AI Agents for Adaptive Anomaly Detection in NASA's Deep Space Network

TLDR: This research paper introduces an agentic AI system designed to automate anomaly detection within NASA’s Deep Space Network (DSN). The system integrates deep learning models (LSTM, GAN, TST) to identify equipment degradation in multivariate time-series data, a reinforcement learning component with human feedback for classifying anomaly severity, and a fine-tuned Large Language Model (LLM) for generating detailed discrepancy reports. The goal is to provide JPL engineers with an AI assistant that can precisely pinpoint anomalies, predict degradation trends, and ensure the continuous operation of the DSN for future space missions.

NASA’s Deep Space Network (DSN) is a critical global system of antenna facilities that ensures communication with spacecraft far beyond Earth. These facilities, located in Canberra (Australia), Madrid (Spain), and Goldstone (California), generate vast amounts of multivariate time-series data. Over time, DSN antennas and transmitters can degrade, leading to costly disruptions and threatening vital connections to dozens of spacecraft. To address this, a recent study explored innovative methods and tools to help Jet Propulsion Laboratory (JPL) engineers pinpoint anomalies and equipment degradation using collected data.

The Challenge of Deep Space Network Maintenance

Maintaining the DSN is a complex task. Engineers need to identify subtle signs of degradation in equipment that operates continuously for years. Traditional methods often rely on human observation of data visualizations, which can miss long-term trends or subtle anomalies. The goal of this research was to develop an autonomous system that could analyze these trends over extended periods, providing precise and accurate anomaly detection.

A Multi-faceted AI Approach

The study proposes an advanced system that integrates several artificial intelligence and machine learning techniques. Instead of relying solely on traditional algorithms, it leverages modern deep learning architectures, a reinforcement learning subsystem, and a Large Language Model (LLM), all orchestrated by an agentic AI system.

Data Sources and Preparation

The system utilizes two primary data sources: DSN Performance Analysis (DSNPAR) datasets from antenna systems and DSN transmitter datasets, which are collected via Microsoft Outlook emails. The DSNPAR data includes raw information from various subsystems with hundreds of thousands of data points and features like Power Carrier Number (PCNO) and Symbol Signal to Noise Ratio (SSNR). Transmitter data, on the other hand, required a separate program to extract and parse information from unstructured email formats, differentiating between older JPL transmitters and newer CEC transmitters.

Before feeding data into the models, extensive processing was performed. This included feature selection using techniques like Principal Component Analysis (PCA) to identify the most correlated parameters, data normalization using Min-Max Scaler, and outlier filtering with Isolation Forests to prevent models from overfitting on anomalous data. Timestamps were converted into sequences for time-series models, and data was prepared as PyTorch tensors for efficient GPU-accelerated training.

Deep Learning for Anomaly Detection

The researchers experimented with three deep learning architectures for reconstruction-type anomaly detection:

Long Short-Term Memory (LSTM) Networks: These are a type of Recurrent Neural Network (RNN) particularly effective at learning long-term dependencies in sequential data. LSTMs use a gating mechanism to regulate information flow, making them ideal for modeling and detecting degradation trends over time.
Generative Adversarial Networks (GANs): GANs consist of two competing neural networks – a Generator that creates data and a Discriminator that distinguishes real from fake. A hybrid GAN-LSTM model was developed to leverage both generative capabilities and temporal dependency analysis for smoother, more realistic temporal transitions in time-series data.
Time-Series Transformers (TSTs): Transformers use self-attention mechanisms to weigh the importance of each data entry relative to others in a sequence. TSTs are adapted for multivariate time-series data, capable of handling long-term dependencies and multivariate sequences as effectively as LSTMs, and are particularly good at identifying globally important patterns.

While preliminary results showed that GANs produced smoother reconstruction errors, TSTs made anomalous points more prevalent and clearer to spot. The study highlighted the need for further research into hyperparameter tuning and optimization to reduce false positives.

Adaptive Anomaly Classification with Reinforcement Learning

To enhance the reliability of detected anomalies, a reinforcement learning (RL) subsystem was implemented. Specifically, a Q-Learning algorithm was used to act as a double-verification system. Given the rarity and ambiguity of anomalies, RL agents can learn to detect new patterns by continuously interacting with the system. The Q-Learning agent analyzes detected anomalies based on factors like weather and wind, and then classifies them by severity. Crucially, this system incorporates human feedback, allowing JPL operators to confirm or reject the model’s classifications. This iterative feedback loop enables the model to improve its accuracy and adaptability over time, especially given the limited historical data on true/false anomalies.

Agentic AI and Large Language Models for Discrepancy Reports

The entire system is wrapped within an agentic AI framework, orchestrated by the LangGraph framework. This allows for complex reasoning and decision-making. A fine-tuned Large Language Model (LLM), specifically the Mistral-7B-v0.1 model, plays a pivotal role. It was fine-tuned using a large dataset of DSN Discrepancy Reports (DRs) provided by NASA JPL. This enables the LLM to extract contextual information, analyze anomalies based on their metadata, and generate human-fluent discrepancy reports, including potential causes and corrective actions. Low-Rank Adaptation (LoRA) was used for cost-efficient fine-tuning, and the LLM was hosted locally on JPL servers to ensure data privacy and security.

The AI agent workflow seamlessly connects the deep learning models, the RL model, and the LLM. It processes data from both antenna and transmitter components, computes anomaly scores, explains decisions, and incorporates human feedback, ultimately generating automated discrepancy reports.

Also Read:

Future Directions

Future work includes optimizing hyperparameter tuning for the deep learning models, potentially using automated methods like Grid Search or Bayesian Optimization. Integrating real-time data pipelines through frameworks like Apache Spark and Kafka is also envisioned to enable continuous monitoring and fully autonomous anomaly detection. Additionally, transfer learning could be employed to adapt models trained on one DSN-SCID pair to detect anomalies in others, further improving efficiency and adaptability. This comprehensive approach aims to provide JPL engineers with an invaluable AI assistant for diagnosing anomalies and proactively managing the DSN’s critical infrastructure.

You can read the full research paper here.

AI Agents for Adaptive Anomaly Detection in NASA’s Deep Space Network