Gravitational wave signal extraction against non-stationary instrumental noises with deep neural network
JournalPhysics Letters B 858, 139016 (2024)
First public: Published online:
- Detection & denoising

Research summary
Tests deep-learning extraction of massive-black-hole-binary signals in data affected by gaps, glitches, and changing noise correlations.
Extended research summary
Tests deep-learning extraction of massive-black-hole-binary signals in data affected by gaps, glitches, and changing noise correlations. Individual disturbances and their combinations are examined to assess robustness beyond ideal stationary-noise simulations.
Extended notes
Highlights
Robust Non-Stationary Denoising: First deep learning model specifically designed to handle three major types of non-stationarities in space-based gravitational wave data: data gaps, glitches, and time-varying noise.
Realistic Mission Scenarios: Addresses practical challenges from routine maintenance and unexpected disturbances that will occur during LISA/Taiji/TianQin science operations.
Maintained Performance: Achieves state-of-the-art accuracy for ideal data while demonstrating remarkable adaptability to anomalous conditions.
Comprehensive Anomaly Coverage: Successfully extracts massive black hole binary signals under individual non-stationarities and their complex mixtures.
Denoising Autoencoder Architecture: Novel bidirectional LSTM-based architecture specifically designed for gravitational wave signal extraction.
Mission-Critical Capability: Provides essential robustness for space-based detectors where non-stationary features are unavoidable.
Key Contributions
1. Addressing Real-World Data Challenges
Space-Based Detector Non-Stationarities
Unlike ground-based detectors, space-based missions face unique challenges:
- Extended Mission Duration: Multi-year operations (LISA: 4+ years) increase probability of anomalies
- Maintenance Requirements: Planned spacecraft operations causing data interruptions
- Environmental Variations: Solar activity, thermal variations, equipment aging
- Remote Operations: Limited intervention capabilities after launch
Three Critical Non-Stationarity Types
Data Gaps
- Routine maintenance windows
- Communication blackouts
- Instrument calibration periods
- Unexpected data loss
Transients (Glitches)
- Instrumental artifacts
- Environmental disturbances
- Micro-meteoroid impacts
- Spacecraft maneuvers
Time-Varying Noise Autocorrelations
- Thermal fluctuations
- Equipment degradation
- Solar activity variations
- Orbital position effects
2. Deep Learning Architecture
Denoising Autoencoder Design
The model architecture consists of:
Encoder
- Input: Normalized and segmented gravitational wave data with noise
- Overlapping subsequences for temporal coherence
- Feature extraction layers capturing multi-scale patterns
- Compression to characteristic feature vectors
Bidirectional LSTM Processing
- Three stacked bidirectional LSTM layers
- Forward and backward temporal context integration
- Captures long-range dependencies in signal structure
- Handles sequential patterns across extended durations
Decoder
- Dense layers for reconstruction
- Gradient-based refinement of waveform estimates
- Output: Cleaned gravitational wave signal
- Preserves phase and amplitude accuracy
Architecture Advantages
- Bidirectional processing critical for handling gaps (uses future context)
- LSTM memory cells maintain coherence across non-stationary regions
- Autoencoder framework learns robust signal representation
- End-to-end training for optimal denoising
3. Comprehensive Robustness Validation
Systematic Testing Protocol
- Individual non-stationarity types tested separately
- Pairwise combinations evaluated
- Triple combination (all three types simultaneously)
- Various severity levels and configurations
Performance Metrics
- Signal recovery accuracy (waveform fidelity)
- Phase preservation (critical for parameter estimation)
- Amplitude estimation errors
- Comparison with ideal-case performance
Methodology
Data Generation
Gravitational Wave Signals
- Massive black hole binary coalescences
- Mass range: 10⁴ to 10⁷ solar masses
- Inspiral-merger-ringdown phenomenological waveforms
- Various mass ratios, spins, sky locations, distances
- Time-domain waveforms in detector frame
Noise Modeling
Baseline Noise
- Instrumental noise from LISA/Taiji sensitivity curves
- Galactic confusion noise from unresolved binaries
- Power spectral density-based generation
- Realistic amplitude and frequency characteristics
Non-Stationary Components
- Data gaps: Random duration and placement
- Glitches: Various morphologies (sine-Gaussians, wavelets)
- Time-varying noise: Modulated autocorrelation function
- Multiple simultaneous non-stationarities
Training Procedure
Dataset Construction
- Training set: Signals with various non-stationarity combinations
- Validation set: Independent parameter draws with anomalies
- Test set: Held-out parameters and anomaly configurations
- Data augmentation: Random time shifts, amplitude variations
Normalization and Segmentation
- Z-score normalization per data segment
- Overlapping windowing for temporal smoothness
- Segment length chosen for computational efficiency and context
- Padding strategies for edge effects
Training Strategy
- Loss function: Mean squared error between clean signal and reconstruction
- Optimizer: Adam with learning rate scheduling
- Early stopping on validation loss
- Batch training for efficiency
Robustness-Focused Training
- Explicit inclusion of non-stationary data in training
- Curriculum learning: gradually increasing anomaly severity
- Regularization to prevent overfitting to specific anomaly types
- Ensemble considerations for production deployment
Inference and Testing
Signal Extraction Pipeline
- Input: Raw detector data with unknown anomalies
- Normalization and segmentation
- Forward pass through encoder
- Bidirectional LSTM processing
- Decoder reconstruction
- Inverse normalization
- Output: Cleaned signal estimate
Evaluation Metrics
- Waveform overlap (match filter between true and recovered)
- Phase error (critical for parameter estimation)
- Amplitude recovery accuracy
- SNR improvement ratio
- Comparison benchmarks: ideal-case models, traditional methods
Results
Performance Under Ideal Conditions
Baseline Validation
- Match with state-of-the-art denoising models
- High fidelity signal reconstruction (overlap >95%)
- Phase errors <0.1 radians
- Amplitude recovery within 5%
- Confirms architecture effectiveness without anomalies
Data Gap Handling
Single Gap Scenarios
- Successfully bridges data gaps up to multiple hours
- Reconstruction using bilateral context
- Minimal degradation in gap regions
- Smooth transitions at gap boundaries
Multiple Gaps
- Maintains performance with multiple interruptions
- Cumulative gap duration up to significant fraction of observation
- Coherent signal reconstruction across fragmented data
Glitch Mitigation
Transient Artifact Removal
- Effective suppression of various glitch morphologies
- Preserves underlying gravitational wave signal
- Minimal residual artifacts
- Works for overlapping signal and glitch
Glitch Characteristics Tested
- Short-duration transients (milliseconds to seconds)
- Various amplitudes relative to signal
- Different frequency content
- Multiple glitches per observation
Time-Varying Noise Adaptation
Varying Autocorrelation
- Adapts to changing noise properties
- Maintains signal recovery despite non-stationarity
- No retraining required for different noise realizations
Comparison to Stationary Assumptions
- Traditional methods assuming stationary noise degrade
- Deep learning maintains robust performance
- Adaptive feature extraction critical
Combined Non-Stationarity Performance
Multiple Simultaneous Anomalies
- Tested with gaps + glitches + varying noise
- Remarkable adaptability to complex scenarios
- Performance degradation minimal compared to ideal case
- Critical validation for realistic mission conditions
Statistical Performance
- Consistent results across parameter space
- No catastrophic failures identified
- Graceful degradation with increasing anomaly severity
- Reliable for operational use
Signal-to-Noise Ratio Dependence
High SNR (>30)
- Excellent recovery even with severe non-stationarities
- Phase and amplitude highly accurate
Moderate SNR (10-30)
- Robust performance maintained
- Slight increase in errors but remains unbiased
- Practical operating regime for LISA/Taiji
Low SNR (<10)
- Challenging but functional
- Higher uncertainty but no systematic biases
- Suitable for detection followed by refined analysis
Impact
For Space-Based Gravitational Wave Missions
Operational Reliability
- Ensures science operations can continue despite anomalies
- Reduces data loss from non-stationary periods
- Enables use of full mission dataset
- Critical for maximizing return on investment
Mission Planning
- Informs maintenance scheduling strategies
- Provides confidence for handling unexpected events
- Supports risk assessment for operations
- Enables more aggressive science timelines
Data Quality Assurance
- Automated quality control through robust denoising
- Reduces need for manual intervention
- Consistent processing across mission duration
- Facilitates reproducible science results
For Deep Learning in Gravitational Wave Astronomy
Robustness Validation
- First comprehensive study of neural network robustness to realistic anomalies
- Demonstrates practical viability beyond controlled conditions
- Establishes testing protocols for future models
- Builds confidence for operational deployment
Architecture Insights
- Bidirectional LSTM effectiveness for handling gaps
- Autoencoder framework for signal preservation
- Feature learning robust to various disturbances
- Design principles applicable to other missions/sources
Methodology Development
- Training strategies for robustness
- Evaluation metrics for anomalous data
- Benchmark datasets for comparison
- Best practices for deployment
For Massive Black Hole Binary Science
Enhanced Detection Confidence
- Reliable signal extraction despite non-ideal data
- Reduced false positives from artifact confusion
- Improved parameter estimation accuracy
- Enables low-latency alerts for multi-messenger follow-up
Population Studies
- Full mission dataset usable for statistics
- Unbiased sample for astrophysical inference
- Maximum event discovery potential
- Support for cosmological applications
Resources
Publication
- Journal: Physics Letters B (2024)
- DOI: 10.1016/j.physletb.2024.139016
- arXiv: arXiv:2402.13091 [astro-ph, physics:physics]
Authors
- Yuxiang Xu
- Minghui Du
- Peng Xu
- Bo Liang
- He Wang
Background on Space-Based Detectors
LISA (Laser Interferometer Space Antenna)
- ESA/NASA collaboration
- Three spacecraft in heliocentric orbit
- Arm length: 2.5 million km
- Frequency band: 0.1 mHz - 1 Hz
Taiji
- Chinese mission concept
- Similar configuration to LISA
- Complementary sensitivity
- Timeline: 2030s
TianQin
- Chinese mission
- Geocentric orbit configuration
- Higher frequency focus
- Verification binaries
Non-Stationarity in Space Missions
Causes of Data Gaps
- Planned communication downlinks
- Spacecraft repointing maneuvers
- Calibration activities
- Solar conjunction periods
Sources of Glitches
- Micrometeoroid impacts
- Outgassing events
- Thermal transients
- Electronic anomalies
Time-Varying Noise Sources
- Solar wind variations
- Thermal environment changes
- Component aging
- Orbital phase dependencies
Related Deep Learning Work
Denoising Methods
- WaveFormer: Transformer-based approach
- Various autoencoder architectures
- Generative adversarial networks
- Diffusion models for signal processing
Robustness Studies
- Adversarial examples in physics
- Transfer learning across conditions
- Domain adaptation techniques
- Uncertainty quantification
Software and Tools
Deep Learning Frameworks
- PyTorch/TensorFlow implementations
- LSTM and recurrent architectures
- Time-series processing libraries
Gravitational Wave Tools
- LISA analysis software
- Waveform generation tools
- Noise simulation packages
- Data handling utilities
Future Directions
Methodological Extensions
- Attention mechanisms for improved gap handling
- Uncertainty estimation for reconstruction confidence
- Active learning for identifying challenging cases
- Transfer learning across missions
Additional Anomaly Types
- Laser frequency variations
- Pointing jitter
- Temperature excursions
- Novel instrumental artifacts
Multi-Source Scenarios
- Denoising with overlapping signals
- Confusion noise mitigation
- Global fit support
- Population-level robustness
Operational Integration
- Real-time processing pipelines
- Automated anomaly detection
- Quality flagging systems
- Mission operations support
Extended Applications
- Extreme mass ratio inspirals
- Galactic binary extraction
- Stochastic background analysis
- Other source types and missions