The rapid expansion of satellite constellations, such as Starlink, has necessitated a shift from traditional humanbased telemetry monitoring to more autonomous systems. With thousands of new satellites overwhelming existing operators, the development of automated satellite health monitoring has become essential. Current systems are largely statistical in nature, but AI-driven solutions ofer the potential to significantly improve the accuracy and eficiency of anomaly detection. However, for AI systems to be adopted, satellite operators must not only trust the decisions made by these systems but also understand the reasoning behind them. This paper establishes a baseline autonomous health monitoring framework for satellite telemetry, with a focus on reliability and real-time anomaly detection. We leverage the EIRSAT-1 telemetry dataset to train and deploy our model and evaluate its performance across various Edge AI devices that either possess flight heritage or are scheduled for future space missions. These devices demonstrate the potential for real-time anomaly response, which is critical for mission success. To further enhance the accuracy of anomaly detection, we incorporate data denoising techniques as part of the postprocessing pipeline. Noise in reconstruction error can obscure subtle anomalies, leading to false positives or missed detections. By employing various denoising methods, we aim to create a clearer signal that allows for more accurate thresholding. We investigate the use of statistical approaches, such as moving averages and filtering, alongside machine learning models, such as clustering methods, for thresholding anomaly detection. These alternative approaches are compared in terms of computational eficiency, robustness, and suitability for real-time, onboard deployment in space environments.
In this paper, we propose a baseline autonomous health monitoring framework designed to operate
on satellite telemetry data. Our system aims to reliably detect anomalies and respond in real time to
support mission-critical decision-making. To demonstrate the viability of our approach, we utilize the
EIRSAT-1 telemetry dataset[
Satellite systems are highly complex hardware devices which generate substantial amounts of data. However, due to down-link bandwidth restrictions, only a fraction of that data makes it to ground. They are also limited when it comes to autonomy. If an error occurs on a satellite, it must go through the down-link cycle and be analyzed by satellite operators on ground, creating a large lag time between any action being taken to rectify the error. This can potentially lead to the loss of the satellite and the premature end of the mission. Our solution aims to be run on-board and in real-time in order to remove these potential problems in the anomaly detection and rectification pipeline. However, this adds more complexity to the mission. Satellites are powered entirely by solar arrays, limiting the power available on-board. Space is also at a premium as mass to orbit costs are in the order of €30,000 per kg. Being in a vacuum environment also adds thermal management issues. Finally, a radiation intense environment is also an aspect that must be considered depending on the orbit of the mission. This paper aims to address these concerns and presents a potential alternative to the current ground based anomaly detection paradigm.
Our methodology includes training the model on the flight test dataset using only nominal (healthy) data. We then evaluate its performance by testing it on the remainder of the flight test dataset, where artificial anomalies have been injected to simulate potential issues. This allows us to assess the model’s ability to detect anomalies under controlled conditions. Following this, we deploy the model on the thermal vacuum (TVAC) dataset, which contains real anomalies. Since the TVAC dataset is more representative of actual satellite conditions, it provides a more realistic test of the model’s robustness and accuracy.
Based on the insights gained from both the labeled flight test data and the real anomalies in the TVAC dataset, we define our thresholding methods. This process is critical for optimizing the model’s sensitivity and reducing false positives. Once optimized, we deploy the model on the unlabeled flight dataset to further evaluate its performance under more realistic, unlabeled conditions, demonstrating its ability to generalize to unseen data.
A key component of our approach is the preprocessing of telemetry data to remove noise, which can obscure meaningful anomalies. Denoising the data helps to enhance the accuracy of the anomaly detection process. In this paper, we explore various denoising techniques, including both statistical methods (e.g., moving averages, Kalman filtering) and machine learning models (e.g., autoencoders, isolation forests). These techniques are investigated for their ability to improve thresholding mechanisms in detecting anomalies. We compare the efectiveness of these approaches in terms of computational eficiency, robustness, and suitability for real-time, onboard deployment, ensuring that the selected method can operate reliably in the challenging conditions of space environments.
By establishing a baseline autonomous health monitoring system and investigating advanced denoising and thresholding methods, we aim to contribute to the growing field of AI-driven satellite health management, paving the way for more reliable and interpretable systems capable of supporting future space missions.
This section describes the basic structure of our experiments and the machine-learning (ML) pipeline deployed for anomaly detection. Based on the challenging aspects of this problem, the ML pipeline can be broken into two major components: dataset preprocessing and model development. We analyze a multivariate time series and simplify it to a less complex feature space, which is then used to identify anomalous data.
We use the EIRSAT-1 dataset as the data source for our experiments [
This section describes the diferent aspects of the ML model that convert our original multi-variate time series into a univariate time series that can be used to detect anomalies reliably.
The first part of the ML pipeline consists of an autoencoder model. An autoencoder is a
machinelearning model built on an encoder-decoder architecture that aims to learn a low-dimensional input
data representation. The decoder then reconstructs the original input data from the low-dimensional
abstraction. Our LSTM autoencoder model[
The model is trained to optimize the Mean Absolute Error loss function with the Adam Optimizer. Early Stopping is used to obtain the optimum model based on the validation loss. The autoencoder is trained only on the nominal data in the Flight-Test dataset. This trains the model to reconstruct the nominal data well while amplifying the reconstruction loss for anomalous data. While the number of epochs while training is fixed at a relatively high number (e.g., 100 epochs), we trace the validation loss after each epoch, and model weights are updated only if the validation loss falls below the best performance during training (up to that point). The model gives an ROC − AU C value of 0.86 on the Flight-Test dataset and 0.98 on the TVAC dataset, thus demonstrating that the model is trained well. The reconstruction error over the whole dataset is integrated into a time series, and we use various thresholding techniques to identify anomalous data.
While the resultant time series can be used for classification, it is still noisy (especially in the case of the Flight data) and leads to false positives on anomalies. Therefore, we use a post-processing function to denoise the time series. This function uses a pre-defined filter X of alternating zeros and ones ([0, 1, 0, ...]) that repeats for a set window size W . In our experiment, we set the size as 20. A significant value for W leads to an oversimplified signal, while a small value retains noisy data. We convolve this filter over our time series to denoise it as shown in Equation 1.
Tdenoised = (X ×
W ) ∗ Tnoisy (1) A comparison of the noisy and denoised signal on the Flight data is shown in Figure 2. The figure shows how significantly the denoising function improves the chances of fitting an appropriate threshold on the time series.
This section discusses the methods we applied to the obtained reconstruction errors from the autoencoder models. Thresholding here refers to establishing a linear or non-linear boundary on the y-axis, which efectively functions as a decision boundary for anomaly detection. Finally, we again used the standard scaler to rescale the signal.
Linear thresholding is a common method to identify anomalies in a time series dataset. We can then use a linear threshold to predict anomalies based on the reconstruction error. The high ROC score, as given in the previous sections, proves that the model is well-trained, and an optimum threshold would return a high F1-Score on Anomaly Detection. We generate an equidistant distribution of fifty points between the 50th and 100th percentile of the TVAC reconstruction error time series to determine this optimum threshold value. Based on the data, we determine the F1 score at each percentile and select the best-performing percentile as the static threshold for the data. This is then applied to the other datasets.
We adopt another linear thresholding method to account for the shortcomings of previous methods, which will be discussed in greater detail in the next section. For this method, we calculate a Z-score that calculates the diference of each point Ei in the time series signal E from the mean value in terms of the number of standard deviations of the signal. The Z-score of a point in the time series is defined by Equation 2:
Z(Ei) =
Ei − µ (E) σ (E) (2)
Here, µ (X) refers to the mean of the signal while σ (E) represents the standard deviation value of the signal. With this equation, the mean of our signal is 0, and the magnitude of each sample is defined in terms of the signal standard deviation. To tune the threshold hyperparameter, we test a range of values between 1 and 2 on the TVAC dataset, giving an optimum value of 1.5.
We also used a K-Nearest Neighbors Classifier as a supervised approach to anomaly detection on the problem. In this case, we used it as a classifier on the resultant reconstruction error from the LSTM autoencoder. The k-nearest neighbors (KNN) algorithm is a non-parametric, supervised learning classifier that uses proximity to make classifications or predictions about the grouping of an individual data point. The intuition behind using KNN here is that nominal data points should look structurally similar so that a KNN can classify anomalies successfully. The model performance in the TVAC dataset is shown in the table 1. KNN is helpful for anomaly detection, but it has a significantly higher inference time when compared to the other models used in this paper. Standard Scaler normalizes each feature before passing it into the model. The value of K is set at 3. We also use 5-fold cross-validation to validate the model performance on the training set. We train the KNN model on a split TVAC dataset and then use the model to make predictions on the Flight data to compare the similarity of predictions between all our detection approaches.
This section discusses the performance of our ML architecture on the TVAC and Flight datasets and the insights recovered from these datasets about the diferent intricacies of the classification/thresholding methods.
Applying our ML architecture to the TVAC data allows us to test the performance of the various thresholding or classification methods and draw a comparative analysis between them. Table 1 shows the performance of the diferent techniques on the TVAC dataset. Here, KNN is shown to do well across the three methods in terms of F-Score, but the other thresholding methods do better on the recall metric. We can further evaluate this by looking at Figure 3. This figure shows that we see a more consistent anomaly prediction on the two thresholding methods compared to KNN, where we notice that the anomalous areas are not consistently reported as anomalous with a small number of nominal predictions. This is reflected in the slightly lower recall value for KNN in Table 1. Additionally, the orange highlighted portion of the figure depicts the ground truth anomalies in the dataset. There are two types of anomalies on the TVAC dataset: • During vibration testing, a damaged solder joint on the battery board and changing pressure in TVAC stopped the communication between the battery and the OnBoard Computer (OBC). The noticeable change in the data for this anomaly is that the Battery Voltage parameter (plat
form.BAT.batteryVoltage) drops to 0. • A communication issue on the I2C Line between the onboard computer and the communications subsystem where the primary indicator for this anomaly is that the Temperature value shown in the parameter (platform.CMC.temperaturePA) abruptly goes to 0. It is important to note that this value can naturally go to 0 if the sensor is facing away from the sun in colder conditions. Therefore, it is also essential to notice the other CMC (Common Mode Current) parameters in the dataset.
We tested the autoencoder-thresholding combination on the scaled flight dataset, revealing interesting information about the flight dataset. The Flight Dataset oficially has no recorded anomalies. However, the model identified some anomalous areas where large chunks of anomalies were predicted. While the predicted chunks of anomalies might not be anomalous, it seemed an Out-of-Ordinary Operation (OOO). Upon investigation, we discovered that the model was identifying the switching states of the GMOD (Gamma Ray Module) and EMOD (Enbio Module) modules, which is not anomalous but not a regular operation that should happen. Table 2 shows a set of OOOs in the Flight Data with a static threshold. Similarly, the detected OOOs on the Flight data can be seen in Figure 4.
While there are three chunks of reported anomalies whose source has not been investigated yet, most chunks are labeled in the figure with their associated time stamps on the x-axis. One chunk of anomalies is not entirely reported as anomalous depending on the threshold value and the thresholding method used. A lower threshold with filtering based on consecutive anomalies solves this problem with some methods. Another interesting note is that the Z-Score predicts a consecutive chunk of anomalies in many cases. At the same time, KNN may give a tiny number of scattered nominal predictions in an OOO chunk. Since anomalies and OOOs do not occur as one-of isolated events in the time series but instead appear in a time chunk, both predictors should be considered valid.
Table 2 shows ares in the flight dataset that were labeled as OOO. After investigating why the model was producing such high reconstruction errors, we found we were detecting the powering on/of of the GMOD and EMOD payloads.
For the targeted low-power small-footprint high-performance hardware to combat the constraints encountered in space-based missions, such as power, radiation tolerance, mass, etc. The developed LSTM model was deployed on two diferent flight-ready hardware devices.
• Nvidia Jetson Orin Nano 8GB [
The LSTM model was deployed on all three EIRSAT-1 datasets, discussed in Section 2.1. The following
results demonstrate the performance of the LSTM model on each hardware devices for each dataset.
The Flight dataset is unlabelled and as a result the inferences per second (IPS) was the only metric
obtained. The hardware selected for this dataset is based on previous work to investigate which edge AI
devices with flight heritage and can efectively run anomaly detection architectures[
As the Orin Nano can run at multiple power profiles, we explored the performance diferences between the nominal 15W profile and the reduced power 7W profile.
Initially, we deployed the model and dataset on the Orin Nano at the nominal 15W profile. Table
3 shows interesting results on the deployment of the LSTM model on the Nvidia Jetson Orin Nano
8GB. As we are utilizing its GPU acceleration capabilities we expect slightly faster inferencing times.
However, GPU acceleration tends not to afect inferencing times as much as training times. The Orin
Nano is capable of 40 TOPS (Trillion Operations Per Second), specifically at the maximum power mode
of 15W [
Table 3 demonstrates further results on the deployment of the LSTM model on the Nvidia Jetson Orin Nano 8GB, which is set to a power mode of 7W and utilizing its GPU acceleration capabilities. There is a clear diference between the IPS in the 15W and 7W power modes. Most notably, the IPS diference on the TVAC dataset which went from to 2226.29 predictions per second for the 15W power mode, to 1270.12 predictions per second while using the 7W power mode. The performance diference in the two power modes is primarily due to reduced power availability in the 7W mode, which limits the GPU’s computational capability, reduces clock speeds, and thus lowers the rate at which predictions can be made by the model. Performance metrics such as AUC score, precision, recall and F1 score are not efected by the hardware power mode as they measure the efectiveness of the model’s predictions, not the speed at which the predictions are made. It is important to highlight that, since the Jetson Nano operates on a full Ubuntu OS, the model’s accuracy remains unafected by factors such as floating-point precision variations. This is because no adjustments to floating-point computations are required in this environment, ensuring that the model maintains its performance consistency during deployment.
The deployment of the LSTM model on the Raspberry Pi 4 [
Table 3 shows the highest inference per second values compared to the other hardware deployments.
The Intel Xeon CPU [
The Long Short-Term Memory (LSTM) model developed in this study is notably smaller in size compared to other publicly available models but maintains competitive classification accuracy. This compact architecture is well-suited for deployment on low-power devices, achieving a consistent inference rate exceeding 1000 inferences per second, except in cases where memory limitations are encountered, such as on the Raspberry Pi. The high inference rate is particularly advantageous for classifying analogue components like reaction wheels, whose housekeeping telemetry rates frequently surpass 100Hz. Additionally, the model’s eficient performance suggests that there is substantial computational overhead available, providing the flexibility to either scale the model for enhanced accuracy or to execute it alongside other on-board artificial intelligence (AI) tasks, such as Earth Observation (EO) algorithms. This integration could further strengthen the robustness and reliability of EO algorithms. A similar approach could be applied to Autonomous Navigation Systems (ANS), potentially enhancing their reliability as well. Overall, this work has shown that a small deep learning model could in fact run on hardware capable of being flown on the same mission the dataset has come from, enabling a next generation of satellite health monitoring and reliability.
This work demonstrates the feasibility of running a small-scale deep learning model on hardware suitable for space missions, thus enabling next-generation satellite health monitoring and reliability. Given the clear distinction between nominal and anomalous events in the dataset, future eforts will focus on refining the model to better characterize specific activities aboard EIRSAT-1. This will involve developing a multiclass classification model to not only detect anomalies but also classify them into predefined categories such as GMOD, EMOD, or unknown classes. Deploying this model on live EIRSAT-1 telemetry at the ground station will enable real-time detection of Out-of-Ordinary Operations (OOOs). In the long term, we aim to test this model on a future nanosatellite mission to enable real-time, on-board anomaly classification, facilitating a fully autonomous satellite health monitoring system.
This work was supported by Irish Research Council PhD Grant EBPPG/2020/11. This work was also supported by an European Space Agency General Science and Technology Program Grant. This work was further supported through the work done by ML Labs PhD program.