Research On Inversion And Recognition Of Dynamometer Card Based On Neural Network

The sucker rod pumping is dominant in crude oil extraction at home and abroad,however, it brings high failure rate, due to limitation of its mechanical structure and effectof complex surroundings, then pumping automatic fault diagnosis system usingdynamometer card is an effective measures to prevent, detect and solve the fault. Combinedwith engineering project, this paper studies on the key of fault diagnosis-inversion andrecognition of dynamometer card. The details are follows:Firstly, based on high-dimensional nonlinear relationship between electrical parametersand load, third-layer BP neural network is adopted to build an inversion of network modelby electrical parameters as input and load as output. Network is trained by measuredelectrical parameters data and load data to achieve n-dimensional nonlinear mapping tom-dimensional space, getting inversion dynamometer card.Secondly, based on geometric characteristics of typical dynamometer card,low-frequency Fourier Descriptors are extracted, adding load variation amount in up anddown stroke, to constitute the comprehensive characteristic parameters, which is standingfor corresponding work condition. RBF neural network is adopted to build a recognition ofnetwork model by comprehensive characteristic parameters as input and type of workingcondition as output. Network is trained by sample set of13kinds of typical dynamometercard of characteristic parameters, to establish a RBF neural network, which can correctlyrecognize the typical dynamometer card.Third, particle swarm optimization model is designed to solve the short of slowconvergence and easy to fall into local minimum value in BP neural network, and theproblem of difficult to determine the Radial basis function data center in RBF neuralnetwork. In order to improve the standard PSO’s defect, premature and slow convergencein later stage, proposing a new parameter setting method, while introducing the variablethoughts of genetic algorithm. Then, Improved PSO applied to optimize initial weight andthreshold of BP neural network to establish PSO-BP neural network. Improved PSO applied to optimize radial basis function data center location, width and weight of thehidden layer to the output layer of RBF neural network to establish PSO-RBF neuralnetwork. Using to inverse and recognize dynamometer card.Lastly, completed simulation work of inversion of dynamometer card by PSO-BP neuralnetwork and recognition of dynamometer card by PSO-RBF neural network base onMatlab simulation platform. Simulation result show that, more than99%match betweendynamometer cards drew by PSO-BP neural network and dynamometer cards drew bymeasured data. The correct recognition rate reach100%and83%, when PSO-RBF used torecognize20samples dynamometer card and6random dynamometer card. Wherefore, thealgorithm of inversion and recognition of dynamometer card base on neural networkproposed by this paper is high reliability, moderate amount of calculation, and easilyimplemented on embedded hardware platforms, which is significant for sucker rodpumping real-time working condition monitoring technology research, timely and accuratetroubleshooting, ensured safe production, and improves the field’s economic benefits.