基于CUM3-CNN的柴油机高压油路故障诊断  

Fault diagnosis of diesel engine high pressure oil circuit based on CUM3-CNN

作　　者：常春 梅检民 赵慧敏 沈虹 王双朋 CHANG Chun;MEI Jianmin;ZHAO Huimin;SHEN Hong;WANG Shuangpeng(Army Military Transportation University,Tianjin 300161,China)

机构地区：[1]陆军军事交通学院,天津300161

出　　处：《振动与冲击》2023年第3期174-180,共7页Journal of Vibration and Shock

基　　金：陆军装备部重点项目(LJ20202A050622)。

摘　　要：针对柴油机故障诊断中噪声干扰强、人工确定特征主观影响大、自动识别准确率低的问题,提出了一种利用卷积神经网络(CNN)识别振动信号三阶累积量灰度图的柴油机故障诊断方法。首先利用三阶累积量抑制高斯噪声的先天特点对缸盖振动信号进行分析,生成抑制噪声后的灰度图像,作为卷积神经网络的输入;用具有动量的随机梯度下降优化算法和学习率退火方法训练卷积神经网络,通过遗传算法优化训练参数,用训练好的网络对柴油机高压油路的5种工况进行故障诊断。试验结果表明:三阶累积量生成的灰度图像既能有效抑制噪声又能全面表现特征信息;用学习率退火方法和遗传算法改进优化的卷积神经网络有良好的泛化能力,相比于传统方法具有更高的准确率和抗噪能力。Here, aiming at problems of strong noise interference, large subjective influences of manually determined features and low automatic recognition accuracy in diesel engine fault diagnosis, a diesel engine fault diagnosis method using convolutional neural network(CNN) to identify third-order cumulant grayscale image of vibration signals was proposed. Firstly, cylinder cover vibration signals were analyzed by using innate characteristics of third-order cumulant to suppress Gaussian noise, and gray images were generated after noise suppression and taken as input of CNN. CNN was trained by using the stochastic gradient descent optimization algorithm with momentum and the learning rate annealing method. The genetic algorithm was used to optimize training parameters. The trained network was used to do fault diagnosis under 5 working conditions of high-pressure oil circuit of diesel engine. Test results showed that gray images generated with third-order cumulant can not only suppress noise effectively, but also fully reveal feature information;CNN optimized with learning rate annealing method and genetic algorithm has good generalization ability, higher accuracy and anti-noise ability compared with traditional methods.

关 键 词：柴油机 故障诊断 三阶累积量 卷积神经网络(CNN) 

分 类 号：TK428[动力工程及工程热物理—动力机械及工程]