Study On Health Assessment System Of MOSEFT Based On Ensemble BP Neural Networks Model

The power devices,which are the core components of power converters,have been widely used in power systems,new energy power systems,aerospace and other fields.However,the power devices have suffered from alternating electrical and thermal stress for a long time when they work,which makes the power devices degrade,and increases the risk of system operation.Due to the lack of awareness of power devices' failure mechanism and monitoring for its health status during operation,most power electronic equipment relies on high cost to make operation safe and reliable.The internal evolution mechanism of power devices,and the evolution mechanism of electrical and thermal parameters in the failure process are studied in this paper.Then a health condition monitoring method and an evaluation model of the power devices are established by comparing and selecting the characteristic parameters that can characterize its health status.And an actual circuit is built to verify the entire analysis.The model could improve the operational reliability of the equipment and reduce the cost of design and maintenance.The main contents of this paper are as follows:(1)A 1:1 three-dimensional finite element model is established using the finite element method.The FE model is used to study the package failure mechanism of the power MOSFET under the coupling of electric field and temperature field.The bonding wire fatigue and the solder fatigue are simulated respectively,and it is found that solder fatigue is the main failure mode of the MOSFET.According to the fact that on-resistance can reflect the fatigue state of the device with higher sensitivity,the damage degree D is constructed as the fatigue evaluation index.The accuracy and feasibility of the damage degree D is regarded as the evaluation index are further verified by simulation.(2)A health state assessment model,which contains multiple BP neural networks model,is proposed.Firstly,the data,which is extracted from FE model,is used to train multiple BP neural networks.The trained multiple BP neural networks model will generate multiple data for each input sample,and these new samples are used as a new sample set.And it is found that the new sample set corresponding from each input sample conforms to the normal distribution.Therefore,the evaluation result and accuracy of model are calculated using the 3? criterion idea.The results show that when the interval width is 2.4s,the model accuracy is 94.47%,and the probability is 98.36%.The ensemble neural network model can evaluate the health status of the power MOSFET with high accuracy.(3)In order to verify the accuracy of the health assessment model in practical applications,a prototype for practical application is designed.Firstly,a BUCK converter used in aerospace area is designed.Sampling methods for model input parameters are designed,too.The experiments are carried out under different current levels and different damage degrees.The sampled data is input into the health assessment model,and the evaluation results are in line with expectations,and the accuracy of the evaluation model is verified.Finally,a health state assessment system,which is based on the health state assessment model and the interface of sampling circuit,is established.