Explainable AI for Industrial Ultrasonics

工业结构中缺陷的检测和表征所必需的数据集越来越多地使用超声换能器相控阵来收集。一些NDT（无损检测）检测利用这种阵列已经是自动化的，但他们收集的数据的解释不是。数据解释的自动化是可取的，因为行业预计将严重缺乏适当的合格和有经验的人员，因为加速无损检测检查和提高其可靠性存在压力。最适合利用阵列数据的方法似乎是TFM（全聚焦法）。这将创建被检查组件部分的可视图像。当它们被噪声污染时，对这些图像的解释就不那么简单了。研究工程师开始探索利用机器学习技术自动解释无损检测数据的可能性。目前，这种方法的价值有限：首先，机器学习技术依赖于大数据，而没有大无损检测数据的存储库；其次，现代人工智能得出的结论往往是无法解释的，而核等安全关键行业不太可能采用这种性质的解释工具。Sound Mathematics Ltd.一直在研究另一种解决方案，这是一种应用软件，它将基于TFM的简单修改的信号处理算法与OpenCV图像处理算法和决策树（AI（增强智能）模块）相结合，该模块模仿人类检查员在编写检查报告时使用的思维过程。与机器学习算法相比，决策树有两个众所周知的优点和一个缺点：它们需要更少的数据集进行训练，它们产生可解释的结果，但它们提出的发展挑战是巨大的。以AutoNDE为例，该应用程序依赖于大约40个不同的参数。我们的研究人员花了数年时间才找到一组在各种配置下都能很好地工作的参数。如果成功完成，该项目将降低运营成本，提高核电站的安全性，这是我们最初的目标市场。在稍后的阶段，这项技术将被转移到海上风力涡轮机上。鉴于目前世界正在努力减少对石油和天然气的依赖，该项目非常及时。以后还可应用于各种压力容器、钢轨、造船用钢板、螺栓等的无损检测。这将降低潜在的运营成本，并提高这些行业的安全性。

Datasets necessary for detection and characterisation of defects in industrial structures are more and more often collected using phased arrays of ultrasonic transducers. Some NDT (Non-Destructive Testing) inspections utilising such arrays are already automated but interpretation of data they collect is not. Automation of data interpretation is desirable, because industry anticipates a severe shortage of suitably qualified and experienced personnel and because there is a pressure for both speeding the NDT inspections up and increasing their reliability.The approach best suited for utilising array data seems to be TFM (Total Focusing Method). This creates visual images of sections of inspected components. When they are contaminated by noise interpretation of such images is not straightforward. Research engineers began to explore the possibility of automating interpretation of NDT data by utilising machine learning techniques. At present, this approach has limited value: firstly, machine learning techniques rely on big data while there are no repositories of big NDT data; and secondly, conclusions reached by modern AIs are frequently unexplainable while safety critical industries, such as nuclear are unlikely to adopt interpretation tools of this nature. Sound Mathematics Ltd. has been working on an alternative solution, an application software that combines a signal processing algorithm based on a simple modification of a TFM with OpenCV image processing algorithms and a decision tree - an AI (Augmented Intelligence) module, which mimicks thought processes employed by human inspectors in writing inspection reports.Compared to machine learning algorithms, Decision Trees have two well-known advantages and one disadvantage: they need orders of magnitude fewer datasets for training, they produce explainable results, but developmental challenge they present is huge. To use AutoNDE as an example, the application relies on about 40 different parameters. It took our researchers years to zero-in on the set of parameters that appear to work well in a variety of configurations.If brought to successful conclusion, the proposed project would decrease operating costs and increase safety of nuclear plants, our initial target market. At a later stage the technology would be transferred to offshore wind turbines. The project is highly timely, in view of the current drive to reduce the world's dependence on oil and gas. Later still the technology can be applied to NDT of various pressure vessels, rails, steel plates used in shipbuilding industry, bolts _etc_. It would reduce the underlying operating costs and increase safety of these industries too.