Abstract: To address the susceptibility of acoustic-signature-based fault detection for wind turbine blades to environmental noise and the scarcity of fault samples, an enhanced edge-level dynamic graph convolutional network (E²LDGCN) method is proposed. The method extracts Mel-frequency cepstral coefficients (MFCC) tailored to blade fault frequency bands, constructs the graph structure by combining feature similarity with temporal proximity, and reduces the influence of low-quality acoustic-signature frames on feature propagation through node-confidence weighting, dual-branch similarity modeling, adaptive temperature scaling, and gated topology updating. Experiments on a dataset constructed in this study show that the method achieves a classification accuracy of 96.0%–97.0% and an area under the receiver operating characteristic curve (AUROC) of 0.965–0.972 for four types of blade faults. The system was continuously deployed at an onshore wind farm for six months, during which it performed online acquisition of blade acoustic signatures and real-time fault identification. The method balances blade fault identification accuracy with field operational requirements and can be used for online condition monitoring of wind turbine blades.