End-to-end multi-scale residual network with parallel attention mechanism for fault diagnosis under noise and small samples

When the fault diagnosis datasets contains noise disturbances, small samples, compound faults, and mixed conditions, the feature extraction capability of the neural network will face significant challenges. This paper proposes an end-to-end multi-scale residual network with parallel attention mechanism to address the above complex problems. Firstly, the adaptive mixing pooling method is employed to facilitate the model’s ability to retain effective feature information present within the timing signal. Then, we propose parallel attention mechanism that can obtain the attention information in both channel and temporal domain of the input features. Moreover, the multi-scale feature parallel fusion can better capture effective information contained in different scale features. The experimental results demonstrate that the proposed model attains $99.67$ $\text{\%}$, $99.83$ $\text{\%}$, $99.71$ $\text{\%}$ and $99.70$ $\text{\%}$ accuracy on four datasets comprising small samples. Furthermore, the accuracy of $60$ $\text{\%}$ to $80$ $\text{\%}$ is sustained when the noise level is increased to 0dB.