Machine Learning and Low Cost Ultrasonic Sensors for the Optimisation of Industrial Mixing Processes

世界正在经历第四次工业革命，人工智能、机器人和物联网等数字技术被用于提高制造过程的生产率、效率和可持续性。工业4.0的基础是数据的获取和智能使用。因此，传感器是这一制造业转型的关键技术。混合是整个制造过程中最常见的过程之一，因为它不仅用于组合材料，还用于悬浮固体、增加热量和质量传递、提供通风和改变材料结构。虽然有几种传感技术可用于监测混合过程，但每种技术都有自己的应用和局限性。超声波传感器是低成本、实时、在线、非侵入性的，并且能够在不透明的系统中工作。以前几乎没有文献使用超声波传感器来监测混合情况。这项研究中采用的工业可应用的、非侵入性的反射模式传感技术，以及所调查的大量数据后处理，使这项工作与现有的文献相分离。学生将在实验室环境中开发几个模型混合系统，并在混合过程中获取传感器数据。监督机器学习(ML)通过将输入数据映射到输出值来进行训练，然后能够从新的输入数据预测输出值。分类ML模型将被训练以预测系统是否是非混合的或完全混合的。将开发回归ML模型来预测混合完成前的剩余时间。确定一个系统是混合的还是非混合的，将为工业过程带来好处，如减少不符合规格的产品和过度混合造成的资源消耗。对剩余混合时间的预测将允许更好的批处理调度，从而提高过程生产率。最大似然模型输入数据的质量影响预测性能。通常，需要一些专业的传感器或过程知识来从数据中设计有用的特征。因此，本研究的另一个方面是使用卷积神经网络(CNN)，它不需要从传感器数据中进行人工特征工程。通过使用CNN，可以减轻在工业过程中部署超声波传感器的操作员的负担。还将探索多传感器数据融合，将多个传感器的输出组合在一起，以产生比使用单个传感器所能实现的更大的最大似然性能。这项工作的进一步研究途径将集中在工业应用上。例如，与工业合作伙伴合作，监测他们的混合过程。此外，可以将重点放在克服可用于在工业环境中培训ML模型的有限产值的问题上。这是因为混合物状态的参考测量通常很难获得，成本很高，或者很耗时。克服这一困难的两种方法是迁移学习和半监督学习。转移学习涉及在类似的系统上训练ML模型，在该系统中更容易获得参考测量，然后使用该模型来帮助预测目标系统。半监督学习使用传感器数据中没有可用的输出值和提供输出值的传感器数据中的信息。

The world is undergoing the fourth industrial revolution where digital technologies such as artificial intelligence, robotics, and the Internet of Things are used to improve the productivity, efficiency and sustainability of manufacturing processes. Industry 4.0 is underpinned by the acquisition and intelligent use of data. Therefore, sensors are a key technology for this manufacturing transformation. Mixing is one of the most common processes across manufacturing as it is not only used for combining materials, but also for suspending solids, increasing heat and mass transfer, providing aeration, and modifying material structure. Although several sensing techniques are available for monitoring mixing process, each have their own applications and limitations. Ultrasonic sensors are low-cost, real-time, in-line, able to be non-invasive, and capable of operating in opaque systems. There is little previous literature using ultrasonic sensors to monitor mixing. The industrially applicable, non-invasive, reflection-mode sensing technique employed in this research, along with the extensive data post-processing investigated, separates this work from the existing literature. The student will develop several model mixing systems in a laboratory setting and acquire sensor data during the mixing processes. Supervised Machine Learning (ML) trains by mapping input data to output values to then be able to predict output values from new input data. Classification ML models will be trained to predict whether the system is non-mixed or fully mixed. Regression ML models will be developed to predict the time remaining until mixing completion. Determination of whether a system is mixed or non-mixed would provide industrial processes benefits such as less off-specification product and less resource consumption caused by over-mixing. Prediction of the mixing time remaining would allow for better batch scheduling and therefore process productivity. The quality of the input data for ML models effects the prediction performance. Often, some specialist sensor or process knowledge is needed to engineer useful features from the data. Therefore, another aspect of this research is to use Convolution Neural Networks (CNN) which require no manual feature engineering from the sensor data. By using CNNs, the burden on operators deploying ultrasonic sensors in industrial processes can be reduced. Multi-sensor data fusion, combining outputs from multiple sensors to produce greater ML performance over that which could be achieved using a single sensor, will also be explored. Further research avenues of this work will focus on industrial application. For example, working with industrial partners to monitor their mixing processes. In addition, focus can be on overcoming the problem of limited output values available for training ML models in industrial settings. This is because a reference measurement for the mixture's state is often difficult, expensive, or time-consuming to obtain. Two methods for overcoming this difficulty are transfer learning and semi-supervised learning. Transfer learning involves training a ML model on a similar system where it is easier to obtain reference measurements, and then using the model to aid in prediction of the target system. Semi-supervised learning uses information from the sensor data with no output values available as well as those with output values provided.