Multi-Instance Multi-Label Semi Supervised Algorithm And Application In The Diagnosis Of Oil Pumping Unit

With the continuous development of the oil extraction technology,the fluid structure in the oil extraction system has become more complicated,which makes it hard to confirm the period and type of the oil pumps’ fault.Faced with the lack of the labeled data,the issue studied the multi-instance multi-label semi supervised method and it’s applied technology by using the numerous condition data which was accumulated in the oilfield’s daily monitoring process,aiming at expressing the data’s integrity and partitioning multiple semantic.The paper came up with the study content and applied technology on the basic of analyzing the current advanced pump condition diagnoses technology and the classic multi-instance multi-label semi supervised algorithm.Countering the problem that using single feature vector to describe the data may bring over simplification and lost of valid information,we built the multi-instance multi-label semi supervised model based on radial basis function neural network,and optimized the method of selecting the model’s nodes by using the Huffman Tree to improve the algorithm’s accuracy.The model is pointed at the problem of timely judgments and expression of the pump’s condition transition,which can output multiple type of condition at the same time to describe the transition between two types,and conduct the confirmation of the model’s parameters by using the clusters’ distance index to make it learning better.Meanwhile,we led into a large number of non-label samples using the semi supervised algorithm in order to assist the analyze of the samples’ space distribution and form the semi supervised clustering multi-instance algorithm and the multi-instance multi-label semi supervised model based on radial basis function neural network.All the works above have improved the original algorithm’s performance.According to the model and algorithm which is faced to the pump’s condition diagnoses,the project achieved the software design and the integration development and got favorable result on the basic of the past monitor data of the pump’s condition.The study has certain practical value and applied vista in some aspects such as the complete description of the condition data,exact expression and judgments of the condition types and so on.