Study On Diagnostic Method For Down-hole Faults Of The Beam Pumping Unit Based On Dynamometer Card

The beam pumping unit is the most common form of artificial lift in oilfield of the world, which is belonged to sucker rod pumping. As the pump working thousands of meters underground, it has complex down-hole working conditions, extremely harsh environment and high failure rate. The dynamometer card can truly reflect the down-hole working conditions, which is commonly used in practical production. At present, it is mainly analyzed by the manual work whose analysis conclusions are used to make a decision whether well-off or well-repaired or adjusting production parameters. However, it is difficult to realize real-time monitoring and make the pumping well work in stable, more efficient and continuous production by this manual form. Therefore, it has very important significance to improve production efficiency of the pumping wells through real-time monitoring and analysis of down-hole working conditions and then developing timely measures.With the development of computer communication and artificial intelligence technology and the demands of high automation for pumping well production, it makes sense to use computer to replace the manual work to monitor the down-hole working conditions in time. Safety, stable and efficient production operation of the pumping wells can be guaranteed by timely adjustments of operation parameters and early warning of failures according to real-time down-hole working conditions. Based on the research background of computer diagnosis for down-hole faults of the beam pumping unit, this dissertation aims to develop advanced methods to improve its comprehensiveness and accuracy. Specially, we have deeply study on feature extraction method, fault location method and uncertain problem caused by irregular dynamometer cards. Our main contributions can be summarized as follows:The more refinement method of the feature extraction is used to deal with the diversity and uncertainty of the dynamometer card. According to the four-point method commonly used in oil production, the graph is divided into four parts which reflects different working states of the pumping unit. The curve moment theory is used to extract 7 moment features of each part to enlarge local features of the graph.28 feature vectors by partition calculation can describe the details of the pump dynamometer card accurately, which can better reflect differences among various graphs. SVM is acted as classifier which is good at small training samples, and at the same time, PSO algorithm is used to choose the best error penalty parameter C and kernel function parameter g.Diagnosis of the dynamometer card by human is usually a process of qualitative analysis, but diagnosis by computer is a process of quantitative analysis. However, uncertain relationships between fault types and graph features are produced by irregularities of the dynamometer card, which cannot be fully explained with simple "Yes" or "No". It is considered that the objective diagnostic conclusions will be acquired only using the method by the combination of qualitative analysis and quantitative analysis. To solve this problem, a method based on extension element theory is proposed in this paper. From angles of qualitative and quantitative,28 feature vectors are used to construct matter-element model of each fault type. Eigenvalues of all feature vectors is expressed as an interval (determined by training samples), which are used to obtain the occurrence degree of the diagnosed sample relative to different fault types through calculation of correlation degree.Multiple faults diagnosis method based on Freeman chain code and DCA is proposed in this paper. Freeman chain code is used to calculate changes in curvature of the data points to visually describe the trend of the curve. By analyzing the dynamometer cards of typical fault types, it can be found that they all have some obvious features. The occurrence of fault can be explained that these obvious features become abnormal. So, some typical characteristics are determined according to typical fault dynamometer cards. Then the occurrence degree of them are calculated and conducted as feature vectors. According to theoretical knowledge and production experience, designed mode set of typical fault types is established. Control chart is used to do abnormal detection of which the upper and lower control lines are determined by the samples under normal working conditions. The observed data is projected to each designed mode to be detected whether the abnormal conditions occur according to the upper and lower control lines of each designed component.Improved fuzzy ISODATA dynamic clustering algorithm is proposed in this paper to solve the problem that supervised learning method relies on training samples. The data set use a self-learning manner to find the relationship between the global distribution patterns and the data attribute to realize automatic clustering of each fault type. Hsim similarity function is used to replace Euclidean distance to calculate the distance between the data as Euclidean distance cannot well measure the distance between high-dimension data. For the determination of best clustering number, the traditional exhaustive search method has enormous calculation and cannot guarantee the global optimal solution. So, the "merging" and "splitting" mechanism is introduced in this paper to dynamically correct the clustering number in classification process. According to the relationship among "the smallest distance between two classes (Md)" parameter, the clustering number (c), validity index (XB) and accuracy rate, SA algorithm is used to search the best Md and XB to obtain the satisfied clustering result.To improve the accuracy and comprehensiveness of the diagnostic conclusions, multi-experts integrated diagnostic method is used to fuse different conclusions given by different methods in order to avoid misdiagnose by single diagnosis method. One important problem is the fusion of conflict conclusions. To solve this problem, an evidences combination method based on weight optimization is proposed in this paper. A calculation method of support degree is used to describe different evidences'credibility. Although it can increase the focusing ability of the combination results, high conflict degree exists. Then Ambiguity Measure (AM) is used to measure evidence's uncertainty, which is inhibited by negative exponential function in order to make evidence distinct. The evidence's weight is calculated by its credibility and uncertainty. The minimum conflict degree is regarded as optimized target and PSO algorithm is used to select reasonable dimension p of support degree and inhibitory factor a of AM. The constraint condition is given to avoid the case that credibility of evidence is smaller and at the same time AM of it is also smaller, which causes the condition that the evidence with lower credibility has bigger weight. An example is given to show that the proposed multi-experts integrated diagnostic method is reliable.