The world of artificial intelligence consists of models, applications, and the cloud. But many AI use cases actually start elsewhere: In the physical world.
Machines produce data. Sensors read environments. Assets transfer between locations. Equipment transforms state.
To make those events valuable pieces of information requires more than an AI model. It requires an architectural structure connecting devices, data, processing, machine learning, applications, and decisions.
This is the concept of AIoT, or Artificial Intelligence of Things.
A (very) simplified AIoT architecture would consist of:
Physical Environment → Sensors → Connectivity → Data Processing → AI/ML → Application → Decision → Action
1. Sensors Create the Data Layer
Sensors form the foundation of many AIoT applications.
Depending on the use case, the physical devices may collect different types of data:
- Temperature
- Equipment state
- Location
- Movement
- Environmental data
- Operational activity
The information collected by the sensors is commonly a stream of measurements or events. In its raw form, the data usually lacks meaningful context for decision-making.
The challenge, therefore, is to collect relevant information with sufficient context.
2. Connectivity Moves Data Between Layers
The next step is to move the data between devices and processing systems.
The choice of connectivity may differ depending on the environment and other variables. Range, bandwidth, latency, reliability, power consumption, and operating conditions can determine which approach to select.
A simple architecture would take the form of:
Device → Connectivity → Processing System
Connectivity is the bridge providing access to the information captured by sensors.
3. Data Processing Turns Raw Into Contextual
The raw sensor data may require additional processing to transform it into valuable information.
A pipeline can include steps such as:
Ingestion → Validation → Filtering → Transformation → Storage → Analysis
For example, a machine may collect temperature data every few seconds. On its own, an isolated reading provides limited insight. However, a time series can indicate whether the temperature remains stable, increases linearly, or differs significantly from the expected pattern.
The historical context, therefore, can help establish what is normal and what is not. This is an important consideration since determining normal and abnormal behavior is a critical step in many AIoT applications.
This is why data engineering plays an essential role in AIoT systems.
4. AI and Machine Learning Analyze Patterns
With the data ready, the next step is to apply AI and machine learning to analyze patterns.
Depending on the use case, a model can help perform multiple tasks, such as:
- Anomaly detection
- Classification
- Forecasting
- Pattern recognition
- Predictive analysis
As an example, a machine learning model can analyze historical data related to a piece of equipment and recognize behavior that differs significantly from the norm.
However, the model is only one component. Other factors, such as the quality of the data, the selection of the algorithm, the deployment environment, operations, and model integration with other systems, are also important considerations.
5. Edge and Cloud Processing
Many AIoT applications involve both edge and cloud processing.
Processing data at the edge can reduce latency, provide local processing power, decrease bandwidth needs, and maintain operations during connectivity loss.
Meanwhile, the cloud can offer additional processing power and storage space while hosting analytical functions, applications, and AI/ML models.
A practical architecture, therefore, would resemble:
Sensors → Edge Processing → Connectivity → Cloud/Data Platform → AI/ML → Application
Depending on the requirements, the workload can be distributed between edge and cloud servers.
6. Applications Connect AI to People
An isolated AI model typically cannot directly impact operations. To affect the real world, the results of the analysis need to be processed by applications.
The applications can range from simple dashboards to complex sets of tools providing actionable insight:
- Dashboards
- Alerts
- Monitoring tools
- Reports
- Workflows
For example, if an AI model recognizes abnormal activity related to a machine, an application can relay this finding to the maintenance team to investigate the cause.
By connecting AI results to an operational process, applications create a meaningful link between machines, people, and the physical environment.
7. From Prediction to Execution
The end to end process of an AIoT application can be summarized in a few words:
Sense → Connect → Process → Analyze → Decide → Act
The actual execution of the final step, however, can take multiple forms.
A person can investigate an alert, a set of tools can generate a ticket for further review, or a machine can autonomously execute specific instructions.
The most important aspect of an AIoT system is that it provides a data-driven basis for action in the physical world.
Designing AIoT Around the Problem
One of the most common mistakes in connected systems design is to reverse the process: Focusing on the tools rather than the problem.
Before selecting the sensors, connectivity means, databases, or AI models, it is helpful to clearly define the question first:
- What physical event needs to be measured?
- What information is needed?
- How much data needs to be collected?
- Where can/will it be processed?
- What pattern is important?
- Who (or what) will perform the action?
Only after clearly defining the scenario can the technical implementation begin.
The Big Picture
AIoT connects multiple fields:
Devices + Connectivity + Data Engineering + AI/ML + Applications
Each plays an essential role. While sensors observe the physical world and connectivity moves the information, data processing adds valuable context to the mix. AI models detect patterns, while applications present results in a meaningful format. Decisions, therefore, can once again affect the physical environment.
Physical World → Data → Processing → Intelligence → Decision → Action
This is the fundamental architecture of many AIoT applications.
Top comments (0)