Research On Life Prediction And Early Warning Technology Of DC-DC Switching Power Supply

With the rapid development of new energy harvesting systems,5G communications and other fields,DC-DC switching power supply is an indispensable energy supply device,so the related reliability research around DC-DC switching power supply has attracted the attention of scientific researchers.The performance warning and life prediction technology for DC-DC switching power supplies has been rapidly developed,but there are also many problems,mainly including:(1)With the development of DC-DC switching power supply’s own modularization and packaging,it becomes more difficult to detect the parameters of internal components in a non-destructive state;(2)The results of multiple predictions based on data-driven prediction algorithms are uncertain;(3)Online prediction algorithms have shortcomings in data feature change perception,prediction timeliness,and anti-interference.Therefore,this paper first starts with the basic research on the failure mechanism of DC-DC switching power supply,and conducts the failure mechanism research on its main components,further demonstrates the feasibility of using the output voltage as a failure characterization parameter for performance in the early warning and remaining life prediction.Secondly,this article has designed an algorithm named WE-OSELM,and a remaining life prediction algorithm based on limited incremental learning and WE-OSELM(LWRLP).In the WE-OSELM algorithm design,so as to enhance the performance of the algorithm that is more sensitively perceive the changes in the prediction data characteristics during the online prediction process,the article designs the information perception weight,and effectively integrates it with the algorithm cost function,then derives the IPW-OSELM algorithm;At the same time,the article has put forward the theory that the online prediction ability of the prediction model is solidified when the model training is completed.By building the footing of the input data and the prediction error of the IPW-OSELM algorithm,the article proposes an error compensation algorithm;then effectively merges the IPW-OSELM algorithm and the error compensation algorithm to form the WE-OSELM algorithm through the double parallel SLFNs.The WE-OSELM algorithm’s online prediction ability is fully verified from the four indicators that are prediction accuracy,single maximum prediction time,generalization and stability,which is through the typical time series data such as Sinc and the Solar Energy engineering data set.The verification upshot shows that the WE-OSELM algorithm displays better than the comparison algorithm in terms of single-step prediction stability and timeliness.In the design process of the LWRLP algorithm,because the single-step prediction ability of WE-OSELM algorithm is accurate and stable,the WE-OSELM’s algorithm design is uesed to solve the problem that based on data-driven life prediction methods are peformed uncertainly in multiple predictions.Based on the the theory that hidden layer nodes are importance to the multi-step prediction effect of the the IPW-OSELM,this article uses the limited incremental learning method to optimize the training model parameter;verifies the remaining life prediction ability of the LWRLP algorithm by using the accelerated degradation data.Finally,the performance experiment under two temperature modes of constant temperature anxiety and gradual heating temperature anxiety are designed respectively with the B1205LS-1W module as the experimental object.This experiment discovered the gradual failure of the output voltage of this type of switching power supply module under complex high temperature stress conditions,and designed the experimental function circuit according to the failure characteristics of the power supply module,and set up an experimental platform.Simulating the lack of historical data of newly developed device,the WE-OSELM algorithm is used to perform online single-step and three-step early warning on the experimental data of DC-DC modules with less training data.The single-step warning result shows that the WE-OSELM algorithm realizes early warning failure in one or two steps in advance.This paper uses the LWRLP algorithm to select part of the experimental data under other temperature conditions as historical data to predict the remaining life of the experimental data under constant temperature stress mode at 135°C.The remaining life prediction results show that the LWRLP algorithm accurately predicts the remaining life within 2.75 seconds and 3.37 seconds respectively.The experimental results show that the relevant design algorithms in this paper have good engineering application capabilities for early warning and life prediction of DC-DC switching power supplies.