Access Official Technical PDF

Share Summary:LinkedInEmail

AI-in-sensor processing is an approach to edge AI that performs computation within or alongside a sensor, rather than sending raw sensor data to a separate processor for analysis. The concept addresses a practical challenge for industrial automation: sensors can produce continuous streams of analog data, while conventional digital AI systems typically require that data to be converted and transported before processing. Moving analysis closer to the source may help reduce data movement and enable decisions to be made nearer to the process being monitored.

A 2019 Frost & Sullivan Technology Innovation Award report described AIStorm’s system-on-chip approach to this challenge. Its example is useful for understanding how charge-domain processing works, what it may offer in edge applications, and which design considerations matter when evaluating sensor-level intelligence.

Why edge AI creates a sensor-data challenge

Edge computing processes information near where it is generated, such as on a device, local computer, or nearby server, instead of relying entirely on a centralized data center. For automation systems, this can place analysis closer to machines, instruments, cameras, or other connected equipment. AI can extend this model by enabling devices to interpret data locally and act on relevant conditions.

However, the path from a physical signal to an AI decision can involve several stages. A sensor produces an analog signal; the signal is digitized; data is transferred to a processor; and the processor runs an algorithm or model. When a system must continuously digitize and process a large data stream, conversion and data movement can contribute to latency and power consumption. The report identified these as challenges for GPU- and NPU-based edge systems, particularly in portable or power-constrained devices.

Another issue is that conventional processing may handle far more input data than is needed for a particular decision. If the system cannot identify useful information until after digitization and transfer, it may spend resources processing data that does not affect the outcome. This matters in applications that monitor video, audio, vibration, motion, or other signals continuously.

How AI-in-sensor processing uses charge-domain computation

The AIStorm technology described in the report uses charge-domain processing. Rather than first converting the complete sensor signal into digital data for an external AI processor, switched-charge circuits control the movement of electrons between storage elements in the chip. The report describes this as a way to perform mathematical operations on analog sensor data at the sensor, with digital communication focused on the final results.

The central architectural idea is to combine sensing and processing more closely. AIStorm’s AI-in-Sensor system-on-chip was described as integrating with sensor types including imagers, microphones, and waveform sensors for vibration or motion. The chip concept also brought together functions such as memory, power management, and communications. The goal was to reduce the need to send large volumes of raw data to a separate processor before a decision could be made.

The report also described an event-driven operating model for imaging and audio. Instead of continuously polling for a condition, an AI-based trigger can signal when an event warrants attention. In principle, this can limit unnecessary processing between relevant events. The specific power and performance results depend on the application and implementation; figures in the award report refer to the technology as described in 2019 and should not be treated as universal benchmarks.

Potential implications for industrial automation

In industrial environments, the broader relevance of AI-in-sensor processing is its potential to analyze signals closer to their point of origin. Applications identified in the report included industrial automation, security and access control, automotive systems, and healthcare. For automation teams, sensor-level processing could be considered where an application depends on timely interpretation of image, acoustic, motion, or vibration data and where transferring every raw measurement is undesirable.

Possible areas for evaluation include:

  • Machine and equipment monitoring: Local analysis of vibration or motion signals may help identify selected events without forwarding an uninterrupted stream for external processing.
  • Vision at the edge: Processing within an imaging module may support applications that need to interpret image data near the camera, including access or occupancy sensing.
  • Event-driven monitoring: Trigger-based operation may be relevant when systems need to respond to specific conditions while avoiding continuous polling.
  • Constrained devices: Integrating sensing and inference may be worth assessing where device size, power use, or communication capacity limits the use of a separate processor.

These are application directions, not guarantees of suitability. Engineers and system integrators still need to assess the sensor signal, target decision, required response time, environmental conditions, and integration requirements for each deployment.

Design considerations for system integrators

Moving computation into a sensor changes the system architecture; it does not remove the need for engineering validation. A useful evaluation starts with the decision the device must make and the data necessary to make it. Teams should determine whether local processing can preserve the information needed by the control or monitoring system, and how results will be communicated to downstream equipment.

Other considerations include the supported sensor and algorithm workflow, compatibility with existing embedded platforms, and how the device will be configured and maintained. The report noted that AIStorm offered a software development kit compatible with TensorFlow Lite and programmable processors. Those details describe the offering at the time of the 2019 report; current availability and capabilities require separate confirmation.

It is also important to distinguish sensor-level inference from a general-purpose edge computer. An integrated sensor processor may be designed for particular inputs and tasks, while a separate CPU, GPU, or NPU can serve broader processing needs. The appropriate architecture depends on the application’s requirements, including data volume, response time, power constraints, and the level of flexibility needed.

Conclusion: move analysis closer when the application warrants it

AI-in-sensor processing offers an alternative to digitizing and transferring an entire sensor stream before analysis. By applying charge-domain computation at the sensor, the approach described in the 2019 report aimed to identify relevant information locally and support event-driven operation. For industrial automation decision-makers, the practical takeaway is to evaluate where sensor-level intelligence could reduce unnecessary data movement or bring a decision closer to the process—while validating integration, performance, and application fit against the actual system requirements.

Access Official Technical PDF

About The Author