Distributed learning with IoT for quality control in manufacturing
< Project overview >
This mini-project built the foundation of online prediction of quality in manufacturing processes by analysing the use of advanced machine learning to integrate process/material data with machine IoT data to improve prediction accuracy levels for aiding quality control. Our project has showed that IoT based machine learning is better than traditional benchmarks in predicting the control limits a product will conform to.
We have validated our developed method via multiple industrial case studies and also developed a demonstrator for how our methods can be deployed using a multi-agent system. This work also kickstarted a collaboration between the University of Sheffield, University of Cambridge and Tinsley Bridge.
The project objectives were as follows:
Identify the state-of-the-art in IoT data capture and AI algorithms that are appropriate to predict quality in manufacturing processes.
Develop AI algorithms that leverage the integration/collaboration between multiple data resources to provide increased prediction accuracy and self-tuning capability.
Evaluate the improvement in the quality prediction substantiated by augmenting machine event data with IoT environmental data.
Validate and verify results through a demonstrator.
We aimed to address the following barriers:
Lack of understanding of the full landscape of possible architectures for a possible IoT solution and how IoT will/can generate value in a given application domain. This project will help companies evaluate the potential IoT data will add on standard quality optimisation approaches through a demonstrative case study.
Incorporation/streamlining of IoT based applications/decisions with existing business processes. Companies worry about their existing business processes being adversely affected by IoT based automation. This project will demonstrate and discuss the extent to which business processes will need to be changed.
What was done?
1. We have captured and consolidated data streams collected from a number of manufacturing environments, comprising machine event data and IoT environmental data to augment existing process data, with several manufacturers involved in the project including Tinsley Bridge and Federal Mogul, research institutes such as AFRC and University of Strathclyde. Methods were further tested with publicly available datasets within the domain of manufacturing quality analytics.
2. Our first project goal was to find a method to make IoT based data work alongside traditional quality control data obtained in manufacturing processes. The reason for this is that when IoT is deployed in a manufacturing system, legacy data still needs to be considered as sub-quality production is a rarity and the manufacturer cannot wait until new cases occur. With this in mind, we have developed a new method that combined deep learning and transfer learning for process condition prediction.
The featured method was applied to an industrial case study from a legacy electrophoresis painting plant where an IIoT system was recently deployed. The results show that our proposed method:
Enables the training of a deep learning model based on a small size dataset containing limited observations of low probability incidents.
Utilises low-quality historical records by use of transfer learning.
Enables zero feature engineering thus reducing dependency on the data processing skills and domain knowledge. The findings from this case study were submitted for publication to a journal, and the dataset was made available on Kaggle to kickstart new research in this area. The company in question adopted the method in its plant.
3. We also developed machine learning methods for control limit violation across multiple processes, by extending and testing previously developed algorithms that leverage the integration of multiple data streams to provide increased accuracy in identifying quality violations using IoT data. A novel deep learning algorithm called Bayesian Autoencoder was developed for this which quantifies predictive uncertainty.
4. We designed and conducted experiments to evaluate how much IoT data can augment limits in these sample sizes. Based on our findings, we offered recommendations on balancing data to achieve greater prediction accuracy. These methods were tested with a SCANIA dataset and findings from it was published in a journal paper.
5. To consolidate our developed methods and aid industrialists for quick adoption of IoT analytics, we further developed a machine learning algorithm testbed using an agent-based system and will open-source the code via GitHub.
Industrial adoption and journal paper one. Developed a novel deep learning and transfer learning method for process condition prediction with IoT which able to train the model with small-size and low-quality dataset. It is now deployed in a paint shop.
Journal paper two. Conducted thorough experiments and recommended data augmentation methods on balancing IoT based manufacturing data sets for achieving better accuracy.
Developed a novel deep learning algorithm called Bayesian Autoencoders which quantify epistemic and aleatoric uncertainties in the data sources. The code has been made available on GitHub.
Easy adoption testbed. Developed a machine learning testbed using an agent-based system which reflects the data flow of the multi-stage manufacturing process.
Demo platform. Developed desktop size experiment and demonstration platform for IIoT and AI integrated technologies.
Deliverables and other tangible outputs
Fathy, Y, Jaber, M, and Brintrup, A (2021) ‘Learning with imbalanced data in smart manufacturing: a comparative analysis’, IEEE Access, 9, pp. 2734–2757. doi: https://doi.org/10.1109/ACCESS.2020.3047838
Song, B, Brintrup, A, Turner, C, Kelly, A, and Tiwari, A (2021) ‘Deep transfer learning for Industry Internet of Things based predictive maintenance’, IEEE Systems Journal (submitted).
3. Agent-based system for distributed quality analytics
software source code
4. Dataset collected from the project has been published at Kaggle
‘E-coating ultrafiltration maintenance dataset’, Kaggle, 2020. doi: 0.34740/KAGGLE/DS/889404
Knowledge transfer between multiple academic partners and industry partners have been established (SCANIA, Tinsley Bridge, Federal Mogul, University of Sheffield and Cambridge). Industry partners introduced their real-world study cases and explored the IoT barriers with academic partners. The IoT techniques have been developed by academic partners and been demonstrated to a mixed community including research groups and industry users.
Further research and industrial projects are planned for extending the algorithms and agent-based testbed to more use cases.
Finalise the experimental platform and demonstrators.
Disseminate the project outcomes in IoT and ML conferences.
Our collaboration manifested in a positive way as the University of Sheffield provided their hardware and software expertise in developing the IoT data capture, University of Cambridge provided advanced data analytics capabilities while Tinsley Bridge provided the factory test environment. By combining our capabilities and knowledge, we learned to overcome challenges together and moved towards a higher quality outcome.
Physical visits to the factory have been limited by the Covid lock down, however, it would be very beneficial if possible to have actual visits to academic partners and industry partners. In particular, visiting the manufacturing shop floor would help the researchers understand the problems, visiting the lab experiment would help the industry partner understand the analytic workflow.
A more effective communication platform during working from home situation between each member of the Pitch-In consortium would have made collaboration more efficient. There is also a lack of recommended data sharing platforms between members of the consortium and for approaching industrial collaborators.
Access to deliverables, resources and media content
Published normalised dataset which was collected and applied to ML algorithm development within this project
What has Pitch-In done for you?
Pitch-In project has provided the opportunities to establish the partnership and network with leading universities and regional SMEs to develop and adopt emerging IoT technology. The feasibility studies and industry use cases gained from this project have become the foundation and expanded into further research. The demonstrators developed from this project have played a key role to introduce digital technology for widen academic and industrial communities.
Dr Alexandra Brintrup – Lecturer in Digital Manufacturing at the University of Cambridge.