基于多任务学习的轨道车辆轴承异常检测方法

Abnormal detection method of rail vehicle bearing based on multi task learning

轨道车辆轴承温度现有异常检测方法的阈值判别指标受到路况、环境等多项外因干扰,并且基于异常检测目的的预测方法需要对轴承进行逐一建模,模型训练耗时且多模型维护困难。针对上述问题,提出一种基于多任务学习的轨道车辆轴承异常检测方法。首先考虑到数据分布在正常与异常时存在差异,把正常工况下的关联轴承温度作为模型输入构建轴承温度预测模型,当实际温度异常时预测值与实际值关联性呈现异常变化,因此该模型具有异常检测功能。其次,考虑到循环神经网络建模时进行递归运算消耗大量时间,引入多头自注意力机制,所构建的模型能够同时对一轴上的轴箱、齿轮箱、电机3类共9个轴承温度进行同时检测。最后,采用极大似然估计方法,将点预测转换为置信区间预测,解释了预测结果的意义。在正常和故障数据上分别对模型进行验证,结果证明本文所提方法具有准确的9个轴承和异常检测能力,并与单任务模型相比能大幅度减少建模时间。

The threshold discrimination index of the existing abnormal detection methods of rail vehicle bearing temperature was disturbed by many external factors such as road conditions and environment.The prediction method based on the purpose of abnormal detection needs to model the bearings one by one,which is timeconsuming for model training and difficult to maintain multiple models.Due to the above problems,this paper proposed a method for abnormal detection of rail vehicle bearings based on multi-task learning.First,considering the difference between normal and abnormal data distribution,the normal data related to bearing temperature is used as the model input to construct a bearing temperature prediction model.When the actual temperature is abnormal,the correlation between the predicted value and the actual value shows abnormal changes,so the model has anomalies detection function.Secondly,considering that recursive operation consumes a lot of time in the modeling of typical recurrent neural network,the multi-head self-attention mechanism was introduced.The model can simultaneously detect the temperature of nine bearings of axle box,gearbox and motor on one shaft at the same time.Finally,the maximum likelihood estimation method was used to convert the point prediction into the confidence interval prediction,which explains the significance of the prediction results.Themodel was verified on normal and fault data.The results show that the method proposed in this paper has the ability to accurately detect the abnormality of nine bearings at the same time,and can greatly reduce the modeling time compared with the single-task model.