Exploiting Artificial Intelligence for PredictingSubcritical Failure of Microstructurally Disordered Materials

当材料在长时间内承受低于其短期强度的载荷时，就会发生亚临界(蠕变)破坏。与时间相关的过程，如塑性蠕变或化学反应，通常是热激活的，然后可能导致微结构损伤的逐渐积累，最终导致材料失效。即使是制造相同的样品，失效时间也可能表现出巨大的分散性。此外，在实验中很难获得较长的故障时间。因此，希望使用监测数据对剩余寿命进行样品或部件特定的预测。这可以避免与过早更换仍在运行的部件相关的成本，并减轻损坏部件在使用中的故障。在过去，这样的预测通常是基于蠕变速率-时间曲线的特征U型曲线，其中蠕变速率首先随着时间的推移减速(阶段I)，然后经过蠕变速率大致恒定的广泛的最小值(阶段II)，然后在失效前加速(阶段III)。然后，样本特定寿命预测可以基于蠕变速率最小值的位置，或者基于失效方法中的蠕变速率的时间依赖性，其可以通过在失效时间的有限时间奇异性在数学上描述。其他监测方法侧重于声发射速率的特征性增加，或最终破坏平面附近的损伤和变形活动的定位。拟议的项目调查是否可以利用人工智能/机器学习的方法，从表征蠕变过程中变形活动的空间和时间演变的监测数据中获得对剩余样本寿命的改进预测。为此，我们将使用从不同复杂程度的材料模型模拟获得的计算机生成的数据，以及从纸样的系列蠕变试验中获得的实验数据，并辅之以声学和光学监测。在每种情况下，我们都在处理描述单个样本蠕变历史的大量数据集(通常在模拟样本的情况下约为10000个，在实验样本的情况下约为100个)。这些数据集的一部分被用来训练所谓的神经网络，通过将蠕变活动的空间和时间模式与预测的故障时间相关联来对所谓的神经网络进行“预测”故障。然后使用剩余的数据来评估这些预测的质量，并将该方法与文献中描述的其他预测策略进行基准比较。

Subcritical (creep) failure occurs when materials are subjected over extended periods of time to loads below their short-time strength. Time dependent processes such as plastic creep or chemical reactions, which are often thermally activated, may then lead to a gradual accumulation of microstructural damage and ultimately to materials failure. Even for identically manufactured samples, failure times may exhibit a huge scatter. Furthermore, long failure times are difficult to access in experiment. It is therefore desirable to use monitoring data for sample- or component-specific prediction of residual lifetime. This can avoid costs associated with premature replacement of still functional parts as well as mitigate against in-service failure of damaged components. In the past, such predictions were often based on the characteristic U shape of the creep rate vs time curve, where creep rate first decelerates over time (Stage I) then passes a broad minimum with approximately constant creep rate (Stage II) and then accelerates in the run-up to failure 8Stage III). Sample specific lifetime predictions may then be based upon the location of the creep rate minimum, or upon the time dependency of creep rate in the approach to failure which may be described mathematically by a finite-time singularity at the failure time. Other monitoring approaches focus on characteristic increases in the rate of acoustic emissions, or on the localization of damage and deformation activity in the vicinity of the ultimate failure plane. The proposed project investigates whether methods of artificial intelligence / machine learning can be exploited to obtain improved predictions of residual sample lifetime from monitoring data which characterize the spatial and temporal evolution of deformation activity during creep. To this end we will use both computer generated data obtained from simulation of material models of different complexity, and on the other hand experimental data obtained from serial creep tests on paper samples accompanied by acoustic and optical monitoring. In each case we are dealing with large numbers of data sets (typically some 10000 in case of simulated samples and several 100 for experimental samples) which describe the creep history of individual samples. Part of these data sets is used to train so-called Neural Networks in ‘predicting’ failure by relating the spatial and temporal pattern of creep activity to a predicted failure time. The remaining data are then used to assess the quality of these predictions, and the method is benchmarked against other forecasting strategies described in the literature.