To address the challenges of insufficient feature extraction, inadequate feature fusion mechanisms, and limited diagnostic capability under complex operating conditions in rotating machinery fault diagnosis, a novel FFT-CNN-Informer fault diagnosis method based on dual-domain feature fusion is proposed. A parallel time-frequency domain feature extraction architecture is constructed. Specifically, Fast Fourier Transform （FFT） is employed to extract frequency-domain features, a multi-scale Convolutional Neural Network （CNN） is designed to capture local temporal features, and a Probabilistic Sparse Self-Attention mechanism is introduced to capture long-range dependencies, enabling multi-dimensional feature extraction of vibration signals. Furthermore, residual learning and an adaptive feature fusion mechanism are incorporated to enhance the ability to extract fault features under varying operating conditions. Experimental results on the Case Western Reserve University bearing dataset demonstrate the superiority of the proposed method. On wind turbine bearing datasets, the proposed method outperforms baseline Transformer and LSTM models by 6.73% and 9.77% in diagnostic accuracy, respectively. Ablation studies further validate the necessity of the dual-domain feature extraction architecture and the adaptive feature fusion mechanism, confirming the effectiveness and robustness of the method in practical wind farm applications.

wind turbines  /  fault diagnosis  /  deep learning  /  frequency domain analysis fusion  /  bearing  /  data fusion