A wavelet residual network-based anomaly detection method for transfer learning under variable operating conditions was proposed to address the insufficient generalization performance of fracturing pump valve anomaly detection caused by dynamically varying operating conditions. A multi-branch residual network incorporating wavelet convolution was constructed to extract discriminative fault features from vibration signals. A multi-branch squeeze-and-excitation attention mechanism was utilized to adaptively recalibrate the channels of key fault features, thereby enhancing feature representation capability. A trainable projected gradient method was adopted to minimize the distribution discrepancy between source and target domains, improving the transfer effectiveness and generalization performance of the model under variable operating conditions. The proposed method was validated using vibration data collected from pump valves of fracturing pumps in a real oilfield environment. The experimental results demonstrated that the proposed method achieved a detection accuracy of 99.51%, outperforming existing state-of-the-art methods by more than 1.73%, and exhibited favorable robustness and generalization capability under variable operating conditions.