Research On Machine Learning-Based Simulation Deviation Revising Method Of MEMS Cantilever Sensors

Serving as a media to connect intelligent software technology and external information,sensors play a significant part in realizing comprehensive social intellectualization and informatization.Distinguished for their high performance and miniaturization capability,MEMS cantilever-based sensors are widely used.With the rapid development of social intellectualization and informatization,MEMS cantileverbased sensors are facing unprecedented demand for higher accuracy and faster upgrading speed.CAE approach have been widely used in various engineering industries and have accelerated sensor's design and optimization.But physical sensors tend to be accompanied with multi-physics interaction,making it hard to perform 3D simulation of the sensors.To ensure the accuracy,3D simulation tend to consume a large amount of time and computing expenses.To solve the problem,a machine learning algorithm-based deviation decreasing methodology is promoted,aiming at decreasing simulation time as well as keeping the accuracy.Taking MEMS cantilever current sensor as an example,the theory of simplifying 3D model to 2D one is established.A magnetic field-solid structure-electric field fully coupled FEM simulation approach is proposed and verified.A series of 1767 groups of 2D and 3D FEM simulations are performed and the deviation between 2D and 3D results are analyzed.The results suggest that FEM simulation deviation mainly comes from solid structure domain,and 2D simulation can reduce 85.7% of simulation time,cutting 65.9% ram expenses while introducing a deviation up to 92%.A deviation decreasing methodology is promoted based on back propagation neural network,which is trained by 1600 groups of FEM results and tested using the remaining 167 groups.The testing result reveals that the systematic error of the well-trained neural network is 2.7e-3.After applying the proposed method to the rough 2D result,the deviation deceases from 0.69 mm to 1.23e-2 mm.The feasibility of a new direct parameter-3D result training pattern is also experimented,the result shows that within the 1600 groups of training datasets,the new method is incapable of reaching a prediction accuracy convincing enough for sensor's pre-design.A new cantilever-based multiple tire status parameter sensor is developed,and its functions are verified both theoretically and experimentally.The results prove that the multi-para sensor is capable of individually sensing tire pressure,rotational speed as well as tire velocity.A total number of 150 sets of FEM is performed and the promoted deviation deceasing methodology is test on this new multi-para sensor,showing that the methodology is also have a good performance on this sensor,suggesting an excellent generalization potential.The deviation revising method proposed in this thesis has high accuracy,high efficiency and excellent generalization capability.The simulation work as well as the method proposed in this thesis is useful to guide the rapid and high-efficient design of MEMS cantilever-based sensors.