Research On Prediction Algorithm Of Improved Extreme Learning Machine Based On Information Fusion

Extreme learning machine (ELM) is an effective learning algorithm, which is a new a type of single hidden layer feedforward neural network developed in recent years. The differences between ELM and traditional learning algorithm are as follows. Firstly, the input parameters of ELM are selected randomly, and it is needn’t to be adjusted. Secondly, the output weights are the least square which is obtained by minimum squared loss function. So ELM has fast learning speed and good generalization performance, and it has been widely used in many fields, such as pattern classification, nonlinear prediction and so on. However, in the process of learning ELM may inevitably exists many drawbacks. The random selection of input parameters of ELM leads to a series of non optimal parameters, as a result, the number of hidden layer nodes is more than the traditional learning algorithm, as well as the generalization performance is effected, and lead to the ill conditioned system; In the learning process ELM Only uses the input parameter information to compute and ignores the very valuable actual output values; When ELM is applied to industrial production, the precision can’t meet the actual standard and so on.In order to overcome the above shortcomings, this paper puts forward a kind of improved ELM-Particle Swarm Optimization Feedback Extreme Learning Machine (PSOF-ELM). The difference values between the actual forecast value of the network and the theoretical expectation are feedback to the input layer based on calman filter, so the Feedback Extreme Learning Machine (FELM) is formed, and the input weights are adjusted according to the feedback difference values. Then the particle swarm optimization algorithm is used to optimize the stochastic parameters of the network and the ill-conditioned system,4 candidate FELM prediction model are generated and combined to be a new network, thereby reducing the overall prediction error. In order to play the advantages of FELM which has a better results in the combination phase, the D-S evidence theory is used to allocate weight for every FELM according to their contribution in the combination forecasting model, and the fusion results are calculated based on these weights.The background of this paper is based on the strip rolling exit thickness prediction in industry. In the actual rolling process, there are many factors affect the strip thickness, and each factor has different effects on the strip thickness according to the tension control methods. The strip data is analyzed and filtered by data processing software ibaAnalyzer, and the values of the mutual information between the input parameters and output are calculated, so the characteristic parameters which have greater impact on the exit thickness are extracted, and then the prediction algorithm proposed in this paper is used to predict the exit thickness.In this paper the rolling production off-line experiment is made under the MATLAB platform. Comparing with the traditional ELM, BP and AI-SVM, the prediction algorithm proposed in this paper has smaller prediction error, better prediction accuracy and over-fitting performance. The fusion prediction results calculated by D-S evidence theory have higher prediction accuracy and stability, as well as smaller volatility comparing with the average fusion in combination phase.