Autonomous anomaly detection and self-healing in a smart test environment

为了到2050年实现净零排放，国际能源署表示，所需技术的50%尚未开发，因此需要在所有部门进行快速测试和开发。然而，测试强度的增加可能会对能源需求产生重大影响，而能源需求已经与向可再生能源发电的过渡相冲突。汽车行业继续支持并在许多情况下增长本已碳密集的运输行业;尽管在COVID-19大流行期间有所停顿，电动汽车销量增加，但道路运输仍相当于全球碳排放量的约28%。因此，支持汽车行业所需的脱碳，同时允许开发新的低碳技术的一个重点领域是提高车辆技术开发测试阶段的效率和功效。 由于实时限制和复杂且技术上精密的系统，物理测试是耗时的，并且因此是高度能源密集型的。此外，设置中的人为错误、故障或错误校准的传感器或不可预见的机械故障通常仅在测试后才被识别，并导致这些测试是冗余的，需要重复。虚拟测试环境在最大限度地减少这些测试方面发挥了作用，允许在开发过程中更早地运行模拟，并考虑更多的用例。然而，如果没有提供强大的物理数据，这些模型将无法准确地模拟硬件响应-因此，它们仍然依赖于用于调整和优化虚拟模型的测试数据的物理测试。为了使技术适用于市场发布，还需要进行一些物理测试，以解释产品差异、未知影响和模拟不准确性。由于动力系统开发过程的总体变化使得难以通过工程经验来补偿测量质量差，异常检测-一种发现数据中意外模式的方法-提供了一种可能的解决方案，可以最大限度地减少物理试验台时间，同时提高真实的数据的可靠性，以馈送到虚拟仿真模型中。如果应用于一系列动力总成单元，无论是内燃机、纯电驱动、燃料电池还是混合动力装置，它都有可能降低车辆测试和开发的能源强度，与此同时，提高了低-碳技术可以被释放到公共领域，以帮助交通部门的大规模脱碳。为了应对这一挑战，该项目将进行三个广泛的方法论。首先，探索性分析在不同背景下构成异常的内容;这将包括关于数据质量是什么以及它在应用程序中是否一致的定性研究。其次，使用历史或合成数据进行方法探索;这将探索不同的异常检测方法，无论是统计还是机器学习（包括但不限于分类，聚类和模糊逻辑），并在现有或人为更改的数据上测试这些方法。最后，一个特定的异常检测方法将在真实的试验台数据上反复完善和测试。在实践中，该项目的成果将通过减少冗余测试，从而提高用于虚拟模型和仿真的数据质量，促进在试验台上更有效地测试汽车技术。更有效的测试将产生两个关键影响，第一，允许新的低碳技术更快地开发和部署到消费者手中，从而有助于向净零排放社会过渡。其次，它将减少移动选项测试和开发阶段的能源强度，因为浪费的测试最少，虚拟模型更有效，这也将通过减少能源需求来支持脱碳目标。

To achieve net zero by 2050, the IEA have stated that 50% of the technology required is yet to be developed, thus, rapid testing and development in all sectors is required. Increased intensity of testing could however have a significant impact on energy demand, which already conflicts with a transition to renewable generation. The automotive industry continues to support, and in many instances grow, an already carbon intensive transport sector; despite lulls during the COVID-19 pandemic and increased electric vehicle sales, road transport still equates to around 28% of global carbon emissions. Therefore one area of focus to support the required decarbonisation of the automotive sector, whilst allowing new low-carbon technology to be developed, is associated with increasing the efficiency and efficacy of the testing phase of vehicle technology development. Physical testing is time consuming due to real-time constraints and complex and technologically delicate systems, and is thus, highly energy intensive. Additionally, human-errors in set up, faulty or mis-calibrated sensors, or unforeseen mechanical failures are often only identified post-test and result in these tests being redundant and needing to be repeated. Virtual testing environments play a role in minimising these tests, allowing simulations to be run earlier in the development process and for more use-cases to be considered. However, if not provided with robust physical data, these models will not be able to accurately simulate hardware responses - they therefore still rely on physical testing for the test data used to adapt and optimise the virtual models. Some physical testing will also still be required for technology to be suitable for market release to account for product variance, unknown effects and simulation inaccuracies. As the general change of the powertrain development process makes it harder to compensate poor measurement quality by engineering experience, anomaly detection - a method of finding unexpected patterns in data - presents a possible solution to minimise physical testbed time whilst increasing the reliability of real data to feed into virtual simulation models. If applied to a range of powertrain units, be it internal combustion engine, pure electrical drive, fuel cell or hybrid setup, it has the potential to reduce the energy intensity of vehicle testing and development, whilst simultaneously increasing the speed at which low-carbon technologies can be released into the public domain to aid large scale decarbonisation of the transport sector.To address this challenge this project will undertake three broad methodological approaches. First, an exploratory analysis of what constitutes an anomaly in different contexts; this will include qualitative studies regarding what data quality is and whether it is consistent across applications. Second, a methodological exploration using historic or synthetic data; this will explore different anomaly detection approaches, be it statistical or machine learning (including but not limited to classification, clustering and fuzzy logic), and test these on existing or artificially altered data. Finally, a specific anomaly detection approach will be refined and tested iteratively on real testbed data.In practise, the outputs of this project will facilitate more effective testing of automotive technology on a testbed by reducing redundant tests and subsequently improving the quality of data being used for virtual models and simulations. More effective testing will have two key impacts, first, allowing new, low-carbon technologies to be developed and deployed to the consumer faster, thus aiding the transition to a net zero society. Second, it will reduce the energy intensity of the testing and development phase of mobility options due to minimal wasted tests and more effective virtual models, which will also support decarbonisation targets through reduced energy demand.