基于深度神经网络的多级高压提水泵故障辨识  

Identification of the faults of a multi-stage high pressure pump using deep neural network

作　　者：秦云[1] 张钊源 袁寿其[2] 司乔瑞[2] 杨宁[1] Qin Yun;Zhang Zhaoyuan;Yuan Shouqi;Si Qiaorui;Yang Ning(School of Electrical and Information Engineering,Jiangsu University,Zhenjiang 212000,China;National Pump and System Engineering Technology Research Center of Jiangsu University,Zhenjiang 212000,China)

机构地区：[1]江苏大学电气信息工程学院,镇江212000 [2]江苏大学国家水泵及系统工程技术研究中心,镇江212000

出　　处：《农业工程学报》2022年第14期27-34,共8页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然科学基金(32171895);国家重点研发计划“重大自然灾害监测预警与防范”重点资助专项(2020YFC1512403)。

摘　　要：多级高压提水泵属于复杂旋转机械,受设备自身结构影响,其故障特征频率易被强噪声淹没,且叶轮、轴承、齿轮等转子部件的故障振动信号相似度极高,导致现有故障辨识方法无法快速准确辨识其故障模式。针对上述问题,该研究提出一种改进的一维卷积长短期记忆神经网络(One-Dimensional Convolution and Long Short-Term Memory Neural Network,1D-CNN-LSTM)自适应故障辨识模型。首先通过贝叶斯优化算法获得给定模型超参数,再输入经互补集合经验模态分解降噪后的振动数据集,通过1D-CNN层自适应提取样本特征并作为LSTM层输入;利用LSTM层学习具有识别性的深层特征并训练模型,最后由输出层Softmax函数完成故障辨识与分类。多级高压提水泵试验台实测数据集对模型进行验证的结果表明:提出的1D-CNN-LSTM智慧故障辨识模型能够快速辨识关键转子部件的故障模式,且准确率可达97%,具有更好的抗噪能力和鲁棒性能,可为智慧应急供水与净水一体化系统的可靠运维技术研发奠定理论基础。Rotor parts have been the central components of"intelligent emergency water supply and water purification integrated equipment in mountainous areas and remote disaster areas".However,various failure modes can inevitably occur,due to high-speed and heavy-duty conditions for a long time in the field.The fault characteristic frequency is also easily submerged by the strong noise,depending mainly on the inner structure of the equipment.There is a high similarity in the fault vibration signals of rotor parts,such as impeller,bearing,and gear.The existing fault identification cannot quickly and accurately detect the fault mode.In this study,an improved adaptive fault identification model was proposed to realize the intelligent fault diagnosis of multi-stage high-pressure lift pumps using a one-dimensional convolutional neural network and long-term memory network(CNN-LSTM).In the CNN network,the last set of convolutional layers used for the feature extraction was then connected to the global mean pooling layer,in order to reduce the number of model parameters and integrate feature information.Relu activation function was then selected for the gradient disappearance,compared with the Sigmoid function.As such,the activation function of the pooling layer was changed to the Relu function in the CNN network.After that,the LSTM time series network was embedded to integrate the one-dimensional CNN and LSTM into a framework structure.Finally,the Softmax function was used as the classification output of the network model.The fault identification was as follows.First,the vibration signal was collected to preliminarily de-noise through CEEMD,where the de-noised data was divided into time series of equal length.Then,the data set was divided into the training,test,and validation set.Secondly,the optimal initial parameters of the model were obtained by Bayesian optimization.The model parameters were also fine-tuned layer by layer using back propagation.The model performance was evaluated by the validation set.Third,the test set was i

关 键 词：故障辨识 卷积神经网络 长短时记忆网络 多级高压提水泵 多尺度特征提取 

分 类 号：TH165[机械工程—机械制造及自动化] TH17