Research On Intelligent Recognition And Prediction Model Of Flatness For Cold Strip Mill

Strip production plays a very important part in the national economy, which is widely used in automotive vehicle, shipbuilding, bridge and household appliances etc. Flatness is the important quality index of strip, with the social development and the progress of science and technology, the requirements for the quality of cold-strip steel products by customers is more and more high, and the flatness problem have become more and more urgent to be solved.Flatness pattern recognition is the key technology of flatness control. Against the problems existed in the current flatness pattern recognition methods, and considering the practical situation of modern mill with many different flatness control means and the improvement of flatness control capability, in order to improve the precision of flatness pattern recognition, firstly, removing the noise in the measured flatness data by the preprocessing of wavelet de-noising technique, and then a flatness pattern recognition model based on quantum-behaved particle swarm optimization algorithm and BP algorithm mixed optimization BP network is established, which takes linear, quadratic, cubic and biquadratic Legendre orthogonal polynomials as flatness basic pattern. The simulation experiments show that, the model has strong anti-interference ability, high recognition precision and speed, it can provide reliable basis for the formulation of flatness control strategy; wavelet de-noising technique reduces actual flatness signal, and improves the anti-interference ability and precision of flatness pattern recognition.Flatness prediction model is important for the design of flatness control system. According to the existing flatness prediction models, in order to further improve the speed and precision of flatness prediction model, a QPSO-RBF network flatness prediction model based on measured data in the production is established. The model uses subtractive clustering algorithm to analyse training sample set and determine the number of cluster by an automatic cluster termination criterion firstly, which solves the problem that the number of hidden nodes is hard to determine, and then the optimization of network parameters is done by quantum-behaved particle swarm optimization algorithm and gradient descent algorithm, which makes the network performance achieve globally optimal. The simulation experiments show that, comparing to the conventional BP network model, the model has high prediction precision and speed, and is more suitable for the actual rolling process.