基于U-net和随机森林的齿轮箱振动时频分析和故障诊断

Time-frequency analysis and fault identification of wind turbine gearbox based on U-net and random forest

对于风电机组齿轮箱(WTG)智能故障诊断算法,孤立的故障分类结果常常面临置信度不足的问题。为了在故障分类的同时提供除了诊断结论之外的更多信息,基于振动时频信息,提出了一种用于风电机组齿轮箱振动时频分析和故障诊断的两阶段框架。首先,在第一阶段中,使用U-net模型自动分割了时频图中与故障有关的特征区域,而无需手动设置分割参数;然后,使用基于形状特征的方法提取了被分割图像中的特征信息;最后,在第二阶段中,利用随机森林算法根据形状特征完成对风电齿轮箱的故障诊断任务,并使用采集自华北某风电场的在役风电机组齿轮箱振动数据验证了上述框架的有效性。实验结果表明:分析算法的F 1分数和诊断算法的诊断精度分别达到了0.942和97.4%,U-net方法与现有方法相比具有更高的综合诊断性能和计算效率。研究结果表明:该框架能够精准地标记时频图中的故障特征区域,并快速有效地诊断齿轮箱故障。

For the intelligent fault diagnosis algorithm of wind turbine gearbox(WTG),the isolated fault classification results often faced the question of confidence.In order to provide more information besides diagnosis conclusion while fault classification,based on the vibration time-frequency information,a two-stage framework for vibration time-frequency analysis and fault diagnosis of wind turbine gearbox was proposed.Firstly,at the first stage,the feature areas related to fault in the time frequency image were marked using the U-net model without manually setting the segmentation parameters.Then,the feature extraction method based on region shape feature was utilized to extract valuable information from the analyzed binary images.Finally,at the second stage,using these shape features based on region information,the random forest algorithm was used to complete the automatic fault identification of the WTGs.The proposed two stage pipeline was used to identify the fault conditions of in-service WTGs located in North China to verify its effectiveness.The experimental results show that the F1 score of the analysis algorithm reaches 0.942,while the diagnostic accuracy of the diagnosis algorithm reaches 97.4%.Comparing with existing methods,U-net method shows higher comprehensive diagnosis performance and computational efficiency.The results show that the proposed method can accurately mark the feature patches and feature bands in time frequency images and quickly and effectively diagnose the WTG faults.