If you say a lot of developers use to think the main thing after saying an 'AIoT' would be the AI component it self:
Which should AI model be used?
How it will infer, operate in its respective operating system or on a given piece of hardware? Where will they deploy? And last, which way would they improve their inferences?
The above listed questions are relevant but, in real world Industrial AIoT systems, there is another piece missing and that is the way we are to pipe consistent and relevant information from the physical world into the "Intelligence layer". A highly efficient AI will only be as good as the input data is contextualized, consistent or relevant to the ongoing operation.
Simplified AIoT Architecture
The layered concept of industrial AIoT architecture can be described as follows:
``text
Physical Assets
Sensors / Devices
IoT Infrastructure
Data Pipelines
Operational Data
AI / Analytics
Applications
Operational Decisions
`
Each layer plays an essential part; Sensors/Devices would capture real world information, which then get managed by the IoT infrastructure and piped through Data Pipelines to the AI/Analytics layer that tries to extract insights or make predictions. The prediction or insights would then be consumed and interpreted by Applications for human or machine consumption which would translate into Operational Decisions and further actions or behaviors of the physical assets being tracked.
Context Is an Data Engineering problem
Consider a given machine that produces temperature readings:
`text
09:00 72C
09:05 75C
09:10 81C
09:15 86C
`
This stream of data, by itself, might suggest something is not right. However without context this data may be misleading.
Was the machine heavily utilized then?
Was production process changed?
Was this machine recently serviced or maintained?
Is the observed reading standard for that operating mode?
Were there other changes in sensor values detected?
These questions imply that real-world data collected has needs far beyond that of simple temporal readings of machine variables. For Industrial AIoT, you may need to include such information as asset ID, its location, its state (e.g. On/off/error/under maintenance), its production context, its maintenance history, prevailing ambient conditions, product stock-relevant context and event histories to derive appropriate values from them. With all of this taken into account the 'intelligence layer' really gains its proper importance.
From events to relations
We could view simpler IoT scenarios with event notifications:
`text
Machine A vibration went high
Machine A temp went high
Production Line 2 output went down
`
The enhanced version of such system could track these relations over time:
`text
Machine A
Vibration on the rise
+
Temp is going up
+
Production yield dropped
+
Standard operating practice based on previous issue investigations
Likely the start of an equipment performance failure trend
``
The goal is obviously not to assume cause with every coincidence but to equip the operator to identify and investigate any correlation worthy of human attention.
AIoT infrastructure relies on sound architecture design:
This has a range of implications for the developer:
1. The question of identity
Each piece of physical asset (machine, device, vehicle or even person) has a true identity which must be captured accurately and consistently. An unable-to-compare signal that can't be reliably linked to its unique entity will pose difficulty down the chain.
2. Identity of Time
Having events properly time-stamped is one major element. For relationships you have to capture precisely how each event or data point was timed to the other.
3. Space relevance
Location often represents crucial data and it could include data from vehicle movement and vehicle identification as well. Understanding where the asset was operating at the point of a sensor reading dramatically improves context.
4. Quality control in data itself
Industrial environment usually produce erroneous (i.e. Missing, duplicated, stale or deviant) readings. It is essential that data pipeline includes data validation and anomaly monitoring mechanisms
5. Interaction between separate system functionalities
AIoT projects will frequently interconnect with production processes, maintenance schedules, and company-wise ERP.
Applications have crucial functionality
Building good analytical systems in AIoT will do no good unless people in operations can find and consume of them effectively and utilize the outcome of AI models. Production ready AIoT application needs to link intelligences to the workflows:
Detection -> Investigation -> Decision -> Action -> Feedback
A feedback component is essential, as actions derived from AI and their subsequent results will only improve further operations with the feedback loop.
Real world problems must define AIoT architecture
A real problem in operations should guide any industrial AIoT architecture. These can range from prediction of equipment failure, to assets monitoring, optimizating production and processes, or enhance worker safety etc. (Please check Aperture Venture Studio's AIoT model for instance). All focus is put on solving a physical world industry problem with the assistance of AI and IoT and packaging these solutions as deployable system platforms.
Key message for developers
There is really no way to equate AIoT with simply 'IoT + an AI model'. A system built for production readiness in an industry necessitates collaboration across physical setup, connectivity, data architecture, contextual information gathering, intelligence and application layers. The strongest AIoT solutions begin by identifying what has to be determined as an operating decision, with the entire design following from this information inwards. For More Info Visit: apertureventurestudio.com
Top comments (0)