Research Of Prediction Model Of Leakage Of Pressure Relief Valve Based On Leakage Monitoring

Pressure Relief Valve (PRV) is one of the most important and the most frequently-used components in industrial production. Continuous leakage or excessive leakage of PRV makes seal faces susceptible to corrosion and definitely produces a partial or whole loss of function of PRV, which results in safety incidents to block industrial production and menace human being's life and society safety. Due to the severity of the leakage of PRV, ISO19973 (full name: pneumatic fluid power - assessment of component reliability by testing) provides test procedures and failure criteria for assessing the reliability of PRV. So it is of constructive and realistic significance for protecting human being's life and society safety and for maintaining healthy and steady development of industrial production, to build mathematical model of the leakage, to study the nature and mechanism of the leakage and master the changing rules of the leakage. Especially if a new time-serial prediction model is developed, which can predict the leakage quickly and precisely, we make the decision timely whether to maintain, and aim at a low cost, high efficiency maintenance system.The process of the leakage of PRV, with the characters of nonlinear, multi-variable, time-varying and highly coupling, along with large quantities of uncertain factors, is a typical complex process. Leakage modeling by mechanism can discover the nature and mechanism of the leakage, and reveal the relationship between the leakage and other physical quantities. However, it is difficult to obtain better prediction result. Support vector machines (SVM), a new artificial intelligence algorithm in prediction modeling, due to its perfect theory basis, has succeeded in engineering applications, and has attracted the worldwide extensive attention. Therefore, SVM launches a possible way for the leakage prediction. On the basis of SVM, this thesis aims at better precision and speed in leakage modeling, and a series of research and work is carried out. The main contents of this thesis are as follows.(1) A mechanism prediction model for leakage of PRV. The total leakage consists of two parts, drainage through leakage hole and leakage through seal faces. For the first part, regard leakage hole as convergent nozzle and drainage through leakage hole can be got according to equation of continuity. For the second part, using fractal theory describes surface topography of seal faces varying with time and the mathematical expression of leakage channel in seal faces is presented. Leakage through seal faces can be calculated by also regarding leakage channel as convergent nozzle.(2) Weight adjusting method based on SVM and hidden information. After long term working, the geometry of both surface topography and seal gap of seal faces changed greatly. So the mechanism leakage model of PRV is difficult to solve. In the domain of SVM, using variable weight describes the case. Since samples with different time contribute differently, each of them is allotted a different weighing coefficient. The concept of hidden penalty coefficient is proposed. Map hidden information and samples into a same high-dimension space. In this space by standard SVR processing hidden information, the bias of samples acted by hidden information is used as adjusting function of weighting function. Simulation verification on regression of Mackey-Glass chaotic time series and experiment verification on regression of leakage of mechanical-seal device are carried out.(3) Time prediction model for leakage of PRV based on double hidden information. For drawbacks of the mechanism prediction models and modern prediction/analysis methods, as well as the characteristic of leakage of PRV, prediction model for leakage of PRV is presented based on SVR and double hidden information, on the basis of support vector regression plus and weight adjusting method. The method and step of building this model is given, including the choice of hidden information, pretreatment of leakage and pressure and usage of the model. Against the background of multi-parameter optimization, the performance difference between Genetic Algorithm (GA) and Particle Swarm optimization (PSO) is discussed. When optimizing 2 or 3 unknown parameters, there is no conclusion verifying which is better, while optimizing 6 unknown parameters, PSO is superior to GA. The experiment of prediction for the leakage failure of PRV is performed, and compared with other prediction models, the proposed model predict the most number of failure samples exactly. Analyze No.3 PRV in detail and explain in theory why the hidden information of No.3 PRV is ineffective.(4) Sequential minimum optimization plus algorithm (SMO+). Support vector regression plus makes full use of the hidden information, and further improves the prediction accuracy. The introduction of the hidden information increases the calculation efforts such that the training time of the prediction model gets longer. SMO+ optimizing support vector machines plus is presented. Based on KKT conditions and by analyzing constraint conditions of support vector machines plus, the standard of judging support vectors or common samples is provided. Violation of KKT conditions of SMO+ along with some situations satisfying KKT conditions is also provided. By performing proper mathematical transformation, it is proved that support vector machines plus and support vector regression plus have the same mathematical expressions, such that SMO+ can optimize both support vector machines plus and support vector regression plus. Prove and explain support vector machines plus with the form of least square does not exist. SMO+ software is developed and used to do prediction experiment on leakage of PRV.