Enhancing the performance of an electrical tomography based multi-phase flow meter using machine learning algorithms

产品配方监测在快速消费品(FMCG)等制造行业的性能和管理中具有重要意义。几十年来，通过质量控制(QC)实验室的直接抽样方法一直是衡量产品质量的主要方法。然而，这些以抽样为基础的方法往往费时且不能代表真实的产品条件。其他控制是基于时间的(例如，清洗15分钟；混合3小时)，当过程测量可以显著节省时。ITS开发了一种工业断层扫描仪，该仪器利用电子过程断层扫描技术，能够提供准确的在线浓度/质量测量。该工业仪表已应用于制造领域，性能良好。据信，同样的技术可以更广泛地应用于快速消费品行业，如在线产品识别，以减少浪费，提高连续生产线的生产率。该技术可进一步扩展到现场清洗(CIP)的实时测量。ITS的层析成像仪器的原理主要基于多相电学测量，因为上述应用都涉及多种成分，这些成分具有不同的电学性质。为了使用ITS层析成像仪监测多相过程，需要一个外部仪器来跟踪初级液体的电导率变化，这样系统输出就可以独立于背景波动，而背景波动可能会受到水介质的离子浓度或温度的影响。理想情况下，电导探头可以安装在取样箱中，探头测量可以潜在地反映传感器中液体的现场状态。然而，在实际应用中，要找到一个能够可靠地指示原始材料的导电性的合适位置是具有挑战性的。此外，由于流动扰动和产品污染的风险，探头不能与工艺管道串联安装，为了克服这一问题，ITS开发了一种机器学习算法，可以根据现有的原始传感器测量结果提取液体电导率的变化，使系统能够准确监测在线液体电导率，旨在消除系统对电导率探头的依赖。此外，对于单相过程，机器学习算法还可以用于识别产品配方或批次过渡过程的完整性，实现相对于产品配方的真正现场流体质量控制。这一创新已在实验室规模进行测试，并被证明在一系列工艺条件下都是稳健的(TRL4)。下一阶段将研究机器学习算法在放大处理环境中应用时的稳健性(TRL5)。萨斯喀彻温省研究委员会(SRC)和伯明翰大学(UOB)的Flow Rips设施为使用他们的基准技术(如高分辨率伽马射线断层扫描和正电子发射断层扫描(PET))表征算法性能提供了理想的平台。来自国家研究委员会(NRC)的机器学习知识和专业知识也将在算法优化方面为该项目增加重大价值。

The monitoring of product formulation is of significance in the performance and management of many manufacturingindustries such as fast-moving consumable goods (FMCG) production.For several decades, the direct sampling methods via a quality control (QC) laboratory have been the principle method ofmeasuring product quality. However, these sampling-based methods are often time-consuming and unrepresentative totrue product conditions. Other controls are time-based (e.g. clean for 15mins; mix for 3hrs) when an in-processmeasurement can make significant savings.ITS has developed an industrial tomography meter that utilizes electrical process tomography technology and is able tooffer an accurate in-line concentration / quality measurement. The industrial meter has already been applied intomanufacturing applications with good performance. It is believed that the same technology can be more widely applied toFMCG industries such as in-line product recognition to reduce waste and increase productivity in continuous productionlines. The technology can be further extended to real-time measurement of clean in place (CIP).The principle of ITS's tomography meter is mainly based on multiphase electrical measurements, as the above-mentionedapplications all involve multiple ingredients, which have distinct electrical properties. For monitoring multi-phase processesusing an ITS tomography meter, an external instrument is required to track the conductivity variations of the primary liquid,so the system output can be independent to the background fluctuations, which could be influenced by either ionic concentration or temperature of the aqueous medium. Ideally, a conductivity probe can be installed in a sampling tank,where the probe measurement can potentially reflect the in-situ status of the liquid in the sensor. However practically, it ischallenging to find a suitable location that can reliably indicate the conductivity property of the primary material. In addition,the probe cannot be installed in series with the process pipelines due to the flow disturbances and the risk of productcontamination.To overcome the issue, ITS has developed a machine learning algorithm that can extract the liquid conductivity changebased on the existing raw sensor measurements, this new algorithm allows the system to accurately monitor the in-lineliquid conductivity, aiming to eliminate the system dependency on the conductivity probe. Furthermore, for single-phaseprocesses, the machines learning algorithm can also be used to identify the completeness of product formulation or batchtransition process, which realises a true in-situ fluid quality control relative to the product formulation.This innovation has been tested at a lab scale and shown to be robust across a range of process conditions (TRL4). Thenext stage will be investigating the robustness of the machine learning algorithm when it is applied in scale-up processenvironments (TRL5). The flow rigs facility from Saskatchewan Research Council (SRC) and the University of Birmingham(UoB) provides an ideal platform for characterising the algorithm performance using their benchmark technologies, such ashigh-resolution gamma-ray tomography and Positron-emission tomography (PET). The machine learning knowledge andexpertise from the National Research Council (NRC) would also add significant value to the project in terms of algorithmoptimisation.