Abnormal detection method of hydro-turbine generator based on vision transformer

Abstract

At present, the stator and rotor components of hydro-turbine generators operate under complex electromagnetic and mechanical stress environments for extended periods, making them prone to micro-cracks, insulation aging, local overheating, or structural wear. Traditional detection methods are not accurate enough in identifying such minor faults. To improve the accuracy and stability of anomaly detection, this paper proposed a stator and rotor surface anomaly recognition method based on Vision Transformer (ViT), which used its self-attention (SA) mechanism to extract long-distance dependencies in the image and capture global and fine-grained features. In this method, the image is first divided into fixed-size patches, and a sequence input Transformer encoder is constructed after linear projection and position encoding. The multi-head SA network and feedforward neural network are combined for feature expression and classification training. The cross-entropy loss function is used for error feedback. The optimizer selects Adam, and the Cosine annealing strategy is introduced to adjust the learning rate to improve the convergence efficiency. At the same time, the generalization ability of the model is improved by L2 regularization. After the SMOTE (Synthetic Minority Oversampling Technique) algorithm is expanded, the dataset is divided into training set, validation set and test set. The evaluation is carried out on 10,462 data including five types of fault images, including cracks, corrosion, material fatigue, deformation and contamination. The experimental results show that the accuracy of the ViT model in crack and corrosion detection reaches 95% and 96%, respectively. When the Patch size is reduced to 8 × 8, the prediction accuracy is still 96.0%, the F1 value is 94.8%, and the AUC (Area Under the Curve) is 0.96, reflecting the ability to identify minor damage. In multiple stability tests, the accuracy rate under different data sets is always maintained above 97%, up to 97.99%, with minimal volatility. Unlike existing ViT-based approaches used in general fault diagnosis, the method in this paper introduces a specialized patch-scale optimization strategy tailored for the inner surfaces of hydro-generator stator and rotor regions, which exhibit highly irregular fault morphology and weak damage gradients. The model configuration integrates patch resolution tuning, cosine learning rate scheduling, and L2-based generalization reinforcement in a unified pipeline explicitly aligned with industrial inspection constraints. This method has significant application value in improving the accuracy, stability and intelligent maintenance level of micro-damage recognition in the stator and rotor regions of hydro-turbine generators, and provides a transferable deep learning solution for intelligent visual inspection in industrial scenarios.