When Lightweight Block Meet Attention: A Lightweight Deep CNN for Fault Diagnosis in Beam Pumping Units

With the acceleration of industrialization, efficient fault diagnosis of key equipment has become an important task to ensure production stability. Beam Pumping Units (BPUs), as the core equipment in the petroleum industry, often suffer from frequent failures due to their complex structure and poor working conditions, which brings great challenges to production efficiency and safety. Traditional fault diagnosis methods rely on manual feature extraction and basic machine learning algorithms, which have the limitations of low efficiency and poor robustness. To this end, this paper proposes a Lightweight Block (LWB) and Lightweight Attention (LA) fault diagnosis method, which realizes efficient feature extraction and classification by combining depth-separable convolution, dynamic channel weighting, and residual linkage. Experimental results show that the model in this paper significantly outperforms the classical models such as ResNet50, DenseNet121, MobileNet-v2 and ShuffleNet-v2 in terms of accuracy (96.63%), recall (91.35%) and computational complexity (0.19 GMac). At the same time, the number of model parameters is only 4.02 M, and the model size is 15.6 MB, which provides high lightweight advantage and deployment flexibility.