In their 'AIQ-Bo - AI enhanced Intelligent - Bolt' project, researchers from the Fraunhofer Cluster of Excellence Cognitive Internet Technologies CCIT have succeeded in integrating a trainable AI model that autonomously identifies anomalies into a vibration sensor. The innovation here is that the AI is operated on the sensor and recognizes deviations directly at the point of data generation. According to the researchers, this enables a future-oriented type of condition monitoring - energy-efficient, decentralized and in the smallest of spaces.

Analysis directly in the edge device

Vibrations are both indicators and frequent causes of wear and damage to drive systems. In 'AIQ-Bo', an optimized AI model in a three-axis acceleration sensor analyses the recorded vibration data. This enabled the researchers at Fraunhofer CCIT to use AI in a microcontroller directly at the point of action in such a way that it evaluates sensor data locally and in an energy-efficient manner.
This means that the vibration data does not have to be sent to the cloud for analysis. The AI is executed on the device and the information is processed via the Fraunhofer 'embeddif.ai' technology where it is generated: in the edge device. The trained AI model in the sensor's integrated microcontroller independently recognizes whether the component is running in normal operation or may reach critical states - without transmission to the cloud. Only when there is a problematic status is it transmitted wirelessly to the cloud. This greatly reduces the amount of data to be transmitted and makes the system extremely energy-efficient.
This is made possible by the open source AI framework 'AIfES'. This can be used to operate and train artificial neural networks (ANN) on almost any hardware in IoT devices.

Wireless AI retraining from the cloud

Computing-intensive tasks such as training the AI algorithm or calculating a new AI model are performed in the cloud. Model parameters can be updated and sent back to the sensor by radio to adapt the anomaly detection to changing environmental conditions, for example. 'mioty' wireless technology is used to transmit the data. Energy harvesting supplies the sensors and the wireless system with energy from the immediate surroundings, using temperature or sunlight, for example, to generate electricity. An alternative supply from small batteries is also possible.
Automated, AI-supported condition monitoring with 'AIQ-Bo' means that the system does not have to be checked manually by technical personnel, which would not be feasible during ongoing operation - for example of an offshore wind farm - and would therefore be associated with high downtime costs.
The technology solution is highly adaptable: "Our aim was to design 'AIQ-Bo' so flexibly that it can be used on different systems with as little training as possible, for example on wind turbines, bridges or overhead cranes," says Dr. Peter Spies, project manager at the Fraunhofer Institute for Integrated Circuits IIS. Training in the cloud and the transfer of data via radio can also be transferred to different application scenarios.

The project partners

The Fraunhofer Institute for Integrated Circuits IIS was involved in the Fraunhofer Cluster of Excellence Cognitive Internet Technologies CCIT project, contributing its embeddif.ai technology and expertise in the areas of energy harvesting and efficient communication and enabling the demonstration in a laboratory environment. In addition, the open source AI framework 'AIfES' from the Fraunhofer Institute for Microelectronic Circuits and Systems IMS was used, which also carried out the optimization of the AI models. The Fraunhofer Institute for Wind Energy Systems IWES was also responsible for the vibration measurements on numerous test benches and wind turbines. The Fraunhofer Institute for Applied and Integrated Safety AISEC is the spokesperson for the Fraunhofer CCIT.