Abnormal operation condition diagnosis of block station based on deep residual network

Oil and gas gathering station is the core part of oilfield ground engineering construction. As the important link of gathering and transportation system, the block station has the production characteristics of centralized equipment and successional production chain, and it is also prone to severe fluctuation of the inflow proportion and equipment operation faults. The diagnosis of the operation condition for block station is crucial to the oil and gas production system, for the abnormal data of simple equipment, the station staff can make a preliminary diagnosis, but for a large number of real-time SCADA monitoring data of the whole station, it is difficult to realize rapid analysis and processing only by experience and knowledge. Compared with the existing threshold alarm method in oil field, data-driven diagnostic approach is more accurate and intelligent. Among the data-driven methods, deep learning method which is good at processing massive high-dimensional data, can automatically extract the nonlinear features of data. Aimed at multiple time series characteristics of data (SCADA) in block station, a fault diagnosis method is proposed by use deep residual network (DRN). In order to identify and classify the abnormal working conditions of block station, a diagnostic model was established by taking 36 monitoring variables of the SCADA system in block station as model input and 5 working conditions as model out. The noise of field data will reduce the ability of the model to identify the working conditions with fewer samples, wavelet decomposition is used to de-noise the data of the block station to reduce the interference of equipment acquisition, enhance model diagnostic performance. Naive resampling is used to enlarge the data capacity to alleviate the difficulty in training the model caused by insufficient sample size of field data. The regularization method is used to punish the weight vector with large values to avoid the dependence of the model on individual variables. On this basis, eight different DRN architectures has proposed to select the optimal diagnostic model for the block station, and the correction between various samples is quantified according to the mutual information method, ensured the validity of the diagnosis results. Verification of real data in field shows that the method can be used quickly and accurately diagnose process status of block station. The average accuracy is 97.3%, which are significantly higher than other machine learning method like support vector machine (93%) and multilayer perceptron (65%). The method has guiding significance for fault diagnosis and anomaly identification of other oil and gas stations.