As AI evolves, it is evident that the most powerful models will be cloud-based, and hosted in data centres that are beyond the control of the average business. Not only that, there is now a reliance on connectivity that makes cloud-based AI feasible only in certain use cases within the manufacturing environment. Essential plant safety and control systems, for example, cannot rely on the internet. Certain control functions and scada might exist in the cloud, but care needs to be taken to ensure that these systems have redundancy, and can fail in a way that does not severely disrupt manufacturing operations or introduce a safety risk. The practical application of AI in manufacturing control and automation will only be possible if some of the computing workloads can be brought onto the plant, inside the firewall and inside the plant network.
Edge computing involves processing data close to where it is generated, rather than relying on the cloud. In practice, this means that sensors are connected to local computing devices using the plant network where latency is low, network congestion can be controlled, and connectivity can be relied on. Edge computing can therefore improve the reliability of advanced computing such as AI in the context of factory systems.
There are several benefits to edge computing in the factory:
• Improved reliability: By minimising dependency on internet connectivity and data centres outside your control, you can ensure that factory operations continue smoothly at all times.
• Enhanced security: Privacy and cybersecurity are of paramount importance to any industrial operation, and you need to be vigilant about information you expose to the internet. Edge computing can bring certain computing workloads inside the firewall, thereby lowering the risk of cybersecurity attacks.
• Cost efficiency: Once the investment in an edge AI solution has been made, there is no reliance on expensive subscription-based software pricing for external services, with the associated uncertainty of runaway costs and future cost of ownership.
• Performance: By reducing latency and having dedicated edge devices, performance levels can be guaranteed within the factory boundary.
Practicality of AI on the edge
‘AI on the edge’ refers to deploying AI models directly onto edge processors. Not all use cases for AI are feasible on the edge, but several are. For example, condition monitoring that is based on machine learning is ideally suited for edge devices, and there are many proven examples where this is already working. Similarly, image and video recognition technologies can usually be run on dedicated edge processors without reliance on the cloud.
In contrast, the most powerful AI transformers such as large language models (LLMs) will certainly need a cloud connection, at least for now. As AI transformers evolve, it is interesting that there may be viable use cases for small language models (SLMs) deployed on edge devices. SLM performance can approximate that of LLMs, provided they are tuned for a specific knowledge domain. It is likely that we will soon see SLM models deployed to mobile phones and other consumer devices.
Types of edge AI technologies
AI is an imprecise term that encompasses a group of loosely related technologies. Each of these AI technologies is similar in some ways, but at their heart they can be quite different in the way they work. AI models can be as diverse as neural networks, transformers (generative AI such as large and small language models), decision trees, or any combination of these. The proliferation and diversity of AI technologies is likely to increase and become even more complex going forward as industry adopts these technologies, while research delivers better ways to combine hybrid models with each other.
Typically, AI on the edge comprises the following:
• Data collection: From sensors and cameras, IoT devices and other real-time data sources.
• Data preprocessing: Cleaning and preparing of the data into a format suitable for the AI model.
• Model interfaces: Feeding the data into the AI model, validating the output, and processing the resulting actions in the form of alerts, changing setpoints or adjusted machine parameters.
• Analysis and visualisation: Allowing humans to easily view and interpret the performance of the AI model. This allows for further fine tuning of the model by closing the loop with human validation.
Edge hardware
Edge hardware is evolving rapidly with AI-centric processors from companies such as NVIDIA. This is clearly a space to watch. However, extreme processing power is not necessarily a prerequisite for edge AI; it is possible to deploy certain AI technologies onto a Raspberry Pi using a Docker container.
Edge software
Edge software architectures are also evolving.
Apart from relatively constrained applications like machine condition monitors or image recognition devices, deploying an advanced and bespoke AI model onto an edge device involves more than just installing software, then setting and forgetting it. You need to consider the entire future lifecycle involving model design, development, testing, deployment, and maintenance. In this regard, it is a good idea to discuss your requirements with your automation software vendor to evaluate if their platform is suited as a standard. Any gaps may need to be addressed with third-party solutions. Evaluating which is best for you may be a complex process, especially as the AI goalposts are always going to be shifting. Do not underestimate the complexity of implementing standardised infrastructure, processes and model architectures, and integrating these into your corporate IT environment where data management and cybersecurity are vital.
Organisational impact
I have written in the past about the importance of bringing together operations technology (OT) and information technology (IT) groups within typical manufacturing organisation structures. With any advanced AI solution, there is another key stakeholder group: the data scientists who will build the AI models. AI models generally need to be trained, tuned, optimised, tested and validated before being deployed into production. Once a model has been developed, deploying it to edge devices and monitoring its performance in the real world, together with the need to recycle models with experience gained, can be a complex process to manage. Who will the data scientists report to in your company?
Conclusion
Edge AI holds many benefits for industrial applications. However, the complexity of deploying these technologies can be high, especially when considering hardware standardisation, software architecture, model development, training and deployment, organisational factors, and finally maintaining the models. In this regard, it is a good idea to partner with your vendor to see how best to leverage their experience and proven platforms before deploying your AI to the edge.
About Gavin Halse
Gavin Halse is a chemical process engineer who has been involved in the manufacturing sector since mid-1980. He founded a software business in 1999 which grew to develop specialised applications for mining, energy and process manufacturing in several countries. Gavin is most interested in the effective use of IT in industrial environments and now consults part time to manufacturing and software companies around the effective use of IT to achieve business results.
© Technews Publishing (Pty) Ltd | All Rights Reserved
printer friendly version