基于PSO-PNN的铸坯全长表面纵裂纹在线预测

表面纵裂纹是铸坯质量缺陷中一种最常见的表面质量缺陷。环境因素使得铸坯表面纵裂纹在线检测精度不高,各大钢厂铸坯质检仍要依赖人工,因此提出一种基于粒子群PSO优化概率神经网络PNN的铸坯全长表面纵裂纹预测方法。首先,建立铸坯生产过程跟踪及数据时空变换模型,构建铸坯生产系统将生产过程数据与铸坯长度方向进行匹配;再利用PNN的Bayes 最小风险准则进行有监督特征学习,并利用寻优算法PSO优化PNN关键参数的选取,得到最终的模型PSO-PNN;最后,利用某钢厂连铸产线铸坯质量缺陷数据和生产过程数据进行试验验证。结果表明,该方法对铸坯整体的质量预测分类精度达到97.5%,铸坯全长的裂纹缺陷的预测精确率和召回率均在92%以上,能有效实现铸坯全长表面纵裂纹的预测,为现场质检人员提供参考。

On-line prediction for surface longitudinal crack of continuous casting slab on length direction based on PSO-PNN

Surface longitudinal crack is one of the most common surface defects on continuous casting slabs. Due to environmental factors, the on-line detection accuracy of longitudinal cracks on the surface of casting slab is not high, and the quality inspection of casting slab in major steel mills still depends on manual work. Therefore, a method of predicting longitudinal cracks on the surface of casting slab based on particle swarm optimization probabilistic neural network PNN is proposed. Firstly, continuous casting production process tracking and data time-space transformation was established to match the production process data with the slab on length direction. The Bayes minimum risk criterion of PNN was used for supervised feature learning, and the optimization algorithm PSO was used to optimize the selection of key parameters of PNN, and the final model PSO-PNN was obtained. Finally, the quality defect data and production process data of continuous casting line in a steel mill are used for experimental verification. The results show that the classification accuracy of the method is 97.5% for the whole slab and precision and recall for surface longitudinal crack of slab on length direction are above 92%, which can effectively realize the prediction of the longitudinal cracks on the surface of the full length of the slab, and provide a reference for on-site quality inspection personnel.