| 引用本文: | 赵书涛,王二旭,陈秀新,王科登,李小双.声振信号联合1D-CNN的大型电机故障诊断方法[J].哈尔滨工业大学学报,2020,52(9):116.DOI:10.11918/201901221 |
| ZHAO Shutao,WANG Erxu,CHEN Xiuxin,WANG Kedeng,LI Xiaoshuang.Fault diagnosis method for large motor based on sound-vibration signal combined with 1D-CNN[J].Journal of Harbin Institute of Technology,2020,52(9):116.DOI:10.11918/201901221 |
|
| 摘要: |
| 针对复杂运行环境下大功率电动机故障诊断准确率不高、算法泛化能力差的问题,提出一种声振信号联合一维卷积神经网络（1D-CNN）故障诊断方法. 首先对采集到的声信号采用背景噪声库联合稀疏表示去除噪声,然后将声音信号进行带通滤波（7~20 kHz）,叠加低频振动信号（7 kHz内）形成频带更完整的电动机状态表征信息. 再对经过滤波提纯处理后的信息进行重叠式数据扩容,获取1D-CNN训练所需大量数据. 最后将数据样本输入1D-CNN进行学习训练,采用局部均值归一化（local response normalization,LRN）和核函数去相关性改进1D-CNN模型结构,降低抽油机正负半周工况波动对电动机诊断准确性的影响. 诊断结果表明:声振信号联合分析的卷积神经网络故障诊断总体诊断准确率达到了97.75%,泛化能力好,与传统的电动机故障诊断方法相比优势明显. |
| 关键词: 电动机 声振联合 1D-CNN 稀疏表示 数据扩容 故障诊断 泛化能力 |
| DOI：10.11918/201901221 |
| 分类号:TM32 |
| 文献标识码:A |
| 基金项目: |
|
| Fault diagnosis method for large motor based on sound-vibration signal combined with 1D-CNN |
|
ZHAO Shutao1,WANG Erxu1,CHEN Xiuxin2,WANG Kedeng1,LI Xiaoshuang1
|
|
(1. School of Electrical and Electronic Engineering, North China Electric Power University, Baoding 071003,Hebei, China; 2. School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, Hebei, China)
|
| Abstract: |
| To deal with the problems of low accuracy and poor generalization ability of high-power motor fault diagnosis in complex operation environment, a fault diagnosis method based on sound-vibration signal combined with one-dimensional convolutional neural network (1D-CNN) was proposed. First, a collected sound signal was denoised by the combination of background noise database and sparse representation. Next, the sound signal was filtered by band-pass filter (7-20 kHz), and low frequency vibration signal (within 7 kHz) was superimposed to form more complete motor state representation information in frequency band. Then, the information after filtering and purification was expanded by overlapping data to obtain a large amount of data required for 1D-CNN training. Finally, data samples were input into 1D-CNN for learning and training. Local response normalization (LRN) and kernel function decorrelation were used to improve the structure of 1D-CNN model, which reduced the impact of positive and negative half-cycle fluctuations of pumping unit on motor diagnostic accuracy. Diagnostic results show that the overall diagnostic accuracy of CNN fault diagnosis reached 97.75% based on combined sound and vibration signal analysis, and the generalization ability was good. Compared with traditional motor fault diagnosis methods, the proposed method had obvious advantages. |
| Key words: motor sound-vibration joint 1D-CNN sparse representation data expansion fault diagnosis generalization ability |
