基于声学特征与自编码器的机械故障检测方法

Mechanical Fault Detection Method Based on Acoustic Characteristics and Auto-encoder

为解决工业中检测机械故障只有少量或没有故障数据时,检测难度大,精确度不高的问题,提出一种基于融合声学特征和自编码器的机械设备故障检测方法。首先将自编码器的潜在特征表示设置为8个,优化网络结构。然后使用MSE作为其重构误差函数。最后分别把MFCC和MAF作为输入特征向量。结果表明,论文提出的方法与BS相比,在减少训练次数的同时,平均AUC和平均pAUC均有所提升,能够较好地完成故障检测任务。

When detecting mechanical faults in the industry,only a small amount or no fault data will increase the detection difficulty and reduce the detection accuracy.To solve this problem,a mechanical equipment fault detection method is proposed based on the fusion acoustic features and auto-encoder.First,the potential features of the auto-encoder are set to 8,and its network structure is optimized.Then MSE is used as its reconstruction error function.Finally,MFCC and MAF are used as input features respectively.The results show that compared with BS,the method proposed in this paper can improve the average AUC and the average pAUC while reducing the number of training,which can be completed better in fault detection tasks.