Having surpassed peak marketing hype, Internet of Things (IoT) technologies are now an established building block, both in the business environment and in industrial applications. An illustration of technology maturity is how combinations with other technologies enable new operational solutions. Gone are the days of deploying exploratory pilots. Now, product managers focus less on connectivity and more on packaging connectivity, data management, and security for operational use. This involves integrating cloud infrastructure services, AI and ML analytics, and visualization dashboards to design end-to-end IoT systems with operational users in mind.
Emerging Requirements for AI in IoT Systems
A straightforward way to view the combination of AI/ML and IoT treats the IoT system as a means of channeling data into intelligent monitoring or decision-making application. In a sense, the IoT and AI/ML technologies are separate and joined via a data delivery API. While this approach solves a technical challenge, it overlooks an emerging set of future AI/ML systems requirements.
As an illustration, consider a condition monitoring application that uses machine data to observe the ‘health’ of the engine and drive-train in a public transport bus. A predictive maintenance component in the application might then signal a developing fault. Assume that the fault notification would require the bus to be taken out of service, potentially inconveniencing the traveling public on an infrequently served country route. These travelers might have to wait a few hours or even to the following day for the next scheduled service. Should the depot manager take the bus out of service?
It would help if the fault detection algorithm had a near-perfect history of detecting major faults. That would make the decision straightforward. Alternatively, the depot manager might have greater confidence in the diagnosis if the AI/ML system explained its reasoning.
This use case illustrates a couple of requirements for future AI/ML and IoT systems. One involves tracking and reporting on an application’s performance record over time. Another is to explain its reasoning and do so in terms that engineering technicians and operational managers can understand. There is also a need to differentiate reported faults between those associated with the machine being monitored and others caused by sensor failures and data transmission errors.
AI for IoT Architecture Models
The condition monitoring and predictive maintenance example provide a framework for thinking about the architecture of AI/ML in IoT systems. The simplest arrangement contains two building blocks. One is the IoT portion, which is responsible for collecting and supplying data. The second contains the application that uses AI/ML functionality for reasoning purposes.
Looking beyond the basic configuration, one architectural enhancement is to integrate AI/ML capabilities within the IoT domain. This could involve a set of background processes to check whether a data stream is valid or possibly corrupted due to a faulty sensor or intermittent data transmission. This arrangement has the benefit of making data quality and sensor integrity check a common function. The same algorithms, using tailored configuration parameters, could be reused as more sensors and data sources are added to the system over time. This arrangement offloads data integrity checks from the AI/ML intelligence associated with the target use-case.
The next architectural progression involves a bi-directional relationship between application and IoT domains. This arrangement would make greater use of semantic descriptors and automated fault-tree analysis techniques to enable higher levels of reasoning. When an AI/ML system reports a fault, an explainable-AI component could create a digital twin representation of the system by automatically detecting its components. It could also reason about possible trigger conditions by querying the underlying sensors and data communication paths, applying fault-tree analysis techniques to the digital twin. Other forms of checks are also possible. An example query might request data about the supplier, a device’s firmware build and, the most recent security update. This could be cross-referenced with industrial databases, as occurs with software security patches, for example.
Industry Frameworks and Standardization
The different architectural models for combining AI/ML with IoT systems highlight a progression of more sophisticated capabilities in sensors, connected machines, and IoT platform components. Early AI/ML products and services need only address the first architectural model. As deployments become business-critical, there will be a need for scalable and trustworthy systems. This would create a market tipping point for improved architectural models and more sophisticated components capable of supporting advanced reasoning forms.
An added requirement in the case of distributed and multi-application systems are interoperability. This applies to systems that combine technologies and components supplied by different vendors. It also involves applications that work across departmental and corporate boundaries. An effective method to enable interoperability is open standardization, as evidenced by mobile and Internet industries’ success. Standardization of AI for IoT is an embryonic topic, especially related to the increasing levels of complexity in the deployment architectures described above.
To this end, ETSI, the European ICT standardization body, recently released two reports from a study on AI implications for IoT. The first technical report analyzes the requirements for new ‘AI for IoT’ capabilities based on common issues to a set of representative use cases. The second report illustrates the practical implications for IoT systems by implementing three ‘AI for IoT’ proofs of concept that build on the oneM2M standard for IoT systems. These two studies lay the foundations for strategic product management of AI and IoT systems.
