KEEPER – knowledge creation for efficient and predictable industrial operations

Background

In this era of technological advancement, the focus has shifted towards more efficient and environmentally responsible monitoring of industrial systems. The integration of real-time analytics is paramount for prompt identification of inefficient and suboptimal operations, especially given the always-increasing complexity of modern systems and the increasing demands on environmentally sustainable operations. The latter creates significant challenges for the industry since it often means technology shifts (e.g., electric vehicles or new fuel mixes) where learning must be done from prohibitively few examples. Previously reliant on labour-intensive human input, the paradigm is shifting towards automated, self-regulating systems requiring minimal human oversight. These systems are characterised by the need to understand and capture context while learning from available operational data, typically with only a handful of human expert labels attached, to identify and address issues autonomously.

Leveraging the Internet of Things (IoT), contemporary industrial setups produce massive and ubiquitous data streams. These data streams are too complex for human operators, or even complete organisations, to comprehend readily. For original equipment manufacturers (OEMs), regardless of whether they are in a vehicle industry like Toyota or Volvo, the maritime sector like Alfa Laval, or the industrial network like HMS Networks, it is becoming increasingly important to gain an in-depth understanding of how the assets they produce are being used by their customers, what challenges they face, and how to best communicate with them. The necessary next step is AI/ML methods that can process the available data, which are largely unlabelled and only superficially understood.

However, the transition to complete autonomy is neither practical nor desirable for industry. The invaluable input of domain experts remains essential in shaping, fine-tuning, and contextualising knowledge creation systems. The symbiosis of AI/ML and expert insights ensures that automated solutions are not only technically sound but also aligned with broader business and societal considerations. KEEPER will develop foundational solutions for innovative services using self-supervised, semi-supervised, and meta-learning, high-fidelity synthetic data, efficiency optimisation of customer operations, and improved understanding of equipment usage patterns.

The project’s goal

The KEEPER project’s main objective is to develop advanced analytics techniques based on AI/ML to create knowledge and industrial value from real-world industrial data. This data is largely unlabelled with significant uncertainties, e.g., the fractions of healthy vs unhealthy equipment are unknown, faults are unverified, activities are only partially recognised, there is variation in the equipment configurations, and the data is noisy. By identifying informative patterns and recognising similar events in data streams, we aim to provide insights to enhance existing services' effectiveness and lead to new ones' creation. The project combines the research capabilities and AI/ML expertise of CAISR scientists at Halmstad University with technology capabilities, real-world data, and business insights from six industrial partners in maritime (Alfa Laval), industrial networks (HMS Networks), forklifts (Toyota Material Handling Europe), and heavy-duty automotive (Volvo Buses, Volvo Group Connected Solutions, Volvo Group Trucks Technology and Volvo Trucks) sectors.

The core KEEPER synergy question is how to best do AI/ML-based knowledge creation from largely unlabelled industrial data that leads to more efficient industry operations. This necessarily includes making sense of all available data, i.e., instead of striving for the unachievable ideal of perfect data that would correspond to well-controlled lab experiments, we learn from the industrial data that exists “in the wild”: either is being collected today, or is realistic to collect in the near future. The “largely unlabelled” phrasing means not only the ratio, i.e., the large amount of available data being unlabelled. It also, or maybe even primarily, means that many concepts of interest lack labels. While the vast majority of recent progress in AI/ML focuses on the supervised learning paradigm, we primarily work with what one could call “semi-unlabelled data.” Most of the data collection in the industry is done for purposes other than analytics, and thus, AI/ML uses data only tangentially relevant to the task at hand. It typically describes the operation of the equipment, or financial transactions and other business processes, but on-target labels are virtually non-existent. Some information can be used as labels, but even those are rare and unreliable. Most often, those labels come from auxiliary sources, and one can consider them “proxy” information, i.e., their meaning does not correspond to the things we are truly after but is associated (to different degrees) with them.

The next step, moving from failure prediction to efficiency optimisation, is even more challenging. While historical repair information is far from perfect, it is still greatly beneficial; a fully unsupervised approach would be completely infeasible. Even bigger challenges exist for energy efficiency, as there are no reliable labels on who the best drivers/operators are. While energy consumption can be used as an indication, one needs to account for confounding factors – which are plenty but often unknown and need to be discovered first and quantified next. Thus, the key challenge to be solved in KEEPER is how to best make use of unlabelled data supplemented with auxiliary information, not excluding supervised ML but accepting its limitations. Realistically, doing that means we need to develop new methods and algorithms that can identify structures, find patterns and create knowledge from data that is only partially understood and only partially relevant to the given task.

An under-utilised resource is high-frequency data streaming onboard vehicles, sea vessels, and within industrial installations. Even with limited labels, a lot can be learned about complex equipment that is continuously generating this data. Taking a system-of-systems perspective, even without perfect labels, one can understand much of the typical operation, building monitoring AI/ML systems capable of anomaly detection. Assisted by Explainable Artificial Intelligence (XAI) techniques, the knowledge acquired by the system can then be conveyed to human experts (operators, managers, technicians, designers), supporting their decision-making. This way, KEEPER will create tools that go beyond simply automating the ”status quo” and instead enable a more efficient industry – as in the example above, not just learn to predict maintenance needs according to current repair practices but provide insight into premature replacements, common diagnostics mistakes, inappropriate configurations, and more.