A Study On The Reliability Of LEDs Using Neural Network

Light emitting diode is strongly supported by the state because of energy-saving and environmental protection.It is widely used in the fields of display, lighting, visible light communication and modern agriculture.The advantages of LEDs include high luminous efficiency, low driving voltage and so on.Currently,the problems such as light distribution and heat dissipating have been to find a more satisfactory solution,cost and reliability become major problems that hinder their development.Reliability is the ability to perform a required function under specified conditions,Combination of theoretical research, reference failure analysis methods and testing standards about LED reliability, we discuss the six levels of LED failure,select the chip and the package as a research object.Different from the current and temperature accelerated life test method,Proposed LED fast and multivariable life prediction model based on neural network.Based on the LED failure theory to choose the ideal factor, junction temperature,color drift as neural network input and measured the sample data.Determine the hidden layer neuron number by the empirical formula,select Levenberg self-learning algorithm to build BP neural network model with the structure is 6-12-1.The correlation coefficient reached 99.8%, it’s a good fit degree of training network.Input test data and the prediction error, the maximum error is 2.97%, Up to standard that the prediction accuracy of the life belong hundred hours.According to the connection weights and thresholds between neurons,Output the weight of parameters associated with LED reliability.Proof white LED reliability have a certain sensitivity with optical, electrical, thermal parameters.Neural network input select scientific and rational, color drift significantly affect the reliability.It confirmed the feasibility of the international LED failuer standards according color drift.Set up LED life prediction model with RBF neural network and fruit fly algorithm.There is fast convergence for Fruit fly algorithmand and less adjustment parameters for RBF neural network.Determines the basis function width of the RBF neural network by the taste function.Studies have shown that when fruit flies iterative algorithm the ninth to find the point of maximum concentration of the odor.Function to get the best width is 0.014.FOA-RBF neural network model constructed contains 20 hidden layer nodes, the largest forecast error is 6.59%, meeting expectations.