Towards Data-constrained Data-driven Digital Twin for Design and Manufacturing

Abstract

This thesis addresses the issue of data availability in building data-driven models for digital twins in manufacturing and design. Digital twins enable the transition of computer models of products, assets and processes from the traditional static to a dynamic real-time representation that allows industry to simulate, predict behaviour and optimize performance. Digital models of the product, asset and process that consistently reflect the real-time status are at the heart of digital twins. The complexity of most processes or products constrains the availability of physics-based models. Therefore, data-driven approaches are critical for building an accurate representation. The challenges in building a data-driven model are due to the availability of data to construct such models being restricted by the available sources of such data, the costs associated with data acquisition and storage, and extending and updating the current models. The thesis addresses three issues with developing data-driven models of products and processes for digital twins in design and manufacturing when appropriate or sufficient data is not readily available. The first issue arises when heuristics or rules drive the process or system, and adequate data must be available to capture this. The second issue is one where a particular data type is critical to building the data model but not convenient or economical to obtain in real-time practice. The idea of a soft sensor is used to address this issue. The challenge here is in identifying and constructing the soft sensor for the parameters that are difficult to obtain. The final issue addressed pertains to extending the applicability of the data model built by using this model to construct digital models for a different configuration in the same domain. Each of these issues has been addressed in the context of specific case studies in which the problem arises. The design of a foam pad for a vehicle seat in the automotive industry is an example of a need for a data model where the process is driven by expert knowledge and the data sources are small. In the machining domain, it is essential to monitor tool wear. However, it is not easy to measure tool wear in situ. The cutting force is a critical parameter in data models for tool wear prediction. However, collecting data on cutting force is quite expensive and becomes a bottleneck in using the data model. A two-stage soft sensor to first predict cutting force and then the tool wear in machining operations is proposed to alleviate this issue. The machining domain also provides the context for the last issue, where the input-data varies due to changes in the process parameters or conditions. An ensemble-based data-driven method is proposed to solve the problem of model drift in predicting tool wear. In each case, data was generated through experiments or from synthetic models, and the data-driven models developed have shown improvement in prediction.