🔭 astrophysics

Multivariate time-series forecasting of ASTRI-Horn monitoring data: A Normal Behavior Model

This study presents a Normal Behavior Model based on a Multi-Layer Perceptron that effectively forecasts multivariate ASTRI-Horn telescope monitoring data up to 6.5 hours ahead with performance comparable to LSTM networks but faster convergence, thereby enabling early anomaly detection and predictive maintenance.

Original authors: Federico Incardona, Alessandro Costa, Farida Farsian, Francesco Franchina, Giuseppe Leto, Emilio Mastriani, Kevin Munari, Giovanni Pareschi, Salvatore Scuderi, Sebastiano Spinello, Gino Tosti

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Federico Incardona, Alessandro Costa, Farida Farsian, Francesco Franchina, Giuseppe Leto, Emilio Mastriani, Kevin Munari, Giovanni Pareschi, Salvatore Scuderi, Sebastiano Spinello, Gino Tosti

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, high-tech telescope named ASTRI-Horn sitting on a mountain in Sicily. It's designed to catch faint flashes of light from deep space. But like any complex machine, it has moving parts (motors) that can get tired, overheat, or break. If it breaks unexpectedly, the telescope stops working, and we miss out on precious scientific data.

The scientists behind this paper asked a simple question: "Can we teach a computer to predict when the telescope is about to act up, before it actually breaks?"

Here is how they did it, explained through everyday analogies:

1. The Problem: The "Silent" Breakdown

Telescopes are like marathon runners. They don't usually collapse suddenly; they get tired, their joints creak, or their temperature rises slowly.

The Old Way: Wait until the runner trips, then call an ambulance (Reactive Maintenance).
The New Way: Watch the runner's heart rate and breathing. If they start panting weirdly, tell them to stop before they trip (Predictive Maintenance).

The problem is that the telescope has broken only a few times in its history. There isn't enough data on "broken" states to teach a computer what a breakdown looks like. So, the scientists decided to teach the computer what "Normal" looks like instead.

2. The Solution: The "Normal Behavior Model" (NBM)

Think of the NBM as a super-smart tutor who has studied the telescope's "perfect days" for two years.

The Input: The telescope sends a constant stream of data: how fast the motors are spinning, how hot they are, how much electricity they are using, and where they are pointing.
The Training: The computer (an Artificial Intelligence) looks at the last 5 hours of "perfect" data and tries to guess what the next 5 hours will look like.
The Goal: It learns the "rhythm" of the machine. It knows that when the motor turns left, the temperature usually goes up a tiny bit, and the electricity usage follows a specific curve.

3. The "Crystal Ball" Test

Once the AI is trained, they put it to the test.

The Setup: They show the AI the last 5 hours of data and ask, "What happens next?"
The Prediction: The AI draws a "ghost line" of what it expects to happen next.
The Reality Check: They compare the AI's ghost line with what the telescope actually does.

If the lines match: The telescope is healthy. The AI is doing its job.
If the lines diverge: The telescope is doing something weird! Maybe a motor is grinding, or a sensor is stuck. This is the alarm bell.

4. The "Brain" Behind the Magic

The scientists tried two different types of AI "brains" to see which one was better at this prediction game:

LSTM (The Memory Keeper): A brain designed specifically for remembering long stories. It's great at remembering the past but takes a long time to learn.
MLP (The Fast Learner): A standard, multi-layered brain. It doesn't have a special "memory" module, but it learns incredibly fast.

The Result: The MLP (Fast Learner) won. It learned the telescope's rhythm just as well as the Memory Keeper, but it did it 10 times faster. It was like finding a student who could learn a whole year's worth of math in a week, just as well as a student who studied for a year.

5. How Far Can It See?

The best part? The AI isn't just guessing the next minute. It successfully predicted the telescope's behavior 6.5 hours into the future.

Why this matters: If the telescope is in a remote, hard-to-reach place (like a mountain top), knowing it will break in 6 hours gives the maintenance crew plenty of time to pack their tools, drive up the mountain, and fix it before the telescope stops working.

6. The "Glitches" in the System

The paper admits the AI isn't perfect.

The "Stop and Go" Problem: If the telescope stops moving completely, the AI sometimes gets confused and predicts it will keep moving. It's like a driver who keeps pressing the gas pedal even when the car is in neutral.
The "Spike" Problem: If there is a sudden, sharp jump in data (like a power surge), the AI tends to ignore it and predict the "average" instead. It's like a weather forecaster who predicts "sunny" because it's usually sunny, missing a sudden thunderstorm.

However, the scientists say this is okay. The goal isn't to predict every tiny flicker; it's to catch the big, dangerous trends that lead to a breakdown.

The Bottom Line

This paper describes a digital crystal ball for a space telescope. By teaching a computer what "healthy" looks like, they can spot the first signs of sickness hours before a breakdown occurs. This turns telescope maintenance from a "fix-it-when-it-breaks" emergency into a "schedule-it-for-Tuesday" routine, saving money, time, and ensuring we keep looking at the stars without interruption.

1. Problem Statement

Modern astrophysical observatories, such as the ASTRI-Horn Cherenkov telescope, are complex systems where unexpected breakdowns can lead to significant scientific data loss and increased maintenance costs. Traditional maintenance strategies (preventive or corrective) are often insufficient for preventing critical failures.

The specific challenges addressed in this study include:

Lack of Fault Data: Historical analysis of ASTRI-Horn revealed that the few documented failures (e.g., gear oxidation, encoder tape wear) did not exhibit clear, gradual degradation patterns in the time-series data, making supervised learning for anomaly detection difficult.
Need for Early Detection: There is a requirement for a system capable of detecting anomalies before they cause operational downtime, ideally providing a prediction window comparable to or exceeding the Mean Time To Repair (MTTR).
Complexity of Monitoring: The telescope generates high-volume, multivariate sensor data (current, voltage, temperature, torque, position, velocity) from its Azimuth and Elevation motors, requiring a model that can capture both temporal dynamics and inter-variable dependencies.

2. Methodology

The authors developed a Normal Behavior Model (NBM) based on deep learning to forecast monitoring time-series data under Normal Operating Conditions (NOCs). The methodology follows a structured pipeline:

A. Data Preprocessing

Data Source: 15 physical variables from the Telescope Control Unit (TCU) collected between September 2022 and July 2024.
Resampling: Data was resampled from variable intervals (approx. 15s) to a uniform 1-minute resolution to ensure consistency and reduce computational load.
Filtering: Only intervals where at least one motor was active were retained (325 active intervals).
Feature Selection: Correlation analysis revealed that motor torque was highly correlated ( $r > 0.99$ ) with motor current. To reduce redundancy, the three torque features were removed, leaving 12 features for modeling.
Cleaning: Outliers (physically impossible readings) were handled via forward/backward filling. Missing values were interpolated.
Normalization: Z-score normalization was applied to all features.

B. Data Segmentation (Sliding Window)

The dataset was segmented into fixed-length windows ( $W$ ).
Each window was split into an input sequence ( $I$ ) and a target sequence ( $T$ ), such that $W = I + T$ .
A sliding-window technique was applied to longer intervals to generate multiple training samples, significantly increasing the dataset size.

C. Model Architecture

Primary Model: A Multi-Layer Perceptron (MLP) (Feedforward Neural Network) was employed to perform multivariate forecasting (predicting all features simultaneously).
- Architecture: Input layer ( $I \times F$ ) $\rightarrow$ $L$ fully connected hidden layers ( $n$ neurons each) $\rightarrow$ Output layer ( $T \times F$ ).
- Optimization: Trained using the Adam optimizer with ReLU activation functions.
Benchmark: The MLP was compared against a Long Short-Term Memory (LSTM) network, a standard architecture for time-series data.
Training Strategy:
- Hyperparameter tuning via grid search (varying layers $L$ and units $n$ ).
- Early Stopping: Training was halted when validation loss ceased to improve (specifically at 112 epochs for the best configuration) to prevent overfitting.
- Loss Function: Mean Squared Error (MSE).

3. Key Contributions

Development of an NBM for ASTRI-Horn: The first application of a deep learning-based Normal Behavior Model specifically tailored to the motor monitoring data of the ASTRI-Horn telescope.
MLP vs. LSTM Comparison: The study demonstrates that while LSTMs excel with small datasets ( $W=60$ ), MLPs can match or slightly exceed LSTM performance on larger datasets ( $W=360$ ) while converging ~10 times faster, making them computationally more efficient for long-horizon forecasting.
Long-Horizon Forecasting: The model successfully predicts multivariate time-series data up to 6.5 hours into the future, a critical capability for predictive maintenance.
Robust Evaluation Framework: The use of Normalized Median Absolute Deviation (NMAD) alongside MSE provides a more robust assessment of typical error dispersion, reducing the influence of outliers.

4. Results

The model was evaluated using 100 independent training runs with random data splits.

Best Configuration:
- Input length ( $I$ ): 300 samples (5 hours).
- Forecast horizon ( $T$ ): 300 samples (5 hours).
- Architecture: 4 hidden layers with 720 units per layer.
- Training duration: 112 epochs.
Performance Metrics (Test Set):
- MSE: $0.019 \pm 0.003$
- NMAD: $0.032 \pm 0.009$
Forecast Horizon Analysis:
- Increasing the forecast horizon from 1 hour to 6.5 hours did not degrade performance; in fact, MSE decreased as $T$ increased up to 300 samples, plateauing thereafter. This confirms the model's ability to capture multi-hour trends effectively.
Feature-Specific Insights:
- Velocity/Current: Showed higher MSE but low NMAD, indicating accurate general trends but occasional large deviations during rapid transients (e.g., rapid repositioning).
- Temperature/Voltage: Showed lower MSE but higher NMAD, suggesting errors were more uniformly distributed, with the model sometimes struggling to track fine-scale variations or constant signals.
- Limitations: The model occasionally predicts motion when the telescope is stationary (vertical strips in residual plots) and struggles with sharp spikes, defaulting to the most probable value. However, these limitations do not hinder the primary goal of trend monitoring.

5. Significance and Future Work

Predictive Maintenance (PdM): The NBM serves as a foundational tool for a Prognostics and Health Management (PHM) system. By comparing real-time sensor data against model predictions, operators can detect statistically significant deviations (anomalies) up to 6.5 hours in advance.
Operational Impact: Early detection allows for scheduled maintenance, reducing unplanned downtime and maximizing scientific observation time for the telescope.
Scalability: The approach is adaptable to other Cherenkov telescopes (e.g., the ASTRI Mini-Array, CTAO Small-Sized Telescopes) and other astronomical infrastructure components (mirror actuators, cooling systems).
Future Directions:
- Extending the model to include longer operational datasets and more diverse components.
- Benchmarking against Transformer-based architectures.
- Developing synthetic fault data to train the model on specific degradation patterns, as real fault data is scarce.
- Investigating online learning or reinforcement learning paradigms for autonomous adaptation.

In conclusion, the paper presents a computationally efficient, robust MLP-based forecasting model that successfully replicates the normal behavior of the ASTRI-Horn telescope, offering a viable pathway toward automated, predictive maintenance for complex astronomical observatories.