Traditional sensors capture data and send it elsewhere for storage or interpretation. That becomes less attractive when continuous vibration, high-resolution video, chemical readings, or other signals produce far more information than a Cloud application needs.
AI-enabled sensors interpret data where it is created. Instead of transmitting every reading and leaving the Cloud to decide what matters, they transmit information that already has meaning attached to it.
A machine-monitoring system may report that a bearing shows signs of an outer-race fault, that the condition is serious and provide a confidence indicator. A vision system may report that a queue has exceeded a defined limit or that somebody has entered a restricted area, rather than upload hours of footage.
This is contextualised information. The sensor output is interpreted in relation to what is normal or expected, where the sensor is located, what time it is or what the system is trying to detect. The Cloud receives a useful observation rather than numbers or pixels that still need interpretation.
This differs from data compression. Compression reduces the number of bits used to represent the same raw data. Edge AI decides which parts of the raw data are worth communicating at all.
What are AI-enabled sensors?
An AI enabled Edge sensor can combine a sensing element with local processing, while a wider Edge AI system may also include smart cameras, microcontrollers, or nearby processing nodes. What matters is where the data is interpreted: close to the source rather than requiring a round trip to a Cloud data centre.
This can make systems faster and more self-reliant where connectivity is patchy, expensive, or unnecessary for routine local decisions. Selected information can then be sent onward for fleet-level visibility, historical analysis, or model management.
An Edge intelligence design needs to consider three closely linked elements: hardware, local analytics, and communications strategy. Intelligence may sit in a sensor package, AI-enabled camera, microcontroller with a neural-processing unit, or industrial gateway. The choice depends on processing demand, power, response time, environment, and the amount of data produced.
Local analytics turns a raw signal into an interpretation. A device may filter input, extract features, and apply an ML model to detect anomalies, classify faults, identify objects, track movement, or combine sensor sources. Models may need to be quantised or otherwise optimised to fit available memory, processing capability, and energy budgets.
The device can send summaries, alerts, diagnostic indicators, and selected evidence rather than every reading. Routine data may remain at the Edge temporarily or be discarded. The Cloud still has an important role, but need not sit in the critical path of every real-time decision.
TinyML can be enough for a narrow task and modest signal: recognising a machine state from vibration, identifying an acoustic event, detecting a gesture, or identifying an unusual temperature pattern. A microcontroller can sample a sensor, run a compact model, and transmit an event within a tightly constrained power budget.
High-resolution images, several sensor streams or frequent low-latency decisions may require an embedded AI accelerator or neural-processing unit. Inference close to the sensor minimises data movement. An MCU offers flexibility for several simple sensors, while a gateway provides more memory and compute for sensor fusion and larger models. The best choice meets the required response time and detection performance within the available power, memory and processing budget.
Edge AI for machine, video, and environmental monitoring
Rotating equipment can generate unwieldy volumes of data. A motor, pump, gearbox, or fan may have a three-axis accelerometer and temperature sensor feeding an Edge node nearby. That node may be a microcontroller with on-chip neural-processing capability or an industrial gateway if the analysis is more demanding. Vibration data is dealt with close to the machine rather than sent in full to a distant data centre before anyone knows whether it matters.
The Edge device gathers vibration readings in short sections. It can filter the waveform, calculate features such as RMS level, kurtosis, and spectral peaks, and use these features or the underlying signal as input to an anomaly detector or a trained one-dimensional convolutional neural network. Rather than report only vibration, it can assess normal operation, imbalance, misalignment, looseness, or possible bearing damage.
Contextual output might include asset ID, identified fault, severity, diagnostic status, and detection time. During normal operation, a machine may send occasional health updates. If it identifies a potentially serious fault, it can retain a waveform segment from immediately before and after the event, then upload it for engineering review.
In one published evaluation of an Edge maintenance approach, transmitting summaries, alerts, and selected waveform extracts substantially reduced the volume of data sent compared with continuously transmitting raw vibration data, while retaining evidence needed for maintenance decisions.
Video makes the case even more clearly. A single IP camera can produce a substantial continuous stream; across a shop, warehouse, factory, or public site, sending, storing, and analysing every frame can become expensive. With Edge AI, a camera or local vision appliance can interpret images where they are captured and produce events or aggregated metrics rather than continuously transmitting video.
The output might report 12 entries during 15 minutes, a queue longer than five people, a vehicle in a restricted area or a worker without required PPE. Where detailed evidence is needed, selected footage can be retained locally or uploaded in response to an event rather than continuously streaming all video.
Environmental sensing follows the same principle. Continuous measurements from gas, wind, chemical, acoustic, or RF sensors can create a substantial data burden, although most readings describe normal conditions. For example, a gas-monitoring system could combine chemical measurements with wind direction, wind speed, recent readings, and baseline conditions to identify a possible methane plume, then report the event and relevant contextual information rather than every sensor value.
Design, data retention, and measurement
An Edge AI design is not automatically lower power because it transmits less data. Radio transmission can be expensive in energy terms, particularly on wireless devices, but sensor sampling, buffers, and inference also consume power. The design task is to balance sensing, processing, and communications.
A model need not run continuously. A low-power device may wake periodically, use a simple threshold to identify possible activity, then run a more demanding model only when needed. Designers can reduce energy use by lowering sample rate where appropriate, shortening the input window, quantising the model, and transmitting exceptions rather than routine measurements.
An overly small model may save power but create too many missed events or false alarms, while a larger model may consume more memory and energy than the product budget allows. The target is dependable system performance under real operating conditions, not maximum model accuracy in isolation.
Sensor fusion can improve the trade-off. Vibration combined with temperature, motor current, and operating speed can provide a more confident diagnosis. A vision system may add LiDAR, UWB, or another ranging technology where poor lighting, dust, reflection, or occlusion makes image-only detection unreliable. Combining sensors adds processing work but can reduce false alarms and improve the information sent upstream.
Raw data remains important for fault investigation, calibration, sensor validation, model-drift checks, and training later model versions. Many systems use a rolling local buffer: they transmit contextual messages under normal conditions, then upload relevant raw-data segments when the inference is uncertain, an anomaly crosses a threshold, an operator asks for evidence, or the system encounters a new condition.
A useful way to evaluate an Edge AI architecture is to compare raw sensor and event-data rates, calculate bytes transmitted per device per day, and identify reductions in communications and Cloud-storage demand. Other useful measures include:
- Inference latency and time from detection to action
- Detection accuracy, missed-event rate and false-alarm rate
- Energy used per inference or contextual event
- Quantity of raw evidence retained or uploaded per anomaly
- Network resilience during a connectivity outage
- Privacy and data-governance benefits of keeping raw information local
AI-enabled sensors can decide what their measurements mean before sending anything onward. That can reduce Cloud traffic, improve response time, protect sensitive raw data, and provide information users can act on rather than another stream of measurements.
For industrial engineers, the question does not stop at the Cloud connection. If more devices transmit compact condition, quality, and diagnostic events instead of continuous raw data, what does that mean for bandwidth, diagnostics, and information flow across the industrial network? That is where technologies such as IO-Link become part of the discussion: not simply as a connection for sensors, but as a way to carry richer device-level information through the automation system.
Explore Farnell’s AI solutions
Farnell provides a range of AI development tools, hardware, and engineering resources to help engineers bring intelligent applications to life. Explore DevKit HQ and other AI resources to support your next design.
