To address the problems of insufficient feature representation in traditional models and high computational cost in deep networks for working-condition diagnosis of sucker-rod pumping wells, this paper proposes an improved LeNet-based working-condition diagnosis method for the real-time diagnostic requirements of the host computer. The proposed method takes indicator diagrams as input and first performs preprocessing on the samples. Then, based on the traditional LeNet architecture, the Mish activation function, Triplet Attention mechanism, and adaptive global average pooling are introduced to enhance the model’s ability to extract local features among similar working conditions, improve prediction accuracy and the interpretability of diagnostic results, reduce the number of parameters, and optimize the model structure. Experiments are conducted using a dataset collected from a sampled oilfield. The results show that the improved LeNet can identify eight common working conditions, achieving a final prediction accuracy of 97.69% and a single-sample processing time of 0.36 ms. Therefore, the proposed method can meet the requirements of fast and stable working-condition diagnosis on the host computer and has certain engineering application value.