Deep learning based real-time inference and parameter inversion system for gravitational wave events

Abstract

This application discloses a method and system for real-time inference and parameter inversion of gravitational wave events based on deep learning. The method includes: constructing a physics-instrument dual-manifold decoupled network; extracting physics signal features and instrument noise features from time-frequency spectra; introducing physical manifold constraints and orthogonally decoupling the features in the latent space; performing parameter inversion on the decoupled physics features using a conditionally normalized flow model; and verifying the consistency of the inference results by regenerating the waveform in reverse and analyzing the residuals. This application distinguishes between high-eccentricity gravitational wave signals and detector transient noise, improving recognition accuracy and reducing false alarm rate.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of artificial intelligence technology and astrophysics, specifically to a method and system for real-time inference and parameter inversion of gravitational wave events based on deep learning. Background Technology

[0002] The rapid development of gravitational wave astronomy, especially with the continuous improvement of the sensitivity of detectors such as the Laser Interferometer Gravitational-Wave Observatory (LIGO), has led to a significant increase in the detection rate of gravitational wave events and a richer variety of detected signals. Among numerous gravitational wave sources, the merger of high-eccentricity binary black holes formed by dynamical trapping channels has become a key target for verifying general relativity and understanding the formation environment of compact objects due to its unique physical processes. Before merging, high-eccentricity binary black hole systems undergo multiple near-point passes, each producing pulsed gravitational wave radiation, which provides unique signal characteristics for their detection.

[0003] However, traditional gravitational wave data processing methods, such as matched filtering, face severe computational challenges when processing high eccentricity signals. Eccentricity, as a new degree of freedom, significantly expands the parameter space of gravitational wave signals, leading to an exponential increase in the computational resources required to construct a complete theoretical waveform template library, which cannot meet the needs of real-time or near-real-time data analysis. Therefore, utilizing deep learning techniques for real-time inference and parameter inversion of gravitational wave events has become an important development direction in this field.

[0004] While deep learning methods have shown promise in processing gravitational wave signals from quasi-circular orbits, they face an unresolved technical challenge when applied to high eccentricity signals: the "morphological degeneracy" between physical signals and instrument noise. Specifically, strain data from ground-based gravitational wave detectors commonly contain non-Gaussian, non-stationary transient noise (glitches). One type of interference, known as short-duration pulse noise (Blip Glitch), manifests in the time-frequency spectrum as short, wide-bandgap energy pulses. This noise morphology is visually and mathematically highly similar to the pulsed gravitational wave radiation characteristics produced near the periapsis of high-eccentricity binary black hole systems. Existing deep learning models often confuse with this morphological degeneracy. For example, the model might misidentify a real physical signal pulse as instrument noise and suppress it, leading to signal loss; or conversely, it might misidentify a short-duration pulse noise as a feature of a high-eccentricity signal, resulting in incorrect astrophysical parameter inferences. Existing technical solutions, such as using a separate denoising network to clean the data before parameter inference, often suffer from signal details similar to noise due to the non-ideal nature of the denoising process. Another approach is to classify noise as a separate category, but this ignores the potential manifold overlap between signal and noise features in the latent feature space, failing to achieve refined feature decoupling. Therefore, in real-time early warning applications requiring millisecond-level latency, effectively distinguishing this time-frequency morphological degeneracy caused by both astrophysical laws and instrumental limitations, based solely on limited single-detector or short-time-window data, is a pressing technical problem in the field of gravitational wave data processing. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the problem of low accuracy and high false alarm rate of deep learning models when performing real-time parameter inversion due to the morphological degeneracy of high eccentricity gravitational wave signals and detector transient noise (especially short-time impulse noise (Blip Glitch)) in time-frequency representation.

[0006] To address the aforementioned technical problems, this invention provides a method for real-time inference and parameter inversion of gravitational wave events based on deep learning, comprising the following steps: Step 1: Construct a time-frequency feature dataset, which contains mixed data generated by mixing theoretical waveforms and transient noise samples, and perform multi-scale time-frequency preprocessing on the mixed data to generate input feature tensors; Step 2: The input feature tensor is input into a physics-instrument dual-flow manifold decoupling network for processing. Candidate physical feature vectors Z_raw are extracted through the physical constraint flow in the physics-instrument dual-flow manifold decoupling network, and latent noise features Z_instr are extracted through the instrument perception flow. The physical constraint flow is driven by an orbital dynamics manifold projector, and Riemannian geometric constraints are introduced during training. The Riemannian geometric constraints map the Fisher information matrix of the gravitational wave physics parameter space to the Riemannian metric tensor of the latent space of the physical constraint flow. Step 3: In the latent space, the candidate physical feature vector Z_raw and the latent noise feature Z_instr are explicitly orthogonalized. This includes minimizing mutual information through adversarial orthogonal loss and performing orthogonalization operations through a feature orthogonalization layer to remove the projection component of the latent noise feature Z_instr in the feature space direction formed by the candidate physical feature vector Z_raw, thereby obtaining the decoupled physical feature vector Z_phys. Step four: Based on the decoupled physical feature vector Z_phys, Bayesian parameter inversion is performed using the conditional normalized flow model to output the posterior probability distribution of the astrophysical parameters to be estimated.

[0007] Furthermore, in step one, the theoretical waveform is the theoretical waveform of a high-eccentricity binary black hole merger, and the transient noise sample is the Blip Glitch sample extracted from the observation data of the gravitational wave detector; the multi-scale time-frequency preprocessing adopts multi-resolution Q transform or continuous wavelet transform.

[0008] Furthermore, the Riemann geometric constraint in step three specifically involves adding a constraint term to the training loss function of the physics-instrument dual-flow manifold decoupling network, which forces the Riemann metric tensor of the latent space to be proportional to the Fisher information matrix of the physical parameter space, wherein the physical parameter space contains the mass, spin, and eccentricity parameters of the binary black holes.

[0009] Optionally, the instrument sensing stream is constructed using a dilated convolutional network.

[0010] Furthermore, the conditional normalized flow model in step four specifically involves: using the physical feature vector Z_phys as a conditional input, and mapping the baseline probability distribution to the target posterior probability distribution of the astrophysical parameters through a neural network composed of a series of invertible transformations. The neural network is based on a neural spline flow or masked autoregressive flow architecture.

[0011] Preferably, the astrophysical parameters to be estimated include the component masses, effective spin, and eccentricity of the binary black holes.

[0012] Furthermore, the deep learning-based real-time inference and parameter inversion method for gravitational wave events also includes: performing an instant consistency check based on reverse regeneration, wherein the check specifically involves: using a decoder corresponding to the orbital dynamics manifold projector to reconstruct the estimated astrophysical parameters inferred in step four and the corresponding physical feature vector Z_phys into a time-domain gravitational wave waveform; calculating the residual sequence between the original input data and the reconstructed time-domain gravitational wave waveform; and calculating a physical consistency score based on the structure of the residual sequence.

[0013] Furthermore, the instant consistency check also includes: comparing the physical consistency score with a predetermined threshold, and when the physical consistency score exceeds the predetermined threshold, confirming the current event as a high-confidence gravitational wave event.

[0014] Furthermore, the deep learning-based real-time inference and parameter inversion method for gravitational wave events also includes: deploying the trained physics-instrument dual-flow manifold decoupling network and the conditional normalized flow model with low latency; and integrating an active curriculum learning module, which continuously collects events that are judged as difficult negative samples by the consistency instant check during system operation and automatically adds them to the training dataset to perform periodic online fine-tuning of the model.

[0015] Another aspect of the present invention provides a real-time inference and parameter inversion system for gravitational wave events based on deep learning, comprising: The time-frequency feature dataset construction unit is used to construct mixed data containing gravitational wave signals and transient noise, and to perform time-frequency preprocessing on it to generate feature tensors; A physics-instrument dual-stream manifold decoupling network unit is used to receive the feature tensor and separate the physics signal features and the instrument noise features from it. This unit internally includes: a physics constraint stream based on a variational autoencoder and Riemannian geometric constraints, wherein the Riemannian geometric constraints establish an equidistant relationship between the latent space geometry of the physics constraint stream and the geometry of the physics parameter space; an instrument sensing stream for capturing noise features; and a feature orthogonalization layer for orthogonalizing the features output by the physics constraint stream and the instrument sensing stream. The Bayesian parameter inversion unit is configured to receive the decoupled physical features output by the physics-instrument dual-flow manifold decoupling network unit, and output the posterior probability distribution of astrophysical parameters using the conditionally normalized flow model.

[0016] The advantages of this invention, achieved by employing the technical solution provided, are as follows: First, by using a physics-instrument dual-flow orthogonal decoupling architecture and introducing physical manifold constraints, it distinguishes between high-eccentricity gravitational wave signals and detector transient noise characteristics, improving the model's accuracy in signal recognition under noisy conditions. Second, the system can output the Bayesian posterior probability distribution of astrophysical parameters, providing a parameter estimation tool for real-time gravitational wave astronomy research. Finally, by introducing a physical consistency-based reverse regeneration verification mechanism, false alarms caused by instrument noise are suppressed, improving the reliability of the gravitational wave early warning system. Attached Figure Description

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating a method for real-time inference and parameter inversion of gravitational wave events based on deep learning, provided in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the physical-instrument dual-flow manifold decoupling network in an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram illustrating the principle of Riemannian geometric constraints used in the physical constraint flow in this embodiment of the invention.

[0021] Figure 4 This is a structural block diagram of a real-time inference and parameter inversion system for gravitational wave events based on deep learning, provided in an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram of time-frequency spectrum comparison provided in an embodiment of the present invention.

[0023] Figure 6 A visualization of the feature space distribution provided for embodiments of the present invention.

[0024] Figure 7 The Bayesian posterior probability distribution diagram provided for embodiments of the present invention demonstrates the inference results for four key parameters of a binary black hole system.

[0025] Figure 8 The residual sequence comparison diagram provided in the embodiments of the present invention shows the differences between real gravitational wave signals and pure short-time impulse noise (Blip Glitch) in residual analysis.

[0026] Figure 9 The training loss curve of the model provided in the embodiment of the present invention.

[0027] Figure 10 This is a schematic diagram of the physical structure of the LIGO detector provided in an embodiment of the present invention.

[0028] Figure 11 This is a schematic diagram of the orbital evolution of a high-eccentricity binary black hole provided in an embodiment of the present invention.

[0029] Figure 12 This is a schematic diagram of the spacetime distortion propagation of gravitational waves provided in an embodiment of the present invention.

[0030] Figure 13 This is a real-time data processing pipeline topology diagram provided for an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the scope of protection of the invention.

[0032] To further understand the technical principles of this invention, the morphological degeneracy of physical signals and instrument noise addressed by this invention will first be explained in detail. For example... Figure 5 As shown, this invention reveals the morphological differences between high-eccentricity gravitational wave signals and short-duration pulse noise (Blip Glitch) through time-frequency analysis. The left subfigure illustrates the typical time-frequency evolution of high-eccentricity gravitational wave signals: three near-star pulses appear within the 0-4 second observation window, with the pulse interval decreasing from 0.9 seconds to 1.1 seconds, the amplitude increasing progressively, and the center frequency rising from 80 Hz to 180 Hz, reflecting the period shortening and frequency increase characteristics caused by orbital energy loss; after 3.0 seconds, it enters the merging phase, with the frequency rapidly sweeping from 200 Hz to 400 Hz, forming a characteristic chirp signal. The middle subfigure illustrates the time-frequency characteristics of short-duration pulse noise (Blip Glitch): a single short pulse appears around 2.0 seconds, with a time width of approximately 0.08 seconds and a frequency coverage as wide as 180 Hz (center frequency 220 Hz). This pulse does not exhibit the physical evolution of gravitational wave signals, but rather manifests as a sudden, broadband energy release. The right-hand subplot illustrates a mixed data scenario: when the pulses of a gravitational wave signal are superimposed with short-duration glitch noise in the time-frequency domain, it is difficult to distinguish the signal from the noise based solely on energy intensity. This morphological degeneracy problem is the key technical challenge that this invention aims to solve. By comparing the time-frequency heatmaps of the three subplots, it can be clearly seen that the gravitational wave signal exhibits the evolutionary continuity of multiple pulses and chirp characteristics, while the short-duration glitch noise only manifests as an isolated single pulse. The physics-instrument dual-flow manifold decoupling network of this invention achieves effective identification based on these differences in time-frequency domain characteristics.

[0033] Example 1 Please refer to Figure 1This invention provides a method for real-time inference and parameter inversion of gravitational wave events based on deep learning. This method can be applied to servers or embedded computing devices in gravitational wave data processing centers. Specifically, the method may include the following steps: Step S100: Construct a time-frequency feature dataset containing high-eccentricity gravitational wave signals and transient noise.

[0034] like Figure 10 As shown, the LIGO (Laser Interferometer Gravitational-Wave Observatory) detector adopts an L-shaped interferometer configuration and is the main source of gravitational wave signal data in this invention. The physical structure of the detector mainly includes the following key components: a laser source (1) generates a highly stable laser beam, which is split into two paths by a beam splitter (3) and propagates along two vacuum arms that are perpendicular to each other at 90 degrees, each arm being 4 kilometers long; a high-precision reflector (4, 5) is set at the end of each arm, and the laser beam returns to the beam splitter after going back and forth within the arm multiple times to form interference; finally, the interference light signal is received by a photodetector (2). When a gravitational wave passes by, it will produce a small spacetime distortion in two perpendicular directions, causing a difference in the length of the two arms, thereby causing the interference fringes to move. The detector captures the gravitational wave signal by precisely measuring this small change (the sensitivity can reach the order of 10^-21 meters). The entire system is placed in an ultra-high vacuum environment to eliminate the influence of air disturbance on the optical path. The signal samples in the time-frequency feature dataset constructed in this invention originate from the observation data of such detectors, including gravitational wave event signals and various instrument noise samples. This step is used to provide high-quality, diverse, and challenging data for subsequent neural network training.

[0035] Specifically, this step first generates the theoretical waveform of a high-eccentricity binary black hole merger. For example... Figure 11The diagram illustrates the complete evolution of a high-eccentricity binary black hole system from its initial elliptical orbit to its final merger. Stage 1 shows the initial high-eccentricity orbit (e approximately 0.8), where the two black holes are closest at their perihelion, resulting in the strongest gravitational wave radiation. As energy is continuously lost through gravitational wave radiation, the system enters Stage 2, a process of orbital decay. The elliptical orbit gradually shrinks, the eccentricity decreases to approximately 0.5, and gravitational wave radiation continues to increase. Finally, in Stage 3, the system evolves into a quasi-circular orbit (e approaching 0), the two black holes rapidly approach and eventually merge, releasing a powerful burst of gravitational waves in the process. The diagram clearly illustrates the gravitational wave radiation characteristics through a pictographic representation of the black holes (black core plus accretion disk) and concentric ripples, revealing the physical mechanisms of orbital circularization and energy dissipation in high-eccentricity systems, providing an intuitive physical picture for understanding the gravitational wave signal characteristics of such systems. Preferably, a waveform model based on the effective single-hole approximation theory, such as the SEOBNRv4EHM model, can be used to generate signal samples covering a wide range of astrophysical parameters. The parameter space can be set to a range where the total mass of the binary black holes is between 20 and 200 times the mass of the Sun, the mass ratio is between 1 and 10, and the effective spin parameter is between -0.9 and 0.9. Specifically, to fully characterize the properties of high eccentricity signals, high-density and uniform sampling is performed on the eccentricity parameter at a 10 Hz reference frequency within the range of 0 to 0.9. This ensures that the generated signal sample library includes various evolutionary forms from quasi-circular orbits to extremely high eccentricity orbits, and incorporates pulsed gravitational wave radiation characteristics generated by multiple near-star transits.

[0036] Simultaneously, this step also constructs a high-fidelity transient noise library. Specifically, from historical observation data of real gravitational wave detectors, such as publicly available data from LIGO or Virgo detectors, a large number of short-duration pulse noise (Blip Glitch) samples are screened and extracted using professional noise identification tools. These samples, in the time-frequency domain, exhibit short-duration, wide-bandwidth energy pulses, highly similar to the near-starburst characteristics of high-eccentricity signals.

[0037] Finally, the generated signal waveform is combined with noise samples via a dynamic injection module. This module injects the signal waveform into a real background noise stream that does not contain significant signals or noise, with a randomly selected signal-to-noise ratio (e.g., a range of 5 to 30) and a random time offset. To improve the model's robustness, this module deliberately constructs challenging samples by adjusting the time delay between the signal pulse and short-time impulse noise (Blip Glitch) samples to make them temporally adjacent (e.g., with a time interval of less than 0.1 seconds) or partially overlapping, thereby simulating the highly challenging overlap scenarios that may be encountered in real-world data processing.

[0038] Step S200: Perform multi-scale time-frequency preprocessing on the mixed data generated in step S100 to generate feature tensors suitable for neural network input.

[0039] While one-dimensional time-series strain data contains all the information, its features are not intuitive enough, hindering direct learning by deep neural networks. Therefore, this step uses time-frequency analysis tools to convert it into a two-dimensional time-frequency spectrum. Specifically, a multi-resolution Q-transform (CQT) can be used. The time and frequency resolutions of the Q-transform adaptively adjust with frequency, exhibiting high frequency resolution in the low-frequency band and high time resolution in the high-frequency band, which perfectly matches the frequency sweeping characteristics of gravitational wave signals. For example, multiple different Q values ​​can be set, Q-transforms can be performed separately, and the resulting multiple time-frequency spectra can be stacked as different channels to form a multi-channel feature tensor, which serves as the input to the subsequent network. Alternatively, a continuous wavelet transform can be used, such as with the Morlet wavelet basis, which can also yield a time-frequency spectrum that preserves the localized structural information of time and frequency.

[0040] Step S300: Construct a physical-instrument dual-flow manifold decoupling network to separate physical signal features and instrument noise features from the input time-frequency spectrum.

[0041] Please refer to the reference. Figure 2 This network employs a physical-instrument dual-manifold orthogonal decoupling architecture, aiming to address the morphological degeneracy problem of signal and noise. The network mainly consists of a physical constraint flow 201 and an instrument sensing flow 202.

[0042] Specifically, the goal of Physical Constraint Flow 201 is to extract and characterize pure astrophysical signal features. Its core component is an orbital dynamics manifold projector. This projector can be structurally implemented using a Variational Autoencoder (VAE), and this VAE is pre-trained only on the pure theoretical waveform data generated in step S100, which is free of any noise. Through this pre-training method, its latent space is forced to learn and encode the high-eccentricity orbital dynamics evolution laws determined by general relativity. The projector outputs a candidate physical feature vector Z_raw, whose corresponding latent space can be viewed as a low-dimensional embedding of the physical parameter space in a high-dimensional feature space, i.e., a physical manifold. Any data that does not conform to the evolution laws of general relativity (such as noise) cannot be effectively projected onto this manifold.

[0043] Furthermore, to ensure that the physical manifold remains consistent with the real physical parameter space not only in its topology but also in its geometric metrics, this invention introduces Riemannian geometric constraints when constructing the physical constraint flow 201. Please refer to... Figure 3The physical parameter space θ (e.g., containing parameters such as mass, spin, and eccentricity) itself possesses an intrinsic geometric structure defined by the Fisher Information Matrix (FIM). Here, h is the gravitational wave waveform, θ_i and θ_j are physical parameters, and angle brackets represent noise-weighted inner products. This matrix describes the degree to which small changes in the parameters affect the waveform signal. This invention adds a constraint term to the loss function during network training, forcing the Riemannian metric tensor of the latent space to be proportional to the Fisher information matrix of the physical parameter space. This is equivalent to imposing an equidistant constraint, ensuring a strict proportional relationship between the geodesic distance d_geo(z1, z2) 303 between any two points z1 and z2 in the latent space and the mismatch 304 between the corresponding waveforms h1 and h2 in the physical parameter space. This deep geometric constraint makes the network extremely sensitive to noise disturbances that deviate from physical laws.

[0044] Specifically, the goal of the instrument sensing stream 202 is to specifically capture and characterize the non-stationary noise features of the detector. Considering that noise such as short-time impulse noise (Blip Glitch) can have complex textures and large spatial structures in its time-frequency spectrum, this stream can be constructed using a dilated convolutional network with a large receptive field. Dilated convolution can effectively expand the receptive field without increasing computational cost, thereby better capturing the global and contextual features of the noise. The output of this stream is the latent noise feature Z_instr.

[0045] To achieve feature decoupling between the two flows, this invention introduces adversarial orthogonal loss during network training and sets up a feature orthogonalization layer 205. The adversarial orthogonal loss uses an auxiliary discriminator network to determine whether there is a correlation between Z_raw and Z_instr; the physical constraint flow and the instrument perception flow are trained to generate independent features, thereby deceiving the discriminator. Furthermore, the feature orthogonalization layer 205 performs explicit orthogonalization processing after the two latent feature vectors are generated. Specifically, the projection components of the latent noise feature Z_instr onto the feature space direction formed by the candidate physical feature vector Z_raw are removed to obtain the decoupled physical feature vector Z_phys. Figure 6 As shown, this invention uses the feature space distribution to visually demonstrate the significant difference in effects before and after decoupling. Figure 6The left subplot illustrates the feature aliasing state before decoupling. A t-distributed neighborhood embedding (t-SNE) dimensionality reduction technique was used to map the high-dimensional features to a two-dimensional space. The black circles represent approximately 500 physical signal samples, and the dark gray triangles represent approximately 500 noise samples. This subplot clearly shows that before decoupling, the two types of samples exhibit severe overlap and aliasing in the feature space, with sample points intertwined and blurred boundaries. Figure 6 The right-hand subplot illustrates the feature distribution after optimization using the orthogonal decoupling loss function. After decoupling, the physical feature Z_phys and the noise feature Z_instr achieve clear orthogonal separation in the feature space: the physical feature clusters cluster in the left region, and the noise feature clusters cluster in the right region, with a clear boundary between the two clusters. The arrows in the figure indicate the orthogonal separation direction between the two clusters, fully demonstrating that the orthogonal decoupling mechanism of this invention effectively separates the essential features of the physical signal from the interference features of instrument noise into different representation subspaces. In this way, it is ensured that the final physical feature Z_phys and instrument feature Z_instr are mutually orthogonal in the latent space, representing two completely decoupled pieces of information.

[0046] Step S400: Based on the conditional normalized flow model, Bayesian parameter inversion is performed on the decoupled pure physical features.

[0047] The pure physical feature vector Z_phys, output from the physical constraint flow 201 in step S300 and verified by the physical manifold and orthogonalized by noise features, is used as conditional input and fed into a conditional normalized flow network. The normalized flow model is a generative model consisting of a series of invertible transformations, capable of accurately mapping a simple baseline probability distribution (e.g., a standard normal distribution) to a complex target distribution. In this invention, the target distribution is the posterior probability distribution p(θ|d) of the astrophysical parameter θ given observational data d (whose information is carried by Z_phys). Preferably, an architecture based on neural spline flow or masked autoregressive flow can be used due to its strong distribution fitting capability. The output layer of this network directly corresponds to the astrophysical parameters to be estimated, such as the component masses, effective spin, and key eccentricity parameters of binary black holes. By sampling from the target distribution generated by this network, the complete posterior probability distribution of these parameters can be obtained, thus achieving fast and comprehensive Bayesian inference of the physical parameters of gravitational wave events. Figure 7 As shown, this invention accurately estimates the key parameters of a binary black hole system through Bayesian inference. Figure 7A 2×2 grid layout was used to illustrate the posterior probability distributions of the four parameters. Subplot (a) shows the posterior distribution of total mass M, with the horizontal axis representing the total mass range of 50-100 solar masses and the vertical axis representing the probability density p(M|d). The peak of the distribution is located at approximately 75 solar masses. Dashed lines indicate the true values, and gray shading represents the 90% confidence interval. Subplot (b) shows the posterior distribution of mass ratio q, with the horizontal axis ranging from 0.1 to 1.0, the peak value around 0.7, and a 90% confidence interval of 0.55-0.85. Subplot (c) shows the posterior distribution of effective spin χeff, with the horizontal axis ranging from -0.5 to 0.8, and a peak value around 0.3. Subplot (d) shows the posterior distribution of eccentricity e, with the horizontal axis ranging from 0.4 to 0.9, the peak value around 0.65, and a 90% confidence interval of 0.52-0.78. All subplots use solid lines to represent the posterior distribution curves, dashed lines to indicate key reference values, and gray shading to represent confidence intervals. By analyzing the posterior distribution of these four parameters, we can comprehensively assess the accuracy and reliability of the parameter estimation, providing a quantitative basis for the physical interpretation of gravitational wave signals.

[0048] Step S500: Perform an immediate consistency check based on reverse regeneration to improve the reliability of the alert.

[0049] To further suppress false alarms caused by noise in the real-time processing flow, this invention also introduces a physical consistency-based reverse regeneration verification mechanism. Specifically, using a decoder (which corresponds to the encoder of the physical constraint flow 201 in step S300), the expected value θ_hat of the posterior distribution of the physical parameters inferred in step S400 and the corresponding latent feature Z_phys are used to reconstruct the gravitational wave waveform h_reconstructed in the time domain. Subsequently, the residual sequence between the original input data d_input and the reconstructed waveform h_reconstructed is calculated: r = d_input - h_reconstructed.

[0050] If the input data d_input contains real gravitational wave signals, its main components will be accurately captured and reconstructed by the model because the signal follows physical manifold constraints. Therefore, the resulting residual sequence should be close to Gaussian white noise. Conversely, if the input data d_input is only Blip Glitch, it cannot be effectively reconstructed by the physical model because it does not follow the physical manifold. The resulting residual sequence will retain its significant non-random, non-Gaussian structure. A "physical consistency score" is calculated to quantify the residual structure; for example, the chi-square statistic of the residual sequence can be calculated and compared with the expected value of a Gaussian distribution to obtain a normalized score. Then, a predetermined threshold is set. Only when the physical consistency score exceeds this threshold is the event confirmed as a high-confidence gravitational wave event, triggering subsequent early warning procedures. Figure 8As shown, the present invention verifies the effectiveness of the signal-noise discrimination method by comparing residual sequences. Figure 8 The experiment comprises two sets of comparative experiments: the upper subplot shows the actual gravitational wave signal, with the black waveform representing the reconstructed waveform from the VAE model, which closely matches the input data. The bottom residual sequence exhibits typical Gaussian white noise characteristics, proving that VAE successfully extracted the principal components of the signal. The lower subplot shows the pure Blip Glitch noise situation, where VAE failed to capture its rapid oscillation characteristics. The bottom residual sequence retains a clear non-Gaussian structure, and the three Blip events are clearly visible in the residuals. This comparative experiment verifies the core principle of this invention: the residuals of the actual gravitational wave signal exhibit Gaussian noise characteristics, while the residuals of pure Blip Glitch noise retain a non-Gaussian structure. By quantifying the statistical characteristics of the residuals, effective differentiation between gravitational wave signals and transient noise can be achieved.

[0051] Step S600: Deploy the trained system with low latency and integrate an online learning mechanism.

[0052] The trained neural network model from the above steps is then processed using model optimization and quantization tools to reduce its computational complexity and memory footprint. For example, high-performance deep learning inference software development kits (SDKs), such as TensorRT (Tensor Real-Time Inference Accelerator), can be used for graph optimization and precision quantization, or it can be converted to the Open Neural Network Exchange (ONNX) format to enable cross-platform deployment of the deep learning model across different frameworks. The optimized model is deployed on the front-end server or dedicated hardware of the gravitational wave data acquisition system to build a low-latency processing pipeline. This pipeline employs a sliding window mechanism to overlap slices of continuous streaming data (e.g., a window length of 4 seconds and a step size of 1 second), and performs real-time inference on each slice, ensuring that the latency of the entire processing flow is controlled within milliseconds.

[0053] Simultaneously, to enable the system to adapt to the long-term evolution of the detector's noise environment, this invention also integrates an active curriculum learning module. During system operation, this module continuously collects events that are identified as difficult negative samples by the consistency check module in step S500 (i.e., noisy events with low physical consistency scores but high initial network output confidence). Figure 9The figure illustrates the convergence characteristics of each loss function term in the training process of the gravitational wave signal detection model of this invention. The X-axis represents the training epoch, ranging from 0 to 200 epochs; the Y-axis uses logarithmic coordinates to represent the loss value. The total loss curve (solid black line) shows a rapid decreasing trend, with the physical reconstruction loss (dashed line), adversarial orthogonal loss (dotted line), and Riemannian geometric constraint loss (dotted line) working together to optimize the model. At epoch 150, the loss terms tend to stabilize, indicating that the model has completed effective training and verifying the stable convergence of the training strategy. These samples are automatically added to a dedicated hard sample database and periodically used for online fine-tuning of the deployed model. This online learning mechanism enables the system to adapt, continuously improving its performance under constantly changing noise conditions.

[0054] Example 2 Please refer to Figure 4 This invention also provides a real-time inference and parameter inversion system for gravitational wave events based on deep learning, which is a device-based embodiment of the above-described method embodiments. For example... Figure 13 As shown, the real-time data processing pipeline of this invention adopts a layered topology architecture, encompassing a data acquisition end (including the LIGO detector 1, analog-to-digital converter (ADC) sampling module 2, and data buffer 3), a preprocessing layer (including a sliding window slicing unit 4, and a graphics processing unit (GPU) cluster 5-7), an inference layer (including a dual-stream decoupled network inference server 8, a parameter inversion calculation node 9, and a consistency verification unit 10), and an output end (including an early warning decision module 11, an observatory notification system 12, and a data archive 13). The end-to-end latency of the entire pipeline from data acquisition to early warning output does not exceed 150 milliseconds, meeting the stringent timeliness requirements for real-time gravitational wave detection. The system may include: The time-frequency feature dataset construction unit 401 is used to generate mixed data containing gravitational wave signals and transient noise, and to perform time-frequency preprocessing on it to generate feature tensors. The specific functions and implementation of this unit can be found in steps S100 and S200 of Embodiment 1, and will not be repeated here.

[0055] The physics-instrument dual-flow manifold decoupling network unit 402 receives the feature tensor generated by the time-frequency feature dataset construction unit 401 and separates the pure physics signal features and instrument noise features from it. This unit internally includes a physics constraint flow based on variational autoencoders and Riemannian geometric constraints, and an instrument sensing flow for capturing noise features. Feature decoupling is achieved through adversarial orthogonal loss and a feature orthogonalization layer. The specific structure, function, and implementation of this unit can be found in step S300 of Embodiment 1. Figure 2 , Figure 3 As shown, it will not be elaborated further here.

[0056] The Bayesian parameter inversion unit 403 is configured to receive the decoupled, pure physical features output by the physics-instrument dual-flow manifold decoupling network unit 402, and quickly output the posterior probability distribution of the astrophysical parameters to be estimated using a conditionally normalized flow model. The specific function and implementation of this unit can be found in step S400 of Embodiment 1, and will not be repeated here.

[0057] The consistency verification unit 404 is used to calculate the residual between the original input and the reconstructed waveform by regenerating the waveform in reverse, and to calculate the physical consistency score based on the structure of the residual, thereby verifying the physical authenticity of the inference result in real time and effectively reducing the false alarm rate. The specific function and implementation of this unit can be found in step S500 of Embodiment 1, and will not be repeated here.

[0058] The system deployment and online learning unit 405 is used to optimize and quantize the trained model and deploy it in a low-latency processing pipeline. It also integrates an active curriculum learning module to collect difficult samples and fine-tune the model online, enabling adaptive learning to new data and changing noisy environments. The specific functions and implementation of this unit can be found in step S600 of Embodiment 1, and will not be repeated here.

[0059] In summary, this invention successfully breaks the morphological degeneracy of high-eccentricity gravitational wave signals and transient noise in the latent feature space by constructing a deep neural network that incorporates a physical-instrument dual-manifold orthogonalization mechanism and introducing physical manifold constraints based on Riemannian geometry. It combines conditional normalization flow for fast Bayesian parameter inversion and improves the reliability of early warning through consistency checks of inverse regeneration. Ultimately, it achieves millisecond-level real-time and accurate inference and parameter inversion for high-eccentricity gravitational wave events.

[0060] Example 3 This embodiment aims to illustrate the processing flow of the method of the present invention in a specific application scenario, so as to demonstrate the necessity of adopting the method of this application. This scenario processes a 4-second segment of detector strain data, which includes a high-eccentricity binary black hole merger signal with a signal-to-noise ratio of 15. In this application scenario, the processing difficulty lies in the fact that the third near-starburst pulse of the signal is temporally adjacent to a short-duration blip glitch with a highly similar morphology, with a time interval of less than 50 milliseconds, causing the characteristic parts of the two to overlap in the time-frequency spectrum.

[0061] For this type of data, conventional methods in existing technologies are difficult to process effectively. If a separate denoising network is used for preprocessing, the high similarity in time-frequency morphology between signal pulses and noise pulses may lead the denoising process to mistakenly suppress real signal pulses as noise, thereby damaging or removing the real signal. If a classification model that treats noise as an independent category is used, when faced with situations where signal and noise features coexist closely, the model often misattributes noise features to the signal due to confusion during the feature extraction stage. This results in serious deviations in the inference results of astrophysical parameters (especially eccentricity and mass ratio), and may even reduce the confidence of the entire event, causing signal loss.

[0062] The specific process of processing this data segment using the system and method of this invention is as follows. First, the data segment is input into the system, which performs a multi-resolution Q-transform on it to generate a multi-channel time-frequency feature tensor. This feature tensor is then fed into a physics-instrument dual-flow manifold decoupling network. In the network, the physics constraint flow, based on the orbital dynamics manifold learned during the pre-training phase, can identify pulse sequence features that conform to the evolution laws of general relativity, i.e., physical patterns where the pulse interval shortens over time and the amplitude changes accordingly. Since the occurrence of this short-duration impulse noise (Blip Glitch) is isolated and does not follow the physical orbit decay law, its features cannot be effectively projected onto the physical manifold defined by the orbital dynamics law, thus its contribution to the physical feature vector Z_phys is effectively suppressed. At the same time, the instrument sensing flow, through its specially learned noise statistics characteristics, successfully captures the non-stationary features of this short-duration impulse noise (Blip Glitch) and encodes it into the instrument feature vector Z_instr. Through feature orthogonalization processing, the system finally generates the potential physical features Z_phys representing the pure gravitational wave signal.

[0063] Subsequently, only the pure physical features Z_phys are passed to the Bayesian parameter inversion unit. Since Z_phys is no longer contaminated by immediate neighbor noise, this unit, based on a conditionally normalized flow model, can quickly and accurately infer the posterior probability distribution of the event, obtaining estimated physical parameters, such as a total mass of 75 solar masses and an eccentricity of 0.65. Finally, the system performs an immediate consistency check. The theoretical waveform is reconstructed using the inferred parameter mean, and this waveform is subtracted from the original input data. The calculated residual sequence clearly shows the structure of background noise and the successfully extracted short-duration impulse noise (Blip Glitch). Because the reconstructed waveform highly matches the physical signal portion in the original data, and the physical consistency score remains above the preset threshold, the system ultimately confirms the event as a high-confidence gravitational wave signal and outputs accurate parameters.

[0064] The results of this embodiment demonstrate that the method of the present invention can effectively decouple signal and noise in a complex scenario where signal and morphologically degenerate strong noise coexist in close proximity, enabling accurate parameter inversion and reliability verification. This solves the problems of signal impairment or parameter inference errors in existing methods on such difficult samples, demonstrating the application value of the present invention in improving the reliability of real-time gravitational wave detection.

[0065] The above description is merely a preferred embodiment of the present invention. It should be noted that it is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for real-time inference and parameter inversion of gravitational wave events based on deep learning, characterized in that, Includes the following steps: Step 1: Construct a time-frequency feature dataset, which contains mixed data generated by mixing theoretical waveforms and transient noise samples, and perform multi-scale time-frequency preprocessing on the mixed data to generate input feature tensors; Step 2: The input feature tensor is input into a physics-instrument dual-flow manifold decoupling network for processing. Candidate physical feature vectors Z_raw are extracted through the physical constraint flow in the physics-instrument dual-flow manifold decoupling network, and latent noise features Z_instr are extracted through the instrument perception flow. The physical constraint flow is driven by an orbital dynamics manifold projector, and Riemannian geometric constraints are introduced during training. The Riemannian geometric constraints map the Fisher information matrix of the gravitational wave physics parameter space to the Riemannian metric tensor of the latent space of the physical constraint flow. Step 3: Explicitly orthogonalize the candidate physical feature vector Z_raw and the latent noise feature Z_instr in the latent space; wherein the latent noise feature Z_instr is a feature extracted by the instrument sensing stream; the explicit orthogonalization process includes minimizing mutual information through adversarial orthogonal loss, and performing orthogonalization operations through a feature orthogonalization layer to remove the projection components of the latent noise feature Z_instr in the feature space direction formed by the candidate physical feature vector Z_raw, thereby obtaining the decoupled physical feature vector Z_phys; Step four: Based on the decoupled physical feature vector Z_phys, Bayesian parameter inversion is performed using the conditional normalized flow model to output the posterior probability distribution of the astrophysical parameters to be estimated.

2. The method according to claim 1, characterized in that, In step one, the theoretical waveform is the theoretical waveform of a high-eccentricity binary black hole merger, and the transient noise sample is a short-time impulse noise (Blip Glitch) sample extracted from the observation data of the gravitational wave detector; the multi-scale time-frequency preprocessing adopts multi-resolution Q transform or continuous wavelet transform.

3. The method according to claim 1, characterized in that, The Riemann geometric constraint in step two specifically involves adding a constraint term to the training loss function of the physics-instrument dual-flow manifold decoupling network, which forces the Riemann metric tensor of the latent space to be proportional to the Fisher information matrix of the physical parameter space, wherein the physical parameter space includes the mass, spin, and eccentricity parameters of the binary black holes.

4. The method according to claim 1, characterized in that, The instrument sensing stream is constructed using a dilated convolutional network.

5. A real-time inference and parameter inversion system for gravitational wave events based on deep learning, characterized in that, include: The time-frequency feature dataset construction unit is used to construct mixed data containing gravitational wave signals and transient noise, and to perform time-frequency preprocessing on it to generate feature tensors; A physics-instrument dual-flow manifold decoupling network unit is used to receive the feature tensor and separate the physics signal features and instrument noise features from it. The unit contains a physics constraint flow based on variational autoencoder and Riemann geometric constraints, wherein the Riemann geometric constraints establish an equidistant relationship between the latent space geometry of the physics constraint flow and the geometry of the physics parameter space. An instrument sensing stream for capturing noise features; and a feature orthogonalization layer for orthogonalizing the features output by the physical constraint stream and the instrument sensing stream. The Bayesian parameter inversion unit is configured to receive the decoupled physical feature vector Z_phys output by the physics-instrument dual-flow manifold decoupling network unit, and output the posterior probability distribution of astrophysical parameters using the conditionally normalized flow model.

Images