The Fault Diagnosis of Rod-Pumped Oil Wells Based on Multi-Branch Time-Frequency Fusion Network

Rod-pumped oil wells are one of the main methods of oil extraction in China, and any failure can lead to significant losses. As the difficulty of extraction increases, traditional condition monitoring methods are becoming less effective in addressing the complexities of the environment. Therefore, achieving accurate diagnosis of well conditions is of great significance for improving oil well production efficiency and reducing the occurrence of accidents. A deep learning-based diagnostic method is proposed, leveraging multi-branch time-frequency fusion. The hybrid architecture synergistically processes temporal and spectral features through parallel TCN-LSTM pathways, where dilated convolutions capture multiscale patterns while gated recurrent units model sequential dependencies, enabling comprehensive spatiotemporal representation learning. Feature fusion is performed using a Cross-Modal Attention mechanism. The model further refines feature representation by optimizing the time-domain and frequency-domain features, enhancing the expressiveness of features through multiple convolutional and residual modules. Empirical validation confirms the developed neural architecture achieves marked improvements in diagnostic precision (94.1%) and operational stability across diverse mechanical failure scenarios.