Prediction Model Research For The Finite Element Thermal Structural Analysis Of Disc Brake Based On The Neural Network

The method of the thermal structure coupled simulation of the vented disc brakeis complicated, mainly because that there are too many cases and parameters need tobe validated in the simulation process, and a long computational time will be cost ineach simulation case. So it is almost impossible to comprehensively know theperformance of the disc brake by taking just one or a few times of finite elementsimulations. To solve this problem, disc brake is studied as the research object in thispaper, and a rapid predicting system, which can quickly give the approximate resultsof FEM simulation, is founded based on the combination of the thermal structuralanalysis technology of disc brake, the principal of the orthogonal experimental designand the algorithm of the BP architecture neural network. The main content of thispaper is listed bellow:1) The thermal structural analysis of disc brake is conducted in Abaqussoftware based on the principal of the thermoelasticity, which worked with the initialbraking velocity70Km/h, and the average braking pressure10MPa, and the frictioncoefficient0.3between the disc and pad.2) The major parameters of the thermal structural analysis of disc brake aredetermined based on the research of the braking procedure and the analysis of theFEM result of disc simulation. The simulation experiment with three level for eachfour factors are determined by the orthogonal design.3) The time costed in the finite element simulations are greatly shorten by thePython scripts and parametric batch methods in Abaqus/CAE developmentenvironment.4) The validity of the predicting system are verified by combination of codingthe BP neural network in Matlab and training the BP neural network which extract itstraining samples in Abaqus software with the typical computational cases.The final result prove that artificial neural network can efficiently approximate and predict the finite element result by training the neural network with typical FEMcomputational results which was arranged by the orthogonal design.