Despite global pressures, manufacturing remains a key industry in Europe, representing over 15% of total GDP* or 2.23 trillion USD. In order to stay competitive, it is critical for European manufacturing companies to continue innovating. Data-driven methods enable them to unlock the significant process and OEE (overall equipment efficiency) improvements needed to maintain their competitive advantage.
This year we had the pleasure to host the AI & Manufacturing track at Applied Machine Learning Days (AMLD) where we brought together 15 experts to discuss how AI is being applied and leveraged in the manufacturing space. The speakers came from different backgrounds, we had speakers from companies with large production lines such as Firmenich or Buhler that are relatively advanced on their digitalization journey and reaping the benefits. We had cloud providers MS Azure and AWS that spoke about problems they helped solve with their cloud offerings. We also had start-ups Amplo, T-Dab or Dida that are working on interesting challenges around helping production line workers, and also VU Engineering or Tornado that are building tools to help in the data labeling space. Finally, we had researchers from EPFL and Imperial College London that are developing digital twin solutions to help simulate factory environments.
Speakers at the AI & Manufacturing track at AMLD
In this article, we want to share and reflect on the main trends discussed during that day.
This was a common theme during many of the sessions. Starting and succeeding with Data and Machine Learning (ML) projects is challenging, particularly in the manufacturing space.
Successful projects require different teams and competencies to be brought together. Domain and business knowledge is coupled with data science and data engineering skills, as well as operator expertise and support from IT. Being able to orchestrate such teams is challenging and it is important to have the correct leadership in place that moves these teams in the same direction in a collaborative effort. Negating particular skills or teams can have detrimental effects on the success of such projects and initiatives. As highlighted by Alexis Roy (Head of BI & Advanced Analytics, Firmenich SA), a project needs strong support from the business to be able to align all the relevant stakeholders. Similarly, data analytics and machine learning cannot succeed in isolation from the domain experts “as the tools are likely to not meet the needs of users”.
Single small scale projects may deliver isolated results, but the scale is critical for manufacturers. Bringing the different teams and skills together is already challenging as mentioned earlier but typical production lines are often different from location to location, with different floor plans, machines, and sensors. Understanding these differences is good, but there is no silver bullet to solve this challenge. “One needs to focus on impactful projects first and, at the same time, build reusable components and tools that will speed up the next project.” according to Telina Reil (Director of Advanced Analytics, Abbott)
Ideally, such reusable components and tools initially start with quality data platforms and processes which allow companies to more easily scale their initiatives. Getting data is often a laborious and lengthy process, especially when it is spread among multiple silos. Once the data has been obtained, datasets often follow different semantics, and extra work is needed to form a coherent picture of a production process. Companies that overcome these challenges are able to build truly impactful Digital twins of their production lines. By federating everyone around a shared understanding of the production process, they become truly data-driven organizations.
“One needs to focus on impactful projects first and, at the same time, build reusable components and tools that will speed up the next project.”
— Telina Reil, Director of Advanced Analytics at Abbott
Dimitris Kyritsis going over the characteristics of a digital twin
The good news is that successful ML & AI initiatives can lead to upwards of 10% gains in increased productivity, as shared by multiple speakers. From stabilizing the production process to reducing stoppage, energy consumption, or optimizing scheduling, each area can unlock significant gains by switching to improved data-driven strategies and analyses.
ML & AI initiatives can lead to upwards of 10% gains in increased productivity.
Despite the organizational challenges, engineers were, on the other hand, quick to apply the latest developments in Machine Learning to solve their problems.
Computer vision neural networks have proven a valuable tool in product and defect inspection as shown by Lucas Vaudroux (VU Engineering) or Samir Araújo de Souza (AWS). Vision systems have started performing better than humans on some specific subtasks and automated inspection systems are becoming increasingly common. A key requirement is to develop the foundational tools to help with the tedious data collection and labeling process by operators. This allows supervised image classification algorithms with enough up-to-date and high-quality training samples to develop great models.
William Heurdier (Microsoft) demonstrated how Reinforcement learning is being applied to process control and scheduling optimization problems. By implementing control feedback loops using Reinforcement Learning, large improvements in process parameter stability can be unlocked. Instead of relying on slow and costly manual experimentation, reinforcement learning opens up the possibility to more rapidly explore the set of production parameters and learn optimal process parameters and stabilize the line.
Finally, semi-supervised learning techniques (a technique to help leverage unlabeled data) were shown to be very helpful when dealing with a small number of samples. Quality tests are sometimes only performed on a small subset of all units produced. The number of defective units returned by customers is often small. By applying autoencoders (e.g. Variational Autoencoders), it is possible to reduce the feature space and train models with a low amount of labels, as was shown by Roman Klis and Enno de Lange (Johnson Electric).
Vision neural network are becoming increasingly common to detect defective units as highlighted by Samir Araújo de Souza (top) and Lucas Vandroux (bottom)
Machine learning does not end when the first model has been developed. In order to truly unlock their benefits, ML models need to be operated over time. There are many challenges to face when moving from the POC stage to production.
First, models are living entities and it is important to have the data infrastructure and MLOps capability to be able to retrain them. From changing production conditions to new products and issues the production environment is constantly changing and so should the model. As highlighted by Ivan Scattergood, adapting to that new reality is hard: the infrastructure to automate retrain, and run models are often missing and companies are often lacking the right skills to build it.
Ivan Scattergood showcasing why having a model is not enough
Commercial, cloud, and open-source MLOps solutions are becoming more accessible but they still take significant effort to put in place.
Second, it is crucial to empower shop-floor workers to interact with the ML models. They need to be able to be a part of the continuous data labeling effort. Production environments and best practices are constantly evolving requiring model adjustment. Operators need to be able to easily report issues with the system and regularly provide updated labels. These new labels can in turn improve the model over time by handling a larger variety of scenarios.
Niels Uitterdijk highlighted that having a user-centric approach and really putting operators at the center of the model development is key. By enabling them to continually provide feedback on the model, one can continue improving the models with relatively low investment.
Productionizing ML Models on the production line cannot succeed without integrating back the right expertise and automating the deployment of new models with the proper infrastructure.
Niels Uitterdijk presenting a data-centric approach putting operators at the center
The last few years have brought their share of impactful events for the manufacturing industry: from a difficult labor market where recruiting operators has been challenging to COVID reshaping the supply chain worldwide. However, these obstacles also highlighted the need for faster adaptation and decision-making using data. In a competitive market, the improved efficiency in manufacturing brought by data and ML is becoming a key differentiating factor.
In parallel, there exists today a large offering from a multitude of players including cloud providers, start-ups, service providers, and others. Their solutions range from predictive maintenance, data platforms, automated quality control, and OEE optimization to data & MLOps. With most companies integrating with the cloud, it has become easier than ever to enter the field and build solutions using a mixture of out-of-the-box tools and custom solutions. Have you started your journey yet?
A big thank you again to all the speakers who have presented and shared their insight during the AI & Manufacturing track!
If you (or your company) are facing some of these challenges yourself (or have solved them), please reach out. We are always happy to exchange on these topics.
All the videos can be found on the AMLD Youtube channel: