Industrial Artificial Intelligence of Things (AIoT) is the combination of connected devices, data pipelines, and analytics that support operational decision-making. While a monitoring system generally captures data for later analysis, an AIoT solution actively performs pattern recognition and drives responses based on identified trends.
However, there is more to an effective AIoT solution than selecting the right sensors and connecting them to an analytics engine. The data acquisition process, system integration, data processing, security, and even the supporting workflows significantly impact the design of an industrial AIoT solution.
The following article details critical architectural choices and considerations when implementing industrial AIoT.
- Define the Industrial Use Case
Before identifying the sensors to use or the analytic model to implement, an industrial use case needs to be defined. Examples of industrial applications that benefit from AIoT solutions include:
• Anomaly detection in rotating machinery
• Tracking work-in-progress
• Increasing visibility into material logistics
• Detecting abnormal energy consumption
• Improving inventory awareness
Depending on the application, specific requirements will dictate aspects of the data acquisition, processing, integration, and presentation layers.
For example, a rotating machinery monitoring solution typically processes data at higher sampling rates than an inventory tracking system, which in turn has more stringent requirements for identification accuracy. Similarly, the choice of identification technology (RFID, vision, BLE beacons) may be dictated by the application’s requirements for read range or accuracy.
As such, the use case should dictate the architecture rather than the other way around.
- Focus on Data Acquisition
Most often, the data acquisition layer is the foundation of an AIoT system, aggregating information from sensors, RFID readers, PLCs, or other industrial equipment. Wherever applicable, data acquisition components should be hardened and support the operating environment (indoor, outdoor, etc).
A typical industrial data acquisition layer contains a mix of sensors, RFID readers, programmable logic controllers, industrial gateways, machine interfaces, and other devices. While some equipment may have built-in networking and data outputs, others may need to be connected through field jacks or other means.
A critical consideration when designing the data acquisition layer is the use case and the requirement for data normalization across heterogeneous systems. In practice, most industrial equipment communicates using manufacturer-specific data formats and conventions.
As such, a practical data acquisition layer should be able to normalize the data it receives while retaining relevant metadata (such as timestamps, location, or units of measurement).
Another critical consideration is data validation and ensuring the accuracy of individual records. Wherever possible, data acquisition components should ensure that the data received follows a logical sequence (no missing records) and does not contain obvious anomalies.
While the data acquisition layer rarely processes data, it may contain rules and logic to identify and flag individual records for additional review.
- Determine What Processing Should Happen at the Edge
While many AIoT designs focus on the cloud, some applications may benefit from processing data closer to the source. Edge processing is generally useful in applications that have strict latency requirements, require offline operation, or have limited connectivity.
For example, an edge server or gateway may be used to process sensor data and identify specific events (such as an abnormal vibration) and raise an alert. Meanwhile, the cloud may be used to aggregate data from multiple facilities, perform analytics across multiple data points, or train machine learning models.
Most practical AIoT designs, however, use a combination of cloud and edge processing, with individual requirements defined by the use case. In practice, such designs follow a similar pattern:
Devices collect individual measurements or events
Edge servers perform data filtering, validation, and processing
The communication layer sends individual data points or aggregated records to the cloud
Cloud data warehouses store the information for processing and analysis
Applications present relevant insights to operational analysts
The division of responsibilities between edge and cloud depends on the specific requirements of the use case.
- Use AI within a Broader Analytics Context
An AI model is generally only a small part of a greater analytics pipeline. For example, a predictive maintenance solution based on AI requires training data that includes sensor readings, equipment states, maintenance logs, and details of actual equipment failures.
More importantly, a practical AI implementation involves continuous data validation, model training, and analysis. As such, an AI model is generally only one element of a broader analytics pipeline.
Another critical consideration when using AI in an industrial setting is distinguishing between an anomaly and a failure. While an unusual sensor reading may indicate an impending failure, such a conclusion can only be made in context.
As such, AI models should be used to identify irregularities, but human expertise is generally required to assess whether such irregularities translate into imminent failures.
For critical systems, such operational judgements should be formally audited to ensure conformity with industry standards and regulatory requirements.
- Ensure Adequate System Integration
Operational insights are only valuable when they are presented to the appropriate decision-makers. As such, an AIoT system should be integrated with relevant operational and enterprise systems that allow for timely contextual response and decision-making.
The choice of systems to integrate with depends on the use case. For example, an inventory tracking system may need to integrate with enterprise resource planning (ERP) software, warehouse management systems (WMS), or manufacturing execution systems (MES).
A rotating machinery monitoring application, on the other hand, may need to interface with maintenance management software to schedule and document repairs. Wherever possible, system integration should consider data ownership, availability, permissions, and potential operational impact.
For example, an inventory application may be unable to update ERP records directly if there is a possibility of erroneous record updates. Instead, the AIoT application may raise an exception that needs to be reviewed by a human operator before any changes to the ERP are made.
In short, system integration should consider potential data conflicts and the downstream impact of data alterations.
- Design With Security and Scalability in Mind
Most often, an industrial AIoT system needs to interface with legacy equipment, modern connected devices, and enterprise software. As such, the security architecture should target all layers with consideration for the requirements of each.
Operational technology (OT) networks, for example, often have specific availability and safety requirements, while connected devices may need assistance with over-the-air (OTA) firmware upgrades and secure credential exchange. Considerations at the network level often include firewalls, segmentation, identity, and least privilege.
Another critical consideration when designing industrial AIoT systems is scalability. Solutions designed for a single production line or facility should take architecture choices that allow expansion without disrupting operations.
Practical considerations for long-term scalability include designing data ingestion and analytics pipelines around standardized formats, using generic integration patterns where possible, and deploying standardized device management practices.
- Evaluate Business Value Beyond System Metrics
While system-level metrics such as data accuracy or latency are important, they are generally insufficient for determining business value.
As such, an appropriate evaluation framework should include both system performance metrics and business metrics. The following table contains examples of system performance metrics and their potential impact on business operations.
System performance metric Potential business impact
Number of data records captured Are the required data points captured?
Alert precision Are the alerts actionable?
Data processing latency Is the information available when needed?
System availability Can the service be relied upon?
Reduction in unplanned downtime Has the solution fulfilled its mission?
Inventory accuracy improvements Has the solution improved inventory accuracy?
With few exceptions, most business metrics require some baseline data before implementation to measure improvements accurately. As such, it is often a good idea to establish baseline metrics before initiating a pilot implementation.
Pilot projects should also have clearly defined success criteria, particularly in terms of cost-benefit analysis.
Visit : "Aperture Venture Studio"
- Conclusion
An effective industrial AIoT system connects operations to analysis through reliable data processing, appropriate data analytics, and effective system integration. Most critically, an effective AIoT solution starts with defining a specific industrial need and targeting it with appropriate technology.
For teams that want to explore how industrial AIoT concepts can evolve into practical technology ventures, Aperture Venture Studio outlines several considerations for building AIoT technology startups that focus on industrial applications.
The fundamental engineering principle behind industrial AIoT solutions is simple – design a system that reliably supports an operational decision, then expand outward as additional evidence and requirements dictate.
