Case study AI4FDIR – AI for fault detection, isolation & recovery
As modern satellite constellations grow in size, traditional, ground-based fault analysis is reaching its limits. In the AI4FDIR project, fortiss is shifting central intelligence directly onto the satellites to detect, isolate and rectify faults in real time.
To this end, the institute is developing an autonomous on-board system for fault detection, isolation and recovery (FDIR). By combining machine learning, deep learning and extremely energy-efficient neuromorphic data processing, this can significantly reduce downtime and recovery times, thereby enabling resilient, self-managing space systems.
Challenge
The autonomous management of large satellite telecommunications constellations places high demands on speed, reliability and resource efficiency. Existing approaches are predominantly based on ground-based systems, which means that faults can only be detected and rectified after a delay. At the same time, the complexity of the systems is constantly increasing, whilst only very limited computing and energy resources are available on board. This makes it increasingly difficult to identify faults in real time, clearly isolate their cause and initiate appropriate countermeasures before there is any noticeable impact on the communication service.
Solution
A key approach is to shift intelligent decision-making processes directly onto the satellites. To achieve this, modern machine learning and deep learning methods are combined with extremely energy-efficient neuromorphic data processing to enable fast and energy-saving inference under real-time conditions. The system continuously analyses heterogeneous telemetry data, detects anomalies at an early stage and automatically initiates measures for fault isolation and recovery.
In addition, various AI architectures are being investigated and evaluated on neuromorphic hardware. Development is proceeding in stages, from the analysis of public ESA anomaly data, through validation using realistic mission and simulation data, to a space-level neuromorphic proof-of-concept that demonstrates the feasibility of autonomous FDIR functionality in satellite operations.
Result
- Development and testing of AI models for classifying anomalies in satellite telemetry data on the neuromorphic BrainChip Akida platform
- Successful identification and utilisation of relevant fault and degradation patterns from the public ESA anomaly dataset for model training
- Derivation of suitable AI model architectures for energy-efficient on-board FDIR applications with a focus on real-time capability
- Validation of the developed approaches using realistic data from the Kepler satellite system and the OHB constellation simulator
- Demonstration of the models’ robustness under varying mission and operational conditions, as well as reliable early detection of anomalies
- Implementation of a neuromorphic proof-of-concept for on-board FDIR at the space system level
- Demonstration of the technical feasibility of AI-supported fault detection, isolation and response directly on board satellites
- Reduction of dependence on ground-based analysis through early, autonomous decision-making capability within the system context
Outcome
fortiss demonstrates that the combination of deep learning and neuromorphic data processing enables autonomous, rapid and energy-efficient fault handling directly within the satellite. This paves the way for future space constellations capable of self-monitoring, self-diagnosis and self-stabilisation.
Project partner
