Predictive Anomaly Detection for Rocket Boosters
The integrity of next-generation rocket boosters is paramount for mission success and safety. Early detection of anomalies, whether structural, propulsion-related, or systemic, can prevent catastrophic failures. This challenge focuses on building a robust, real-time predictive anomaly detection system for a rocket booster during its pre-flight testing phase, mirroring real-world scenarios like the recent Starship booster incident. Participants will develop an AI model that processes streams of multivariate sensor data (e.g., pressure, temperature, strain, vibration) from a simulated rocket booster. The system should identify subtle deviations from nominal operational parameters that indicate an impending anomaly or failure, long before it becomes critical. The solution will leverage advanced machine learning techniques, including time-series analysis and pattern recognition, to forecast potential issues and provide early warnings to ground control.
What you are building
The core problem, expected build, and operating context for this challenge.
Build an AI system using Gemini 2.5 Flash and Seldon Core to predict rocket booster anomalies from simulated sensor telemetry, employing automated feature engineering.
How work is evaluated
The evaluation will assess the anomaly detection system's accuracy, timeliness, and robustness. Participants will submit their model and deployment configuration, which will be tested against a simulated telemetry stream containing various known and unknown anomalies.
Challenge datasets
Public datasets, plus the private datasets you have access to.
What you should walk away with
Orchestrate a data ingestion pipeline for high-frequency telemetry data from a simulated rocket booster, handling missing values and sensor noise using Python (e.g., Pandas, Dask).
Master automated feature engineering for time-series data using Featuretools, extracting relevant statistical, temporal, and spectral features to enhance anomaly detection capabilities.
Implement a predictive anomaly detection model using Gemini 2.5 Flash, by framing the problem as identifying unusual patterns in sequences of sensor readings, potentially using its generative capabilities for synthetic anomalous data generation or pattern matching.
Design and integrate a digital twin concept by simulating booster behavior under various conditions and feeding this synthetic data alongside real-world-like telemetry for model training and validation.
Deploy the trained anomaly detection model as a real-time inference service using Seldon Core on a Kubernetes cluster, ensuring low-latency predictions.
Develop robust monitoring and alert systems for the deployed model, integrating with Grafana or Prometheus to visualize predictions, confidence scores, and system health.
Optimize model parameters and deployment configurations for resilience against sensor failures and data drift, crucial for aerospace applications.
Reference links and supporting material
Sample data for 3 tasks
Challenge setup
The evaluation will assess the anomaly detection system's accuracy, timeliness, and robustness. Participants will submit their model and deployment configuration, which will be tested against a simulated telemetry str...
Challenge input
JSON or CSV files of simulated booster telemetry data for training (e.g., sensor_data.csv, anomaly_labels.csv).
Agent execution
No agent tool is listed for this challenge.
Evaluated output
Serialized model file (e.g., .pkl, .joblib, or TensorFlow SavedModel format) and Featuretools entity set/pipeline.
- Verify the submitted model and feature pipeline can be loaded successfully.
- Ensure the Seldon Core deployment configuration is syntactically correct and deploys successfully.
- Check if the inference API endpoint responds to requests within a specified latency.
- F1Score target: 0.85
- Python execution harness
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- Verify the submitted model and feature pipeline can be loaded successfully.
- Ensure the Seldon Core deployment configuration is syntactically correct and deploys successfully.
- Check if the inference API endpoint responds to requests within a specified latency.
Metrics
- F1Score · Target 0.85
- TimeToDetection · Target 10
- FalsePositiveRate · Target 0.02
Workflow tasks
- Task 1
Input: JSON or CSV files of simulated booster telemetry data for training (e.g., sensor_data.csv, anomaly_labels.csv).
Output: Serialized model file (e.g., .pkl, .joblib, or TensorFlow SavedModel format) and Featuretools entity set/pipeline.
Execution harness: Python
- Task 2
Input: A directory containing the serialized model and Featuretools pipeline.
Output: Kubernetes YAML configuration file(s) for Seldon Core, including a custom inference runtime if necessary.
Execution harness: Docker
- Task 3
Input: JSON object representing a single timestep's sensor readings: {"data": {"ndarray": [[timestamp, sensor_1, sensor_2, ...]]}}
Output: JSON object with prediction, e.g., {"data": {"ndarray": [[anomaly_score, is_anomaly]]}}
Execution harness: Docker
Recipe freeze marker
No recipe freeze time is recorded.