A novel wind turbine fault diagnosis method based on improved TFMST and DSC-CNN-GRU model

In wind turbines, rotating components serve as critical parts and are also prone to failures. The fault signals of wind turbines represent typical non-stationary and nonlinear signals susceptible to noise interference. Existing time-frequency analysis methods exhibit insufficient energy concentration when extracting time-varying non-stationary fault features, making feature extraction from signals more challenging. The primary drawbacks of single-data fault diagnosis methods lie in their limited information scope, poor robustness, lack of redundancy and fault tolerance, and difficulty in handling complex or multi-dimensional fault patterns. To address these issues, this paper proposed a model based on Improved TFMST and DSC-CNN-GRU. Firstly, the original Time-Frequency-Multisqueezing Transform (TFMST) technique was enhanced by optimizing its window function, introducing multi-scale adaptive thresholding to improve robustness, and relaxing the curvature criterion to enhance feature sensitivity. Furthermore, eps protection was incorporated throughout the algorithm to ensure numerical stability. Secondly, two datasets were constructed: one comprising two-dimensional data derived from the improved TFMST and the other containing one-dimensional raw data. Subsequently, a dual-input DSC-CNN-GRU model was developed, and both datasets were fed into it. Notably, the proposed model adopts a lightweight design. Finally, information from both data branches is fused and delivered to the classifier for the fault diagnosis task. To demonstrate the effectiveness of the proposed method, comparisons with other relevant methods were conducted on various datasets, indicating that the proposed method achieved desirable fault diagnosis accuracy.