基于稀疏局部嵌入深度卷积网络的冷水机组故障诊断方法  

Fault diagnosis of chillers using sparsely local embedding deep convolutional neural network

作　　者：刘旭婷 李益国[1,2] 孙栓柱[3] 刘西陲[1,2] 沈炯[1,2] LIU Xuting;LI Yiguo;SUN Shuanzhu;LIU Xichui;SHEN Jiong(School of Energy and Environment,Southeast University,Nanjing210096,Jiangsu,China;Key Laboratory of Energy Thermal Conversion and Control of Ministry of Education,Southeast University,Nanjing210096,Jiangsu,China;Jiangsu Frontier Electric Technology Co.Ltd.,Nanjing211102,Jiangsu,China)

机构地区：[1]东南大学能源与环境学院,江苏南京210096 [2]东南大学能源热转换及其过程测控教育部重点实验室,江苏南京210096 [3]江苏方天电力技术有限公司,江苏南京211102

出　　处：《化工学报》2018年第12期5155-5163,共9页CIESC Journal

基　　金：国家自然科学基金项目(51476027)~~

摘　　要：针对于冷水机组提出一种基于稀疏局部嵌入深度卷积网络(sparsely local embedding network,SLENet)的故障诊断方法。采用稀疏局部嵌入方法代替卷积核,对输入数据进行特征选择,避免了复杂的训练和调参过程。另外采用空间金字塔最大池化作为网络的输出层,减少了网络的输出维数和分类器的计算量。针对美国采暖、制冷与空调工程师学会提供的冷水机组的典型故障数据进行分类,结果表明,该方法相比深度卷积网络(CNN)和支持向量机(SVM)方法具有更高的故障诊断精度。A fault diagnosis method for chillers based on a sparsely local embedding deep convolutional neural network(sparsely local embedding network,SLENet)is proposed.The sparsely local embedding filters are adopted for feature selection in the first two layers of SLENet,as a result complicated training and adjusting process are avoided.In addition,the spatial pyramid max pooling layer is employed to construct the output layer of SLENet to reduce the output dimensions of the proposed network,and then alleviate the computationally burden of classification.The experimental data from ASHRAE RP-1043are used to validate the fault diagnosis method.The results demonstrate that the method achieves high diagnostic accuracy compared to convolutional neural network(CNN)and support vector machine(SVM)methods.

关 键 词：算法 神经网络 安全 故障诊断 稀疏局部嵌入 深度卷积网络 空间金字塔最大池化 

分 类 号：TQ277[化学工程—有机化工]