Pump Indicator Diagram Intelligent Identification Based On Deep Learning Network

To monitor the working conditions of oil well in real-time can discover the oil pumping unit breakdown timely and avoid invalid producing. And this is significant for reducing the production cost and improving the oil well productivity. It is difficult to check the working conditions of the oil pump unit, because the pump is often installed under the ground in the depths of thousands meters. The main method to identify the working conditions of the oil well currently is to recognize indicator diagrams. And it’s often to measure the indicator diagrams at the suspended point. The artificial recognition method is often used to analyze the well working conditions in the traditional operation. But this method is inefficient and greatly dependent on the judgment of individual, and it can’t realize real-time automatic diagnosis about the oil well working conditions. The oil well indicator diagrams are greatly affected by the working environment, so the accuracy is often low when using traditional machine learning model to identification indicator diagrams. To solve the above problems, it’s put forward a method to recognize indicator diagrams based on deep learning neural network in this dissertation. The main research contents and results are as follows:It’s analyzed the working principle of the rod pump unit and the generation mechanism of pump indicator diagram. The common working condition categories and the corresponding diagram forms were introduced in the dissertation. The generation background and basic ideas of deep learning neural network were expounded and the structures and principles of three different deep learning models were introduced in detail. It had a deep study on the structure principles and algorithms of the DBN and CNN.The sample indicator diagrams were normalized firstly. Then recognition models of the pump indicator diagrams were built based on DBN and CNN separately. The recognition models that based on DBN and CNN were tested and verified in separate. The sample diagram data were divided into two parts: training samples and testing samples.. The training samples were used to train the diagram recognition models and then test samples were used to test the trained models. It’s used the SVM as classifier to experiment based on DBN and CNN. Then the experiment results of the pump recognition methods based on deep learning were compared and analyzed. Finally the deep learning methods in this dissertation were compared with the shallow structure diagram recognition models in other documents. The experimental results proved the deep learning neural network would automatic learn the characteristics of samples in the training process without complicated artificial extracting the characteristics. At the same time, deep learning neural network had a better recognition effect than other traditional shallow models.Finally, the pump indicator identification methods in this article were applied to the intelligence fault diagnosis system of the oil well, using lab windows cvi platform to develop the service software that had the ability to identify indicator diagram automatically. A complete set of intelligence fault diagnosis system was designed successfully at last.