Improved characterization of failure behaviours of sheet metals based on pattern recognition methods

该研究项目的目的是借助模式识别方法对板材在成形过程中的失效行为进行客观分类。为了将FLC扩展到失效阶段，专家对成形过程的视频序列进行了注释，并将序列的单个图像归类为以下不同的失效类别：均匀成形、扩散颈缩、局部缩颈和裂纹萌生。基于专家注释，开发了一种有监督的模式识别方法，提供对各个故障阶段的自动分类。特别是，对于局部颈缩类别的灵敏度高达92%，而应变分布中专家注释和伪影的一致性负面地损害了结果的质量。为了实现专家的独立性和对新材料的可移植性，还开发了一种无监督模式识别方法，使FLC具有不同的失效概率，并在相同安全裕度的情况下实现了比传统评估方法更高的成形极限。这些结果还得到了金相研究的进一步证实。在新的项目阶段，将研究将该方法扩展到更多的材料表征试验。挑战不仅是机器学习方法向潜在更健壮的深度学习方法的发展，而且还包括在不同应变条件下对材料机制的综合分析。通过单轴拉伸试验、缺口拉伸试验和液压胀形试验对评价方法的适用性进行了评价，并将其应用范围扩展到弯曲试验。通过机器学习确定的实验的极限应变通过金相调查进行了验证，并在单独的实验中另外确定了材料特定的行为模式。此外，在第一阶段，可以确定应变分布的不连续性，例如局部最大值，可以被自动检测。然而，客观失效定义的确定是复杂的，而且还不能最终确定客观失效定义。一方面，金相结果只能有限地转移到模式识别，因为观察到的不连续在微米范围内，因此只能通过应变测量部分捕获。另一方面，不连续性是可能与材料破坏不一致的应变局部化。在此基础上，研究了不连续性对方法学的影响。为了量化材料的性能，在100多个拉伸试验的基础上进行了随机评估，以确定在材料老化过程中可以检测到哪些材料的性能、成型阶段和模式。

The aim of the research project is an objective classification of failure behaviour of sheet metals during forming processes with the aid of pattern recognition methods. In order to extend the FLC with failure stages, the video sequences of the forming processes were annotated by experts and the individual images of the sequences were assigned to the following different failure classes: homogeneous forming, diffuse necking, localised necking and crack initiation. Based on the expert annotations, a supervised pattern recognition method was developed that provides automatic classification into the respective failure stages. In particular, a sensitivity of up to 92% was achieved for the local necking class, whereas the consistency of the expert annotations and artifacts in the strain distributions negatively impaired the quality of the results. In order to achieve expert independence and transferability to new materials, an unsupervised pattern recognition method was additionally developed, which as a result provides an FLC with different failure probabilities and achieves higher forming limits compared to conventional evaluation procedures with the same safety margin. These results were additionally validated by metallographic investigations. In the new project phase an extension of the methodology to additional material characterization tests will be investigated. The challenges are not only the development of the machine learning approach to potentially more robust deep learning methods, but also the comprehensive analysis of material mechanisms under different strain conditions. The transferability of the evaluation method is evaluated by means of uniaxial tensile, notch tensile and hydraulic bulge tests and the application spectrum is additionally extended to the bending test. The limit strains of the experiments, determined by machine learning, are validated with metallographic investigations and material-specific behavior patterns are additionally identified in the individual experiments. Furthermore, in the first phase it was ascertained that discontinuity in the strain distribution, such as local maxima, can be detected automatically. However, the determination of an objective failure definition is complex and an objective failure definition cannot yet be conclusively established. On the one hand, the metallographic results can only be transferred to pattern recognition to a limited extent, since the observed discontinuities are in the micrometer range and are therefore only partially captured by the strain measurement. On the other hand, the discontinuities are strain localizations that may not coincide with the material failure. Hereby, the effects of the discontinuity on the methodology are investigated. In order to quantify the material properties, a stochastic evaluation based on more than 100 tensile tests is carried out to determine which material properties, forming stages and patterns can be detected during material aging.