基于小波散射深度序列神经网络的制动噪声分类识别

Braking Noise Classification Based on Wavelet Scattering Deep Sequential Neural Network

摘要:

为实现对制动噪声的智能化识别，研究了一种小波散射结合深度序列神经网络的识别方法。采用3层小波散射变换构造出制动噪声相应卡钳振动信号的小波散射多维特征向量。首先，以单层一维卷积神经网络（1DCNN）和单层双向长短时记忆网络（BiLSTM）为基础，将小波散射特征以序列形式和分别输入方式进行训练和测试；结果显示，与短时能量和短时平均过零率这类一维序列输入相比，小波散射变换多维特征输入能够大幅提高分类准确率。其次，针对网络欠拟合状况，建立的4层深度1DCNN与3层深度BiLSTM网络相比，其基础网络具有更强的特征捕捉能力，均进一步提高了制动噪声分类准确率。根据分类性能指标F₁，4层1DCNN的整体性能均超过3层BiLSTM网络，并且具有训练参数数量较少的优越性。

Abstract:

A data driven classification method based on wavelet scattering collaborative deep sequential neural network is studied to achieve intelligent braking noise recognition. Wavelet scattering transform is used to extract the features of caliper vibration signals relating to braking noise， which form the multidimensional wavelet scattering feature vectors. These feature vectors are input respectively to a standard one layer 1D convolutional neural network （1DCNN） and a one layer bi-directional long short-term memory （BiLSTM） network. The testing results show that the accuracy for braking noise classification can be improved significantly by both networks input with multidimension features from wavelet scattering transform compared with one-dimension features from Shot-time Energy and Shot-time mean zero crossing rate. Modified 4-layer deep 1DCNN and 3-layer deep BiLSTM network are further proposed to reduce training underfitting from standard one layer networks to enhance the capability of feature capture and further improved the classification accuracy. According to F₁ indicator the 4-layer deep 1DCNN shows a better overall performance than that of 3-layer deep BiLSTM network， which also has advantages of fewer training parameters.