Abstract: When dealing with low signal-to-noise ratio and non-stationary axial flow pump sound signals, traditional models exhibit limitations in terms of single receptive fields and poor time tracking ability. To address these issues, this paper proposes a fault diagnosis method combining discrete wavelet transform and a multi-scale convolutional temporal fusion network (DWT-MSFANet). First, the model introduces the Discrete Wavelet Transform to preprocess the original acoustic signals, effectively filtering out intense environmental background noise. Second, a parallel convolutional architecture with four scales of convolutional kernels is constructed to simultaneously capture high-frequency, microscopic transient mechanical impacts and low-frequency, macroscopic hydraulic pulsation characteristics within the fault signals. Subsequently, the cross-scale spatial features are fed into a Gated Recurrent Unit to further excavate the long-range temporal dependencies of fault evolution under continuous hydraulic excitation. Finally, a self-attention mechanism is employed to execute a weighted focus on critical fault frequency bands. Case analysis demonstrates that the proposed method consistently outperforms baseline models across all evaluation metrics.

Key words: axial flow pump, multi-scale convolutional neural network, gated recurrent unit, fault diagnosis