基于1D卷积与特征融合的深度学习轴承诊断算法研究

Research on Deep Learning Bearing Diagnosis Algorithm Based on 1D Convolution and Feature Fusion

摘要: 轴承损伤严重影响设备正常运行，减小设备寿命。对轴承损伤的有效识别可以帮助设备进行维护。为此，本文创造性地提出了一种基于1D卷积与特征融合的互扰神经网络(Interference neural net-work, IFNN)深度学习模型来实现轴承诊断。该模型由7层大尺寸卷积核单元、传统特征计算单元、融合单元和Softmax分类器组成。为了检验所提出方法的有效性，采用10倍交叉验证、4个数值指标和ROC曲线面积来评估分类结果。实验结果表明，本文提出的IFNN模型准确率为96.09%，精度为96.48%，召回率为96.10%，F1-score为96.08%，ROC值为1。通过与随机森林、支持向量机等模型的对比，IFNN模型的性能明显优于其他模型。因此，所提出的IFNN模型能有效地完成轴承损伤诊断任务。

Abstract: Bearing damage seriously affects the normal operation of the equipment and reduces the life of the equipment. Effective identification of bearing damage can help equipment maintenance. For this reason, this paper creatively proposes a deep learning model of Interference Neural Network (IFNN) based on 1D convolution and feature fusion to realize bearing diagnosis. The model consists of 7-layer large-size convolution kernel unit, traditional feature calculation unit, fusion unit and Soft-max classifier. In order to test the effectiveness of the proposed method, 10-fold cross-validation, 4 numerical indicators and ROC curve area are used to evaluate the classification results. The experimental results show that the accuracy rate of the IFNN model proposed in this paper is 96.09%, the precision is 96.48%, the recall rate is 96.10%, the F1-score is 96.08%, and the ROC value is 1. Through comparison with models such as random forest and support vector machine, the performance of the IFNN model is significantly better than other models. Therefore, the proposed IFNN model can effectively complete the task of bearing damage diagnosis.