre-training for forward propagation layer by layer to reduce the complexity of the mode from the direct one-way training.Then,the supervised backpropagation was used to optimize the parameters.The output layer used a Softmax classifier for fault pattern recognition.The proposed method was verified by the fault diagnosis experiment of rotor system and compared with FVS-DNN method and single channel-DBN method.The experiment results show that the proposed method is superior to FVS-DNN method and single channel-DBN method and can fuse multi-channel information well and has high recognition rate.

关 键 词：深度置信网络 全矢谱 故障诊断 信息融合 自特征提取 

分 类 号：TH17[机械工程—机械制造及自动化] TG115.27[金属学及工艺—物理冶金]