Research On Multiple Neural Networks Soft-sensor Models Of Grinding Granularity

As a result of the complex conditions of grinding classification field and the expensive detecting device, not only prediction accuracy of grinding granularity which is used on-line detection is not high but also the price is high.Therefore this paper proposes a new prediction method has important significance. The new prediction method combines the artificial neural network and soft measurement.The specific works are as follows:First of all,inspiring by the thought that the multi-source information fusion technology can improve the prediction accuracy and robustness of model, a multiple neural networks soft sensing model of grinding granularity which is based on kalman filter is put forward. According to the characteristics of grinding process and accuracy requirements of soft measurement,wavelet neural network, support vector machine(SVM) regression and RBF neural network model which is based on the orthogonal least squares are established respectively.And then kalman filter algorithm is adopted to compromises outputs of the three neural network models.Finally using the fused model and three neural network models to predict the grinding particle size separately and comparing the results of four models.Second,according to the idea of spatial data clustering segmentation of model can improve the prediction accuracy and robustness, a multiple BP neural networks soft sensing model of grinding granularity is proposed which is based on fuzzy c-means(FCM) clustering algorithm. FCM is used for clustering segmentation of grinding granularity data space according to the characteristics of grinding process and accuracy requirements of soft measurement. For single subspace of grinding granularity, BP neural network soft measurement model is set up and optimized by improved PSO(Particle swarm optimization, PSO) algorithm. The advantage of improved PSO is that it has adaptive mutation operator and quadratic subsection inertia weights operator. And then the multiple BPNN models which are optimized by PSO to forecast grinding granularity,simulation results show that the proposed model can satisfy the real-time control of grinding production process requirements.Finally,according to characteristics and a lot of experimental data in the process of actual grinding classification,mechanism model of grinding granularity is deduced.And then BPNN model and WNN model are built. Mixed multi-model switching is realized which is based on hysteresis switching strategy.Namely according to the switching performance index, selecting the optimal as the current model at each sampling instant, so as to realize the adaptive control of the whole operation. To improve the prediction accuracy,performance index of hysteresis switching strategy is optimized by cuckoo search algorithm(cs).Simulation results show that the proposed model has better generalization results and the prediction precision.In conclusion,according to the simulation results,three kinds of multiple neural networks soft-sensor models of grinding granularity can achieve good prediction,and prediction accuracy and robustness of grinding granularity are improved than a signal model.In a word, those multiple neural networks soft-sensor models can satisfy requirements of grinding process real-time control.