A lightweight vision transformer framework integrated with flow visualization for incipient cavitation diagnosis in centrifugal pumps

Abstract

Long-term incipient cavitation in hydraulic machinery poses significant reliability and safety risks, yet its timely detection remains challenging due to imprecise cavitation criterion, weak signals, and noise interference. To address these issues, this study proposes a lightweight and interpretable diagnostic framework based on Vision Transformer (ViT) for early-stage cavitation identification under complex and noisy conditions. Accurate cavitation labels are obtained through high-speed flow visualization, while the synchronously acquired multi-channel vibration signals are subjected to dimensionality reduction and fusion, and subsequently transformed into time–frequency representations to enhance cavitation-related features. A novel Adaptive Convolution (AC) block is developed to extract global and local information from time-frequency images, and combined with Mobile ViT block to construct the lightweight AM ViT model. The proposed model achieves 100 % accuracy across all eight cavitation states in noiseless conditions, with diagnosis time of 15.4 ms. Compared to the pure Transformer ViT-Base model (86M parameters) and the ResNet-18 model (11.7M parameters), the parameter of the proposed model is 5.8M. Even under signal-to-noise ratio (SNR) of −10 dB, the model improves the accuracy by 19.2 % compared to the baseline model. The model trained under flow rate 25 $m^3/h$ generalizes effectively to 20 $m^3/h$ and 30 $m^3/h$ without retraining. Attention map shows that the model focuses on the 3000–5000 Hz band, which contains key cavitation-related features. Comparative experiments confirms that the proposed model consistently outperforms alternatives under various noise conditions. Sensor contribution analysis further indicates that x- and y-axis sensors near the impeller inlet provide the most informative features, offering guidance for optimal sensor layout. The proposed method establishes a correlation between vibration signals and cavitation visualization, offering potential applications in real-time, noise-resistant diagnostic systems.