基于二进离散小波——核降噪自编码的电机振动信号多级降噪

Multi-level Denoise of Electric Machinevibration Signal Based on Discrete Dyadic Wavelet Transform--the Kernel Method-Denoising Autoencoder

在电机故障预测中振动特性信号常被噪音掩盖难以直接使用,因此针对含有噪音的振动信号本文提出一种基于二进离散小波变换(DDWT)--核降噪自编码(KDAE)的局部-全局多级降噪模型,该模型结合小波降噪在局部信号处理领域的优势和深度神经网络对于多维特征全局优化的能力。局部降噪过程选用sym5小波基函数,降噪阀值0.06,降噪后的信噪比到达23.1,信号的样本标准差降低一个数量级;为了弥补二进离散小波降噪后数据的重建损耗波动幅度大的缺点,文章进一步提出了针对全局降噪优化的基于核方法的降噪自编码模型,模型的误差只有0.16%,通过优化后信噪比在上级降噪的基础上又达到了18,同时样本标准差降低为0.0004,并且重构损耗的时序曲线在时间尺度上具有明显的光滑性。本文所构建的降噪模型对含有多维特征的的机电振动信号降噪提供了有效的解决方法。

Features are often very masked by noise in fault prognostics.Therefore,a local-global multi-level denoisemodel based on dyadic discrete wavelet transform(DDWT)-kernel denoising autoencoder(KDAE)is proposed for vibration signals containing noise.This model combines the advantages of wavelet denoise in the field of local signal processing and the ability of depth neural network for global optimization of multi-dimensional features.In the process of local denoise,sym5 motherwaveletis selected,and the threshold value of denoise is 0.06,andthe signal-to-noise ratio reaches 23.1,and the sample-standard-deviation is reduced by an order of magnitude.Due to the lack of wavelet de-noising in the reconstruction of loss timing,this paper further leads to a kernel based de-noising self coding model for global de-noising optimization.The error of the model is 0.16%,the SNR is 18,and the sample-standard-deviation is reduced to 0.0004,and the time series curve of reconstruction loss has obvious smoothness on time scale.The denoise model constructed in this paper provides an effective solution to the denoise of electromechanical vibration signals with multi-dimensional characteristics.