Soft-measuring Of Hot-rolled Strip's Mechanical Properties Based On Improved BP Algorithm

Nowadays, hot-rolled steel producers concerned more about the quality of their products and flexible production capacity of production lines. The traditional steel production is faced with challenges, and the traditional technology needs to be improved. The critical and necessary means to solve the above-mentioned problem is using the soft-sensing model in hot rolling. Compared with other methods, the BP algorithm is more suitable. So, the paper establishes a soft-sensing model based on BP algorithm.The primary work of establishing the soft-sensing model is arranging the data to be used in network training. Aiming at the shortcomings of site-collected data that it contains noise, the paper uses the data mining theory for reference, and deals with the site-collected data so that it contains less noise to provide reliable and sufficient data samples for the model based on BP algorithm. Then, based on the traditional experience, the necessary input and output parameters of BP algorithm are selected from the data samples. The input parameters contains the original chemical composition and hot-rolled parameters, and the output parameters contains the mechanical properties parameters. BP network shows the relationship between the chemical composition process parameters and mechanical properties.In this paper, the improved BP algorithm model is established based on soft-sensing technology, using C++language. The model contains three hidden layer, which can make the training process more complete and improve precision of the model. However, it raises the training time, which rises from 0.07583s/time to 0.6127s/time. The model setted is then used to predict the mechanical properties of steel, and the test results is satisfied. It reaches to the purpose that the precision is improved. The test-error of yield strength reaches to 16.8098Mpa, and the test-error of tensile strength reaches to 10.7411Mpa. At the same time, the test-error of extension rate reaches to 4.1932.Finally, aiming to solve the problems that the training of BP algorithm is slow and its precision is low, the paper uses several methods to modify the model separately, and then the modified models are separately used in experiments to test their performance.