【作者】 王永慧；

【导师】 颜云辉；

【作者基本信息】 东北大学 ， 机械设计及理论， 2009， 博士

【摘要】 随着市场变化及其结构的调整,高品质的板带钢在钢铁工业中的地位日益重要。它在汽车、家电、造船、航空航天等相关领域有重要应用并且需求量上升迅猛。由于连铸钢坯、轧制设备、加工工艺等多方面的原因,会导致板带钢表面出现焊缝、夹杂、抬头纹、氧化皮、结疤、辊印、刮伤等多种类型的缺陷。这些缺陷的存在,一方面降低了产品的抗腐蚀性、耐磨性和疲劳强度等性能,另一方面也使其无法应用于外观要求较高的领域。虽然我国的钢铁产量居世界第一,但每年却需要大量进口高品质的钢板。提高板带钢表面缺陷的检测与控制水平,具有重要的战略意义。模式识别是基于图像信息的检测技术的重要步骤。目前应用于板带钢表面缺陷检测系统的模式识别方法存在环节众多、识别率低、识别速度较慢、推广性差、适应性差等不足。其原因主要在于:就研究对象的特征而言,一方面不同类缺陷的图像在特征上并不存在很明确的界限,另一方面,同一类缺陷的图像在特征上存在较大差异;就常用模式分类方法而言,其分类机制与特征空间中类别的分布特性在表现方式、尺度、扩张性等方面存在的多种不一致,产生了聚类中心周围含有其它类别杂点、拟合划分面切开同类样本等问题,增加了误识的可能。对于以上存在的问题,可以通过构建适应性良好的分类模型加以解决。因此,本文在分类模型及具体的分类器实现方法上展开研究,具体研究内容和取得成果如下:(1)对模式分类机制进行深入研究,总结分析目前常用的分类方法存在的问题。针对这些问题,提出类别的本征空间和认知空间的概念,进而提出与类别在认知空间的分布状态的一致性较好的以多个单一类别的扩张体为基本元素的多体分类模型,给出了该模型的具体构建方法。通过理论分析和实例分析表明该模型与常用方法相比的在分类机制上优越性。(2)结合多体分类模型,对传统SOFM神经网络进行改进和优化,提出了WTM-SOFM分类法。该方法采用跟踪SOFM网络训练历史轨迹的方式克服了SOFM神经网络对类别间隔狭小或空间怠点分类能力的不足。实验研究表明它相比SOFM具有更强的边界冲突调解及边界扩张能力。(3)研究并提出一种新的模式分类方法,即版图分类法。该方法受到人类历史上版图形成过程中多体划分效果的启发,结合多体分类模型的要求而构建。与WTM-SOFM方法相比,它对未知分类空间的估计划分更具合理性。实例分析表明,在众多其它方法很难解决或根本无法计算的类边界复杂交错、多尺度等分类问题上,仍然可以表现出很高的识别率和识别速度。在板带钢表面缺陷分类实验中,对六种缺陷的识别率明显高于其它常用分类方法。更多还原

【Abstract】 As the market changes and restructuring, high-quality sheet steel in the steel industry’s position is becoming increasingly important. Its comsuption rapidly increasing in automotive, home appliance, shipbuilding, aerospace and other important applications concerned. As results coming from continuous casting billets, rolling equipment, relavent process, and many other reasons, there are weld, inclusion, wrinkles, scale, roller mark, scratches and various defects in strip surface. With these defections, the corrosion resistance, wear resistance, fatigue strength and other important performance of the product is reduced; on the other hand, it can not be used in the scopes with appearance quality demanded any more. As the world’s biggest iron and steel production country, plenty of high-quality steel strip is imported in China, every year. It becomes an important strategic significance to improve the surface quality of strip.Pattern recognition is an important step in detection technology based on the image information. Currently pattern recognition technology which is used in strip surface defect detection system has a lot problem such as complex recognition steps, low recognition rate, slow recognition speed, weak generalization ablility, poor adaptability and so on. The reason is mainly due to: the characteristics of the study, on one side, the image of different types of defects does not exist the clear boundaries in the characteristics space, on the other hand, the character of same type of defects might be really different; on common pattern classification methods, the classification mechanism and the characteristics of classes distribution in feature space exist in a variety of inconsistent in the way of expressions, scale, expansion feature and so on which produced the problem such as the situations of miscellaneous points around the cluster center, fitting surface divided into same class samples and so on, causeing higher possibility of error. In view of the above problems, it is possible by constructing a good model of adaptive classification process to be addressed. The classification model and the specific realization of the classifier is focused on for further improvment in the study.The research content and results are as follows:(1) Though the in-depth research of the pattern classification mechanism, the common problems and limitations of current classification methods is analysed. To solve these problems, the concept concerned classes eigen space and cognitive space are proposed, and then the multi-space classification model which has better consistency of the class distribution of cognitive space with the basic elements of many single type expansion subspace is established. The construction method is given for this model. Through theoretical analysis and case analysis shows that the superiority of the classification mechanism comparing with the common methods.(2) According to the requirments of the mulit-space classification model, WTM-SOFM classification method is proposed as the optimization of SOFM neural network. The method used to track the training history of SOFM network approach to overcome the less capacity of SOFM neural network for classification of the situations of cramped and idel space among classes centre in the feature space. The experimental study shows that it is a classifier with more cpapcity of the conflict mediation and the border expansion compared to SOFM.(3) proposed a new pattern classification method, namely, the territory classification method. This method is elicited from the multi-space division effect in the formation of the territory of the history of mankind, based on the request of mulit-space classification model. Comparing the method of WTM-SOFM, the division and estimation rules for the unknown classification space are more reasonable. Experimental analysis shows it can still show very high recognition rate and recognition speed on the situations of the complex boundary staggered, multi-scale and other issues which is difficult to solve or impossible to calculate for some comon ways. In the experiment of strip surface defects classification, the identification of the six defects was significantly higher than other common classification methods.更多还原