KI-PREDICT to use distributed AI in sensor-based process

Seven partners are taking part in this project, one being the Fraunhofer Institute for Integrated Circuits IIS. It is developing a sensor interface application-specific integrated circuit (ASIC). Fine-tuned for sensors that monitor the condition of equipment and help control processes in real time, this ASIC will enable energy-efficient feature extraction and local, on-sensor signal processing.

Microelectronics, when combined with sensors and embedded software, can capture and process data generated by industrial plants. This information can serve to digitalise manufacturing processes and operational workflows for Industry 4.0.

There is a catch, though: Today’s electronic data acquisition and signal processing systems have yet to be optimised for this application, so they are fairly pricey in comparison to the components they are supposed to monitor.

The cost, space requirements and power consumption of digital signal processors (DSP) and field-programmable gate arrays (FPGA) are too high for many frequently used Industry 4.0 sensors, so manufacturers seeking to digitalize their processes cannot simply swap sensors in a one-to-one replacement scheme.

The KI-Predict project addresses this problem in a holistic manner. The idea is to combine new AI methods with optimized, integrated hardware to provide intelligent process monitoring with local signal processing and feature extraction. This new quality of on-sensor data processing enables reliable, decentralized analysis and prediction with a defined, low latency. Researchers engaged in this project are developing a dovetailed hardware and software architecture to this end. It will fuse, reduce and assess sensor-related data, and interpret anomalies to detect faulty sensors.

Aside from executing the usual functions, such as capturing digital data on electrical current, position, vibration, acoustics, pressure, force and temperature, it will also provide the machine learning (ML) capabilities needed to process and reduce data locally.

The interface is able to detect features even in high-frequency sensor signals in an energy-efficient way. It delivers these features to the control level for sensor data fusion, or uses them directly for classification, clustering or anomaly detection. Aggregated features are extracted from the data stream at the sensor to reduce the amount of data. This reduction is necessary to connect to standard industrial interfaces and networks.

These features can also serve a higher purpose: When forwarded to process control or enterprise resource planning (ERP) software, they help determine the equipment’s status, assess product quality and track trends using sophisticated AI and ML methods. This hardware is not tailored to any specific applications, so automated routines can ‘train’ it for new use cases.