Sparrow search algorithm-optimized variational mode decomposition-based multiscale convolutional network for cavitation diagnosis of hydro turbines

This article presents an intelligent cavitation diagnosis method for hydro turbines, employing the sparrow search algorithm (SSA), variational mode decomposition (VMD), and deep learning. Initially, hydroacoustic signals from the hydro turbines are decomposed using SSA-VMD, with the key parameters of VMD optimized by SSA. Subsequently, the decomposed signal components are all fed into a multiscale convolutional neural network (CNN) to adaptively extract features and classify states effectively. This method combines the optimization capabilities of SSA with the adaptive signal decomposition characteristic of VMD and the potent classification capacity of CNN, eliminating the need for complex manual feature extraction typically found in existing approaches. The proposed multiscale CNN based on SSA-optimized VMD (SSA-VMD-MSCNN) offers three primary advantages. Firstly, the added SSA-VMD layer legitimately processes nonstationary hydroacoustic data to adaptively obtain components at multiple characteristic scales. It fully capitalizes on VMD, thereby enabling the CNN to extract multiscale features. Secondly, the SSA-VMD layer directly feeds multiscale components into the hierarchical CNN, effectively extracting rich cavitation information while avoiding the loss of valuable information owing to hand-crafted feature extraction. Finally, the number of channels remains constant through zero-padding, addressing the issue of uncertain component quantity caused by adaptive decomposition of the raw signal. Hydroacoustic data captured by hydrophones, provided by a research institute, demonstrate that the SSA-VMD-MSCNN method presented in this study outperforms traditional CNN, wavelet packet decomposition-based multiscale CNN (WPD-MSCNN), and the shallow machine learning method support vector machine (SVM) for cavitation diagnosis.