Current mechanical oil production system diagnostics suffer from limited single-sensor information, heavy reliance on manual feature design, and difficulties in edge deployment. This paper proposes an intelligent diagnostic method integrating multi-source sensor data with lightweight convolutional neural networks (CNN). By deploying multi-source sensors such as load, current, and vibration, a spatio-temporal alignment algorithm based on hardware clock synchronization and linear interpolation is designed to address inconsistent sampling frequencies of heterogeneous data. A feature-level fusion input structure is constructed based on the pumping unit stroke cycle. Multi-source data are concatenated channel-wise into a two-dimensional matrix as input to a one-dimensional CNN. A lightweight CNN model based on deep separable convolutions and residual connections is further designed, significantly reducing parameters and computational overhead while enhancing edge device deployment feasibility. Experiments using six months of real-world data from 150 oil wells across seven typical operating conditions demonstrate 96.9% accuracy. The model requires only 17.7% of the parameters compared to traditional fusion methods, achieving single-sample inference in just 6.5 ms on Jetson Nano edge devices—meeting field requirements for millisecond-level response. Visual analysis indicates that the features extracted by the model exhibit excellent inter-class separability, validating its effectiveness and practicality under complex operating conditions.
Yaping Wang (Sun,) studied this question.