Combining support vector machines with neural networks for fault diagnosis in oil and gas measurement and control systems

Oil and gas measurement and control systems operate under high-temperature, high-pressure, high-noise and dynamic conditions, making them highly susceptible to various failures such as tubing string leaks, pump jamming, sand production and insufficient fluid supply. Traditional methods based on mechanism modeling and shallow learning struggle to achieve efficient and stable fault diagnosis under conditions of strong noise and small sample sizes. To address this paper proposes a hybrid fault diagnosis method (TsRNet-SVM) combining Two-stream Residual Convolutional Neural Networks (Two-stream ResNet, TsRNet) with Support Vector Machines (SVM). This approach utilizes a two-stream residual architecture to concurrently extract multi-scale features, employs Principal Component Analysis (PCA) for feature dimensionality reduction and finally replaces traditional Softmax classifier with SVM to achieve multi-class fault discrimination. Simulation and field data validation demonstrate that method maintains high-precision fault recognition even under low signal-to-noise ratio conditions. Compared to baseline methods such as CNN, LSTM and CNN-LSTM, it exhibits significant advantages in accuracy, robustness and convergence speed. In practical application at the Western South China Sea oilfield, early warning system successfully detected 354 well-specific faults in advance with 325 valid alerts yielding a success rate of 91.18%. This reduced the average response time for corrective measures by 8.4 days, significantly minimizing production losses.