Cavitation state recognition via optimized decomposition and logarithmic-scale multiscale convolutional neural network

Blade cavitation is a common and severe phenomenon in hydroturbine operation. Its evolution produces strongly non-stationary hydroacoustic signals with spectral overlap and blurred boundaries, which hinder accurate state recognition. To address the difficulty of separating multi-frequency components and the limited feature-extraction capability of conventional methods for complex cavitation signals, this study proposes a cavitation state recognition method that integrates the Crayfish Optimization Algorithm (COA), Variational Mode Decomposition (VMD), and a Multiscale Convolutional Neural Network (MSCNN), termed COA–VMD–MSCNN. In this method, COA adaptively optimizes VMD’s key decomposition parameters, enabling more robust extraction of narrow-band modal components with higher frequency separation. These multiscale components are then fed into the MSCNN as input features. At its front end, a logarithmically designed multiscale convolution module effectively captures characteristic features ranging from long-period modulations to rapid local disturbances, thereby enabling accurate recognition of cavitation states. Experiments on hydroturbine hydroacoustic data demonstrate that COA–VMD–MSCNN effectively identifies four typical states: non-cavitation, incipient cavitation, developing cavitation, and severe cavitation. With an overall accuracy of 98.58% and a macro-averaged F1 score of 0.9841, the method outperforms traditional machine-learning approaches, baseline deep models, and other optimized variants. It also shows superior reliability in critical states, such as incipient cavitation, by reducing misclassification, confirming robustness and adaptability for complex non-stationary hydroacoustic signals.