论文部分内容阅读
在应用C-Mn钢工业大数据进行神经网络建模时,如果将大量原始数据不加处理或者经过简单的剔除异常值处理后进行建模,很容易建立满足一定精度要求的模型。但是,如果进一步研究模型的规律性,却常常有违背客观规律的情况。这是由于原始数据中大量的数据相互干扰和生产数据的离散分布造成的。因此在建模过程中,需要将冗余和误差较大的数据剔除,保证训练数据和预测数据的均匀分布,这样能够在减小建模的计算量的同时保证数据具有显著的规律性,从而建立出合理的模型。文章利用Bayes神经网络建立了多种牌号C-Mn钢力学性能预测模型,并对影响屈服强度的工艺参数进行了分析。经统计,屈服强度和抗拉强度的预测数据中分别有96.64%和99.16%的数据预测值和实测值绝对误差在±30 MPa之内,伸长率的预测数据中有85.71%的数据预测值和实测值绝对误差在±4%内。
When using C-Mn steel for industrial big data to model neural network, if a large amount of raw data is not processed or simply treated by removing abnormal values, it is easy to build a model that meets certain accuracy requirements. However, if we further study the regularity of the model, we often violate the objective laws. This is due to the large amounts of data in the raw data that interfere with each other and the discrete distribution of production data. Therefore, in the modeling process, redundant and error-prone data need to be removed to ensure the uniform distribution of training data and prediction data, so as to reduce the computational load of the modeling and ensure the data has a significant regularity, thereby Establish a reasonable model. In this paper, Bayesian neural network was used to establish the mechanical properties prediction models for various grades of C-Mn steel. The process parameters affecting the yield strength were also analyzed. The statistics, yield strength and tensile strength of the predicted data were 96.64% and 99.16% of the data predicted and measured values of the absolute error within ± 30 MPa, the elongation of the predicted data 85.71% of the data predicted And the measured value of the absolute error within ± 4%.