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

AI summary

Overview

Research area: Astronomical instrumentation and machine learning — specifically, time-series forecasting of telescope monitoring data (astro-ph.IM).

Technical level: Intermediate. The core idea (predicting what "normal" looks like, then flagging deviations) is easy to grasp, but the terminology around normal behavior models, multivariate forecasting, and the MLP/LSTM comparison assumes some familiarity with machine learning.

Scope (one sentence): The paper builds and evaluates a neural-network model that forecasts the normal operating behavior of the ASTRI-Horn Cherenkov telescope's motor-related sensor readings, as a foundation for detecting anomalies and eventually supporting predictive maintenance.

What This Paper Is About

Telescopes like ASTRI-Horn produce continuous streams of monitoring data from their control systems, and operators need to know when something is drifting away from healthy operation. The problem is that "unhealthy" is hard to define directly, so the authors instead model what normal looks like. They train a model to predict the near-future values of 15 motor-related physical variables under normal conditions; when real readings diverge from those predictions, that divergence becomes the signal of a possible fault. The goal is a reliable, hour-scale forecasting tool that could later be embedded in an online anomaly-detection and predictive-maintenance system.

Key Contributions

1. A Normal Behavior Model (NBM) for the ASTRI-Horn telescope — a data-driven baseline of healthy operation built from real monitoring telemetry, rather than from physical simulation or hand-written thresholds.
2. A multivariate forecasting setup — a single Multi-Layer Perceptron trained to predict all 15 features simultaneously rather than one model per sensor, using a sliding-window scheme over fixed-length intervals.
3. A systematic evaluation across input/forecast-length configurations — the model was tested under different combinations of input sequence length (I) and forecast horizon (T), and benchmarked against an LSTM network.
4. A specifically tailored pipeline — cleaning, resampling, feature selection, and correlation analysis applied to telescope control-unit data spanning September 2022 to July 2024, establishing a reproducible recipe for this kind of instrument telemetry.

Main Findings

• The MLP performed consistently: Results held up across different features and across the various I-T (input-length / forecast-length) configurations tested, rather than working only for one setting.
• It matched the LSTM while training faster: The MLP reached comparable accuracy to the LSTM benchmark but converged more quickly — a practical advantage for a model that might need periodic retraining.
• Best configuration reported: 4 hidden layers, 720 units per layer, with input and forecast lengths of 300 samples each — equivalent to 5 hours at 1-minute resolution.
• Test-set accuracy: An MSE of 0.019 ± 0.003 and an NMAD of 0.032 ± 0.009 under that best configuration.
• Longer horizons did not hurt: Extending the forecast window to 6.5 hours — the maximum the configuration allowed — did not degrade performance, supporting the claim of reliable hour-scale prediction.
• The abstract does not report which individual features were easiest or hardest to forecast, the relative size of the training/validation/test splits, or detailed LSTM numbers; those specifics are not in the abstract.

Methodology in Plain English

The authors started with telescope control-unit records of 15 physical variables measured at the Azimuth and Elevation motors between September 2022 and July 2024. They cleaned the data, resampled it to a uniform cadence, selected which features to keep, and checked how the variables correlate with one another. They then chopped the continuous record into fixed-length intervals: the first I samples of each interval served as the model's input, and the next T samples were the values it had to predict. Because a single pass through the data yields relatively few intervals, they used a sliding window — stepping through the timeline in overlapping chunks — to manufacture many more training examples. A Multi-Layer Perceptron was trained to output all features at once (multivariate forecasting), and its accuracy was measured with two error metrics, MSE and NMAD, and compared head-to-head with an LSTM network doing the same job.

Why This Matters

Impact on research: The work demonstrates that relatively compact feed-forward networks can serve as credible normal-behavior models for astronomical facility telemetry, converging faster than recurrent alternatives while achieving comparable accuracy. It provides a template for instrument monitoring that other observatories could adapt, and it links astronomical operations research to the established PHM (prognostics and health management) literature.

Real-world applications:

• Telescope operations: Flagging sensor readings that deviate from predicted normal behavior before they escalate into failures or lost observing nights.
• Predictive maintenance scheduling: Using deviation trends to plan interventions during daytime or weather downtime instead of reacting to breakdowns mid-observation.
• Multi-instrument observatories: Extending the same NBM approach to other subsystems (mirrors, cooling, cameras) or to the Cherenkov Telescope Array's larger instrument set.
• Industrial condition monitoring: The general recipe — clean telemetry, sliding-window multivariate forecasting, NBM as an anomaly baseline — transfers to motors, drives, and rotating machinery outside astronomy.

Industry relevance: The NBM-plus-anomaly-detection pattern is standard in industrial predictive maintenance, so this paper shows that a technique developed in that world ports effectively to scientific infrastructure. The finding that a simpler architecture matches an LSTM at lower training cost is directly relevant to operators who need models that can be retrained routinely on modest hardware.

Future Directions

• Closing the loop to anomaly detection: The abstract presents the NBM as an enabler; actually deploying it to flag anomalies on live ASTRI-Horn streams is the stated next step.
• Building a full prognostics and health management system: Moving from detecting deviations to predicting remaining useful life and recommending maintenance actions.
• Broadening the feature set: The current model covers 15 motor-related variables; whether it extends to other telescope subsystems is not addressed in the abstract.
• Online operation questions left open: How the model handles data drift, seasonal or environmental shifts over years, and retraining cadence are not discussed in the abstract.

Target Audience

Instrumentation and operations researchers in astronomy, particularly those working on telescope monitoring and observatory automation; machine learning practitioners interested in time-series forecasting and normal-behavior modeling for anomaly detection; and industrial predictive-maintenance engineers looking for a validated, real-world case study of the NBM approach on scientific infrastructure.

Authors’ abstract

This study presents a Normal Behavior Model (NBM) developed to forecast monitoring time-series data from the ASTRI-Horn Cherenkov telescope under normal operating conditions. The analysis focused on 15 physical variables acquired by the Telescope Control Unit between September 2022 and July 2024, representing sensor measurements from the Azimuth and Elevation motors. After data cleaning, resampling, feature selection, and correlation analysis, the dataset was segmented into fixed-length intervals, in which the first I samples represented the input sequence provided to the model, while the forecast length, T, indicated the number of future time steps to be predicted. A sliding-window technique was then applied to increase the number of intervals. A Multi-Layer Perceptron (MLP) was trained to perform multivariate forecasting across all features simultaneously. Model performance was evaluated using the Mean Squared Error (MSE) and the Normalized Median Absolute Deviation (NMAD), and it was also benchmarked against a Long Short-Term Memory (LSTM) network. The MLP model demonstrated consistent results across different features and I-T configurations, and matched the performance of the LSTM while converging faster. It achieved an MSE of 0.019+/-0.003 and an NMAD of 0.032+/-0.009 on the test set under its best configuration (4 hidden layers, 720 units per layer, and I-T lengths of 300 samples each, corresponding to 5 hours at 1-minute resolution). Extending the forecast horizon up to 6.5 hours-the maximum allowed by this configuration-did not degrade performance, confirming the model's effectiveness in providing reliable hour-scale predictions. The proposed NBM provides a powerful tool for enabling early anomaly detection in online ASTRI-Horn monitoring time series, offering a basis for the future development of a prognostics and health management system that supports predictive maintenance.