Federated Learning for IoT promises to help transform the security, resiliency and even sustainability of our Critical Infrastructure across Energy, Utilities and Telecoms.
A problem of scale
Our present and future is one of IoT-driven, intelligent automation. There are already more IoT-connected devices on Earth than there are human beings. By 2030, there will be an estimated one trillion such devices with even more connections in the networks that they comprise. These devices have a vital role in the future of Critical Infrastructures such as energy, telecommunications, energy and renewables, water and utilities. Critical Infrastructure is a demanding environment. Here, IoT systems must operate robustly at a vast scale with high efficiency and security.
While AI-enabled connected devices have the potential to deliver greater efficiency, resilience and sustainability, the scale at which these systems must operate presents a significant challenge to current machine learning technology. Machine learning – particularly deep learning – depends on large volumes of data to learn complex statistical patterns and relationships. IoT systems exist in the real world, affected by changes over time of the devices themselves and the environment they exist in. They are often complex, with high degrees of diversity and heterogeneity of devices, environment, and data.
To achieve sufficient performance in the face of such complexity, machine learning must happen on a large scale. However, the distributed nature of IoT systems naturally creates data silos which are reinforced by barriers of ownership, regulation, security, and cost. The current paradigm relies on a cloud centric, centralised learning platform meaning large volumes of data must be moved from distributed devices, via paid for networks, before being stored and processed on computers sold by cloud infrastructure vendors. The result is that machine learning at scale rapidly runs into significant barriers of risk and cost, which if not overcome, results in poor performance and low system resiliency.
Cutting-edge AI – Federated Learning
OctaiPipe started as a prototype in 2019. We had been working for some years in industrial machine learning and IoT. Conversely, we had been undertaking R&D for complex machine learning in IoT systems with Dr Pedro Baiz at Imperial College, London, as part of our Innovation Sandbox initiative. Our work started in the field of deep reinforcement learning in marine autopilot systems. However, from late 2019, this progressed onto machine learning in IoT systems using Federated Learning – which was at that point a relatively novel concept in this context.
Federated Learning Operations (FL-Ops) enables the deployment of AI to the edge and the management of distributed learning across a network of intelligent devices. Rather than move data from Edge devices to the cloud to train AI algorithms – with all the data storage costs and security concerns that entails – Federated Learning (FL) instead trains algorithms on-device at the Edge with data shared between devices in a decentralised network for continuous, distributed learning.
This then reverses the normal centralised paradigm. Individual machine learning models are trained on many devices, then model parameters are returned to a central node to be aggregated into a single model. This Federated approach maintains the privacy of data at the device level, while also massively reducing the amount of data moved on the network and complexity and cost of computation in the cloud.
We undertook benchmarking studies on the well-known NASA C-MAPSS data set of turbofan engine failures. We satisfied ourselves that we could deliver performance equal to – or in excess of – the best cloud-trained deep learning architectures at a fraction of the cost in terms of network and cloud computing usage, and also with a massive advantage in data privacy and data security over conventional approaches.
Protecting critical industrial systems
Around the same time, we were introduced by our customer Crown Packaging to Carnaud Metal Box Engineering. CMBE was not only a subsidiary of Crown Packaging, but also the original equipment manufacturer (OEM) for the critical machines in the aluminium can manufacturing process. The world food supply chain and global FMCG industry depend on these processes for sufficient supply. In 2020, with the emergence of COVID, this was particularly true. Federated Learning, with the ability of machines to intelligently take automated actions to prevent failures, is able to increase production, improve energy efficiency and reduce wastage of material.
We have worked closely with our customers and community of data scientists over the past 12 months to understand their requirements. At OctaiPipe, we make it even easier to train, deploy and manage on-device Federated Learning at scale, while further enhancing data security, system resiliency, and lessening the network and cloud dependency. Society depends on the resilience, performance and security of our Critical Infrastructure, and by making Federated Learning for IoT easy to deploy, OctaiPipe is ensuring Critical Infrastructure can continue to be trusted in the age of AI.
If it can be trusted, it can be scaled
In late 2021, OctaiPipe was approached by Domin Fluid Power and Bath University to help them develop next generation electro-hydraulic servo-valves for actuation systems, and such systems are critical in everything from transportation to telecommunications and robotics.
To do so, machine learning needed to learn at a real world scale. Such a scale can only be obtained by learning across such a number of devices that again this must span multiple customers. Funded by Innovate UK we further developed OctaiPipe, in particular the ability of our Federated Learning engine to deliver edge Federated Learning on more resource constrained devices.
Deploying trusted AI to predict and reduce rare events
The release of ChatGPT coincided in 2022 and 2023 with an increasing social and political discourse surrounding the trustworthiness of AI systems. In 2023, Innovate UK BridgeAI released the two-phase Accelerating Trustworthy AI grant. This grant is designed to support the development of technologies and platforms that enable the development of Trustworthy AI systems. OctaiPipe was awarded funding for the first phase, the aim of which was to fund feasibility studies into challenges of trustworthy AI and assemble a consortium of companies to work together to solve these.
Nowhere is there a greater need for AI systems that can be trusted than in applications that keep human beings safe. Increasingly, blue chip manufacturers are installing computer vision systems to detect health and safety critical events, near misses and risk. In addition, there are solutions being developed to use computer vision to assess environments for safety and compliance risks. Thankfully, such events are rare.
However as these events are rare, this poses a challenge for machine learning models to learn and predict the future risk of rare events. Meaning devices must learn across multiple organisations in order to observe a sufficient number of cases.
Federated Learning across edge computing systems connected to CCTV provides a sufficiently high level of privacy preservation to reduce the risk associated with data sharing between different or competing entities. Federated Learning therefore builds trust between parties, enabling collaboration to achieve common goals such as workforce health and safety.
The impact of Innovate UK
The role of Innovate UK in supporting the development of OctaiPipe has been fundamental. As has long-term funding in consortium with partners and end customers, which has enabled the development of our technology in the context of real-world settings with sufficiently long timescales for meaningful progress. This in turn has allowed us to achieve sufficient market traction to successfully attract venture capital funding and a leading position in the market.
