Health-AI On-Board Health Monitoring System For Satcom

  • Status
  • Status date
  • Activity Code

Traditional implementations of Failure Detection, Isolation and Recovery (FDIR) systems rely on monitoring fixed thresholds. As these thresholds are defined before launch, they are conservative and include margins. Moreover, satellite ageing alters the sub-systems’ behaviour and requires adaptation in detection and classification rules. Conversely, data-centric approaches powered by Artificial Intelligence (AI) algorithms can improve anomaly detection timeliness, enable onboard anomaly classification, predictive health monitoring, and a reduced dependency on ground operations.

The Health-AI project develops an innovative AI-powered FDIR system, exploiting recent innovations and developments in Deep Learning technology. The underlying objective is to improve the health monitoring of the different platform sub-systems and software elements of the spacecraft, including ADCS, EPS, and OBC.

The main goals of the activity are:

  • Design an AI-based FDIR system for onboard execution that is reusable and highly adaptable in different missions.

  • Test and benchmark the designed solution on industry-driven use cases obtained from real flight telemetry, covering a variety of satellite subsystems.

  • Assess the impact of AI-based FDIR systems on onboard computational requirements and hardware, including the employment of AI accelerators.

  • Demonstrate the execution of the FDIR system on a relevant test bench, targeting a Technology Readiness Level equal to 4.


Employing AI algorithms in critical applications poses reliability challenges due to the black-box nature of most machine learning approaches, which makes it difficult to provide proof of minimum guaranteed performance. To mitigate this, it is foreseen that advanced systems will work alongside traditional FDIR to augment performance, rather than substitute existing implementations altogether.

Additionally, data availability and data quality play a crucial role in the development of AI algorithms. In the activity, the role of Tyvak as a satellite integrator and operator is crucial to provide real-mission use cases and data for training and testing the Health-AI system.


When compared to traditional implementations of onboard FDIR, the Health-AI system is capable of:

  • Early detection: threshold violations are often anticipated by abnormal behaviour within the nominal range that can be detected by advanced algorithms.

  • Expanded detection capabilities: the Health-AI system can detect failures that do not result in threshold violations and go unnoticed by classical fault detection systems.

  • Easier root cause analysis: improved detection capabilities make it easier to identify root causes directly onboard.

  • Improved multivariate analysis: artificial intelligence algorithms receive multiple data streams as input and learn correlations and dependencies among variables.

  • Reduced ground intervention: improved onboard autonomy can result in a prompt reaction that prevents the interruption of service and a reduced workload for ground operators.

  • Support to mission data analysis: outputs of anomaly detection and classification can help further investigations from operators on the ground. Moreover, telemetry analysis can spot ageing effects and support in-orbit tuning of subsystem configurations. 


The main system function is anomaly detection and classification from onboard satellite telemetry. In the detection step, deep learning models predict the future nominal evolution of selected satellite’s telemetry points, which is compared with the real-time measurements. A non-parametric dynamic thresholding detects significant error increases and issues alerts for potential anomalies.
In the classification step, a Knowledge-Based System infers: i) whether the alert is a false-positive; ii) the anomaly class that fits the data more; and iii) the confidence level of the classification output. The inference results are packed into an FDIR event message which is transmitted to the satellite flight computer.

System Architecture

The proposed system design leverages orbital_OLIVER, an onboard satellite operation automation software developed by AIKO. orbital_OLIVER is an AI-based software designed to enhance the autonomy level of a spacecraft, based on a distributed microservices architecture, which also comprises modules for telemetry forecasting and FDIR.
The Health-AI system leverages two main orbital_OLIVER services, conveniently configured and tailored to address the project use cases. The Sensing Service extracts relevant information from data available onboard through inferences of DL models. Within the Health-AI project, it analyses telemetry streams to detect anomalous behaviour or predict the future evolution of target quantities. The Reasoning Service implements a Knowledge-Based System that performs reasoning tasks, inferring on a structured database of rules and parameters. For the Health-AI project, its role focuses on anomaly classification starting from the platform state and the output of the Sensing Service. It outputs the FDIR events that are then transmitted to the platform.
orbital_OLIVER’s modular design allows a high degree of adaptation to diverse hardware configurations. During the project, the Nvidia Jetson Nano and Ingeniars’ GPU@SAT have been explored for system deployment and testing.


Over a duration of 18 months, the project activity encompasses the detailed definition of technical requirements, use cases, and dataset generation, followed by the complete system design and development. Finally, the project concludes with deployment on target hardware configurations and a thorough testing campaign.

Current status

The project underwent a successful Final Review on 9th November 2023 and is now completed.
During the activity, the consortium developed an on-board FDIR software for anomaly detection and classification, and telemetry forecasting. The system reached TRL 4 through an extensive testing campaign, which proved its suitability to operate onboard with positive performance. Tested use cases cover different satellite subsystems: ADCS, EPS, and OBC.
The system has been deployed in different configurations on representative hardware, including Artificial Intelligence accelerators. In particular, IngeniArs’ GPU@SAT has been proven suitable to support onboard FDIR.

Prime Contractor