Identification Method of Dynamometer Cards Based on Improved MobileNetV2 in Fault Identification of Sucker Rod Pumping System

Existing fault identification methods still face problems such as insufficient lightweight design, inadequate multi-scale feature representation capability, and difficulty in model training under small-sample conditions, making it difficult to meet the requirements of oilfield production sites. For the dynamometer-card-based fault identification task of the sucker rod pumping system, this paper makes targeted improvements to MobileNetV2 and a lightweight fault identification method based on transfer learning and improved MobileNetV2 is proposed. The method uses MobileNetV2 as its basic framework, improves the activation function, and introduces a multi-scale feature extraction module and an attention mechanism, thereby improving the model’s identification capability while maintaining a lightweight design. Experiments were carried out using a constructed dynamometer card dataset based on competition data and measured oilfield data. The proposed method was compared with representative deep learning methods, and a cross-source external validation experiment was further conducted using independently measured oilfield samples. The results show that the improved MobileNetV2 achieved a mean accuracy of 95.01%, with a standard deviation of 0.09%, 2.28 M parameters, a model size of 9.01 MB, 536.72 M MACs, and a CPU inference latency of 29.62 ms per image. In terms of mean accuracy, the improved MobileNetV2 outperforms EfficientNetV2, ResNet34, DenseNet121, MobileNetV3, ShuffleNetV2, and the original MobileNetV2, while achieving a mean accuracy of 90.16% on the cross-source external validation set. This method achieves a relatively good balance between identification performance and model complexity on the dataset used in this study, and it can provide a useful lightweight model reference for dynamometer-card-based fault identification in a sucker rod pumping system.