A Combination of Dilated Self-Attention Capsule Networks and Bidirectional Long- and Short-Term Memory Networks for Vibration Signal Denoising

Abstract

As scalar neurons of traditional neural networks promote dimension reduction caused by pooling, it is a difficult task to extract the high-dimensional spatial features and long-term correlation of pure signals from the noisy vibration signal. To address the above issues, a vibration signal denoising method based on the combination of a dilated self-attention capsule network and bidirectional long short memory network (DACapsNet–BiLSTM) is proposed to extract high-dimensional spatial features and learn long-term correlations between two adjacent time steps. An improved self-attention module with spatial feature extraction ability was constructed based on the random distribution of noise, which is embedded into the capsule network for the extracted spatial features and denoising. The dilated convolution is integrated into the improved capsule network to expand the receptive field to obtain the spatial features of the vibration signal. The output of the capsule network was used as the input of the bidirectional long-term and short-term memory network to obtain the timing characteristics of the vibration signal. Numerical experiments demonstrated that DACapsNet–BiLSTM performs better than other signal denoising methods, in terms of signal-to-noise ratio, mean square error, and mean absolute error metrics.

A receptive field transfer strategy via layer-aligned distillation learning for fault signal denoising

“…In addition, such methods can be adapted and optimized to the particular circumstances of the datasets. However, such methods are difficult to identify complex high and low frequency mixed noise, and are susceptible to model selection and parameter settings, thus requiring a certain level of expertise and experimental experience [21,22].…”

A multi-scale spatial–temporal capsule network based on sequence encoding for bearing fault diagnosis

“…Wang et al [25] presented a time CapsNet encoder-assisted classifier method, which can jointly optimize subspace learning and fault detection. Qin et al [26] proposed a CapsNet model with an LSTM mechanism to improve remaining life estimation by capturing temporal correlations in time series data. Wang et al [27] introduced the Bidirectional-LSTM into the CapsNet, enabling the network to handle the periodicity of vibration time series vibration signals.…”