Artificial Intelligence (AI) and the Internet of Things (IoT) are frequently seen as two distinct technology sectors.
IoT links devices, sensors, equipment and physical surroundings. AI makes use of the data, finds patterns and enhances choices.
The integration of these abilities is often summarized as AIoT – connected devices that emit data which can then be analyzed and utilized by AI models to support real-world activities.
The engineering problem is not just hooking up a model to a sensor. The real engineering problem is creating a system where data ingestion, processing, AI inference, connectivity, apps and human decision-making combine reliably.
Begin with the Problem, Not the Model
An initial failure point when designing AI:
Which AI model should we use?
For AIoT: At least these should be done:
What is our decision or operational problem?
What if you had a machine in a factory you wanted to connect to AI. Sensors in the machine might generate temperature, vibration, pressure, location, or other operational data. An AI system could find patterns in that data.
However the model alone does not alleviate the concern.
A useful system also needs to answer:
Where is data collected?
How frequently is it processed?
Where does inference happen?
What happens when connectivity is unavailable?
How is an alert delivered?
Who validates the result?
What action follows a prediction?
These questions bridge an AI experiment into an engineering solution.
Edge or Cloud?
Computational location is another critical architecture decision.
Cloud processing may offer a centralized infrastructure and facilitate analysis of numerous interconnected devices. Edge processing brings portions of the processing to the physical environment.
Edge computing may be helpful if latency, connectivity, bandwidth, or responsiveness at a local level are considerations. Cloud infrastructure may be helpful for central analysis, storage, and more-demanding workloads.
There is no universal answer.
An AIoT architecture may use both. The device may do some preliminary processing locally, send data to a platform, and use cloud infrastructure to do more extensive processing or models.
The appropriate design depends on the application.
Data Quality Still Matters
AI does not eliminate traditional data challenges.
A connected system produces a huge amount of information and can still go wrong if the information is incomplete, inaccurate, has no context, is mis-labelled or just plain wrong.
Developers should consider:
Sensor trustworthiness: Are measurements being reproducibly measured?
Data context: is the system aware of when and where the data is created?
Data quality: How does the device deal with missing data and outliers? - Workflow: What is the workflow of the model?
– Training data: Is it also realistic?
- Feedback: How will the 'day-to-day' stuff be used to assess?
It might be has importance as the model chosen.
Human Oversight Is Within the Framework
AIoT systems can be used where the decision results in physical ramifications.
For example, even if a model detects a rare machine state, an engineer may still want to know if it is a true fault.
So, human oversight may be incorporated into system design.
It's not always about you taking the people out. In a lot of manufacturing circumstances, that's about giving the people better data at the time when they need it to make a decision.
Sometimes Integration Is Difficult Than the AI
A prototype that functions smoothly in a laboratory setting may have difficulty functioning in a deployed environment.
New systems can use a variety of protocols, databases, data formats and processing procedures. A production AIoT system might also need to integrate:
sensors and connected devices
gateways and networks
edge computing
cloud infrastructure
databases
machine-learning models
APIs
dashboards
enterprise applications
monitoring systems
So AIoT by its nature is a multidisciplinary field. Software engineering, data engineering, networking, machine learning, embedded systems, cybersecurity and expertise within the relevant domain can all be equally critical for implementation of the technology.
Design for Failure
Failure to failure should be part of the model as well as failure to normal operation.
What happens when a sensor stops reporting?
What happens when network connectivity disappears?
So, what do you do When Your AI Model Gets Bullied Data?
What if a prediction does not match the observation of the operator?
For important components, teams can define:
Normal behavior
Expected failure conditions
Detection mechanisms
Fallback behavior
Recovery procedures
Monitoring requirements
Early consideration of these situations can reduce the complexity of the systems.
Measure the Result, Not Only Model Performance
Although model accuracy may be a gauge, it may not be the determinant of whether an AIoT system succeeds.
Teams might also consider:
response time
false-alert frequency
system availability
data quality
operational efficiency
maintenance outcomes
human review rates
integration reliability
A bad model that generates useless alerts might be worse than no model at all. It might be better to have a simple model that can sure up a key workflow.
Learning From Real-World AIoT Experience
The AIoT is reshaping industry & enterprise The AIoT is already being implemented across an industrial landscape and enterprise level, so its worth having some on-the-ground expertise.
Technical presentations, research papers, system demos, and case studies may introduce developers to issues hard to replicate in isolated prototypes.
The Aperture Ventures Summit speaker information lists a global event forum dedicated to real world AI, IoT, and AIoT discussions of systems and industries. Featuring technical presentations and research as well as system displays.
Aperture Venture Studio - Aperture Venture Studio
Due to our team’s recent breakthroughs in physical AI and AIoT, this website and the websites of all 64 portfolio companies are being substantially upgraded
apertureventurestudio.com
The general takeaway is it doesn't take a single incident for us to know that it can be worthwhile to bring software concepts and AI concepts together with the physical homes in which the systems will ultimately be used.
The Bigger Engineering Question
AIoT is not simply "AI plus IoT."
It is a systems-engineering problem.
What really matter are questions such as how is data transported through the system, where are decisions made, how are failures and errors addressed, how do humans interact with the results, and does the technology solve a really important operational problem?
As AI continues to permeate all connected physical systems, developers will have to cross-cut.
The next generation of better AIoT solutions will depend on more than just better models, it will depend on robust architecture, dependable data, proper integration, smart monitoring, and a well understood problem space.
Top comments (0)