随机丢弃和批标准化的深度卷积神经网络柴油机失火故障诊断

Diesel Engine Misfire Diagnosis with Deep Convolutional Neural Network Using Dropout and Batch Normalization

针对已有的柴油机失火故障诊断方法需要精细且耗时的时频特征提取过程,且对实际含有噪声的样本诊断准确率低的问题,提出了一种随机丢弃和批标准化的深度卷积神经网络柴油机失火故障诊断方法。在不同的转速工况下进行柴油机失火故障模拟试验,将采集到的原始缸盖振动信号作为深度卷积神经网络的输入,并在输入端引入随机丢弃抑制输入噪声,通过一维卷积操作自动提取失火故障模式特征,接着在各卷积输出层对特征信号进行批标准化处理,以减少信号在深度卷积层内部的方差偏移,最后基于多分类函数完成失火故障分类。不同噪声环境和方法的对比试验结果表明,所提方法的分类准确率最高可达100%,同时在保证算法准确率的前提下,其鲁棒性优于依赖时频提取特征的方法。

The existing diesel engine misfire fault diagnosis methods require a fine and time-consuming time-frequency feature extraction process, and have low diagnostic accuracy for the actual noise-contained samples. An end-to-end intelligent diagnosis method with deep convolutional neural network using dropout and batch normalization mechanism is proposed. The simulation test of diesel engine misfire fault is conducted at different rotating speeds, and the original cylinder head vibration signals are acquired. The dropout mechanism is introduced to suppress input noise at the input end, and the features of misfire fault are automatically extracted by using one-dimensional convolution. Then the batch characteristic signals of each convolution output are normalized to reduce variance offset. Misfire fault classification is completed based on the multi-classification function. According to the comparative experiment results of different noise environments and evaluation methods, the maximum classification accuracy of the proposed method is up to 100%, and the robustness is better than the methods depending on feature extraction.