ABSTRACT A method for condition monitoring and localization of defects in mass-produced structural members using supervised learning is presented. An example for the effectiveness of the developed method comprises cantilevered carbon composite plate. In a numerical finite element model, the plate is partitioned into zones and a point mass is put on several locations within each zone. Point mass is treated as a pseudo-defect locally modifying structural properties of the plate. For each act of mass application, strain values are recorded and serve as defect-sensitive feature. Two variables of classification are tested – two different supervised learning algorithms (linear discriminant and non-linear k-nearest neighbours) and a limited number of strain data points per class which is varied in the range of 2 to 9 points. Several query points are simulated and subjected to classification in terms of belonging to particular zones of the partitioned plate. This step can be treated as a defect localization. It is shown that only 2 strain readings per class are sufficient for defect localization. The methodology is experimentally validated on a cantilevered carbon composite prepreg of the same dimensions and properties.

中文翻译：

---

基于应变特征的线性判别和最近邻分类的二维结构缺陷定位状态监测

摘要 提出了一种使用监督学习对大批量生产的结构构件中的缺陷进行状态监测和定位的方法。所开发方法的有效性的一个例子包括悬臂式碳复合板。在数值有限元模型中，板被划分为多个区域，并在每个区域内的多个位置放置一个点质量。点质量被视为局部修改板结构特性的伪缺陷。对于大规模应用的每个动作，应变值都被记录下来并用作缺陷敏感特征。测试了两个分类变量——两种不同的监督学习算法（线性判别和非线性 k 最近邻）和每类有限数量的应变数据点，在 2 到 9 个点的范围内变化。几个查询点被模拟并根据属于分区板的特定区域进行分类。这一步可以被视为缺陷定位。结果表明，每类只有 2 个应变读数就足以定位缺陷。该方法在具有相同尺寸和特性的悬臂式碳复合材料预浸料上进行了实验验证。