Abstract: The rapid diagnosisof incipient faults is an effective approach to preventing and reducing significant faults.Intelligent fault diagnosis methods based on Convolutional Neural Networks (CNN) have become a hotspot in the field of marine machinery.However,the complexity of the current 2D-CNN algorithm is too high and the model parameters are too large,which is not suitable for the rapid diagnosis of micro faults.A novel improved 1D CNN—Global Average Pooling (1DCNN-GAP) algorithm was proposed for fast fault diagnosis,which introduced 1D-CNN to deal with multi-sensor data fusion problem.Then,a 1D (1D-GAP) layer was designed to improve the structure of the fully connected layer and reduce the model parameter quantity and diagnosis time of the existing 1D-CNN.The proposed method was used to diagnose 2-channel vibration sensor fault data from rolling bearing under the 1-horsepower,2-horsepower and 3-horsepower.The diagnostic accuracy was 99.84%,99.51% and 99.33% respectively.The proposed method was compared with the mainstream Support Vector Machine (SVM),K-Nearest Neighbor (KNN),Deep Neural Networks (DNN) and 2DCNN-FC algorithms.The experiment results confirmed that the proposed method had better diagnostic performance and was more suitable for the rapid diagnosis of incipient faults.

Key words: fault diagnosis, 1D convolutional neural network, multi-channel data fusion, marine machinery, rolling bearing, deep learning

摘要： 微小故障的快速诊断是预防和减少重大显著性故障发生的关键。近年来,基于卷积神经网络(CNN)的智能诊断方法已成为船舶机械领域研究的热点。然而,现行的基于图像处理框架的2D-CNN算法在处理多传感器、多通道故障数据时存在检测时间长、数据融合效率低的不足。为此,提出一种改进的1DCNN-GAP的深度学习新算法,用于船舶旋转机械的故障快速诊断。该方法首先引入1D-CNN处理多传感器数据融合问题,然后设计了一维全局均值池化层(1D-GAP)改进全连接层结构,减少传统CNN的模型参数量和诊断时间。通过将提出的方法用于滚动轴承在1马力、2马力和3马力多种负载工况下的2通道振动传感器故障监测数据进行诊断,诊断精确率分别为99.84%、99.51%和99.33%。通过与主流的SVM、KNN、DNN和2DCNN-FC算法进行对比验证,结果表明,所提方法具备更加优越的诊断性能,更适用于多传感器监测环境下微小故障的快速诊断。

关键词: 故障诊断, 一维卷积神经网络, 多通道数据融合, 船舶机械, 滚动轴承, 深度学习