A novel hierarchical BP model for strip flatness pattern recognition

In order to overcome the weakness of traditional flatness defect pattern recognition by least squares method (LSM) proximity algorithm which is illegible on physical meaning and poor robust stability, as long as the low accuracy of common BP neuron network, a novel parallel flatness defect pattern recognition model based on binary tree hierarchical BP neural network and Legendre orthodoxy polynomial decomposition were presented, each node in the binary tree has the same structure but different weights. The precision of novel model was improved dramatically by classifying the prediction range and setting the binary tree depth. Experiment results show this novel hierarchical BP network performances are improved not only in precision but also in robust stabilization.