Load diagrams directly characterize the operational status of pumping wells and serve as a core diagnostic tool for identifying malfunctions. To address the limitations of traditional load diagram recognition—which relies on manual expertise, suffers from low efficiency, and lacks sufficient accuracy—this paper proposes an intelligent load diagram recognition method based on an enhanced MobileNetV3 architecture. This approach replaces the SE module in the original MobileNetV3 network with a CBAM attention mechanism, thereby enhancing feature expression capabilities across both channel and spatial dimensions. Using a real-world indicator diagram dataset from an oilfield, classification experiments were conducted on nine typical operating conditions. Experimental results demonstrate that the improved MobileNetV3_CBAM model outperforms other comparison models in accuracy, precision, recall, and F1 score metrics. Specifically, precision reaches 95.4%, with overall recognition accuracy improving by approximately 1.6% compared to the original model. The research findings demonstrate that the proposed improvement method effectively enhances the accuracy and robustness of load curve recognition, providing reliable technical support for pumping unit fault diagnosis and intelligent oilfield development.