CN-121997078-A - Gravity lens gravitational wave pairing method and system based on deep learning

CN121997078ACN 121997078 ACN121997078 ACN 121997078ACN-121997078-A

Abstract

The invention discloses a gravitational lens gravitational wave pairing method and a gravitational wave pairing system based on deep learning, wherein the method comprises the steps of obtaining time sequence strain data of a gravitational wave detector, performing frequency domain whitening treatment on the data, converting the data back to a time domain, and intercepting to obtain a whitened strain fragment with a fixed length; inputting the images into a twin encoder, selecting a plurality of other events with highest similarity for each gravitational wave event based on cosine similarity among feature vectors to form a candidate pairing set, calculating waveform cross-correlation by adopting a generalized cross-correlation-phase transformation algorithm to obtain refined similarity scores and relative time delay estimation among event pairs, constructing a weighted graph, and obtaining a final pairing result by solving the maximum weight matching of the graph. In the invention, the twin encoder is utilized to reduce the dimension of the high-dimension time domain data, and the catalog is prevented from being subjected to full quantity pairwise comparison by a coarse screening strategy, so that the searching speed on a large-scale data set is greatly improved.

Inventors

ZHANG QIKAI
ZHANG FAN

Assignees

浙江大学

Dates

Publication Date: 20260508
Application Date: 20260410

Claims (8)

1. The gravitational lens gravitational wave pairing method based on deep learning is characterized by comprising the following steps: Step S1, time series strain data of a gravitational wave detector are obtained, the data are subjected to frequency domain whitening treatment and then are converted back to a time domain, and a whitened strain segment with a fixed length is obtained by interception; S2, inputting the whitened strain segments into a pre-trained twin encoder and mapping the whitened strain segments into low-dimensional feature vectors; Step S3, selecting a plurality of other events with highest similarity for each gravitational wave event based on cosine similarity among feature vectors to form a candidate pairing set; step S4, for each pair of events in the candidate pairing set, calculating waveform cross-correlation by adopting a generalized cross-correlation-phase transformation algorithm to obtain a refined similarity score and a relative time delay estimation between the event pairs; And S5, constructing a weighted graph by taking the gravitational wave event as a node and the refined similarity score as an edge weight, and obtaining a final pairing result by solving the maximum weight matching of the graph.
2. The gravity lens gravitational wave pairing method based on deep learning according to claim 1, wherein in step S2, the twin encoder comprises two one-dimensional convolutional neural network branches with identical structures and shared weights, each branch performs feature extraction on input time domain data through a plurality of convolutional blocks and a pooling layer, and finally outputs a 128-dimensional unit norm feature vector.
3. The gravity lens gravitational wave pairing method based on deep learning according to claim 1, wherein in step S2, a contrast learning mode is adopted in the pretraining process of the twin encoder, a normalized temperature scaling cross entropy loss function is used, different images generated by the same gravitational wave source through the gravitational lens effect are used as positive sample pairs, and signals of different gravitational wave sources are used as negative sample pairs for training.
4. The gravity lens gravitational wave pairing method based on deep learning according to claim 3, wherein the expression of the loss function is: in the formula, And Respectively outputting characteristic vectors of the sample i and the sample j through a projection head of the twin encoder; Representing cosine similarity; is a temperature parameter; Is the positive sample pair number; in order to indicate the function, Index for all samples in the batch; Is the first in the batch The feature vectors of the samples are output by a projection head of the twin encoder; by minimizing the loss function, the positive pairs of samples are brought closer together in feature space and the negative pairs are brought farther apart.
5. The gravity lens gravitational wave pairing method based on deep learning according to claim 1 is characterized in that in step S4, the generalized cross-correlation-phase transformation algorithm is implemented by performing amplitude normalization processing on cross power spectrums of two signals in a frequency domain, calculating a cross-correlation sequence based on normalized phase information, taking a peak value of the sequence as a refined similarity score, and taking a peak value position as a relative time delay estimation between two events.
6. The gravity lens gravitational wave pairing method based on deep learning according to claim 1, wherein in step S5, the construction process of the weighted graph is as follows: all gravitational wave events are regarded as nodes of the graph, and each event corresponds to a unique node in the graph; For any two different event nodes i and j, obtaining a calculated refined similarity score If the score exceeds the preset threshold lambda, establishing an undirected edge between the node i and the node j, and weighting the edge Is arranged as If the score does not exceed the threshold lambda, no edge is established between the nodes; All nodes and edges meeting the condition together form a weighted undirected graph Wherein For a set of nodes, Is a collection of edges.
7. The gravity lens gravitational wave pairing method based on deep learning according to claim 1, wherein in step S5, the maximum weight matching algorithm searches a set of edges in the weighted graph, any two edges in the set of edges do not have a common node, and the sum of the weights of the selected edges is the largest.
8. A deep learning based gravitational lens gravitational wave pairing system for implementing the method of claim 1, said system comprising: the preprocessing module (1) is used for whitening, time domain conversion and segment interception of the original time domain data of the gravitational wave detector; The feature mapping module (2) is deployed with a pretrained twin encoder and is used for converting the preprocessed time domain data into low-dimensional feature vectors; The coarse screening module (3) is used for calculating cosine similarity among the feature vectors and generating a candidate pairing list containing a plurality of most similar events for each event; a fine analysis module (4) for performing generalized cross-correlation-phase transformation calculations on event pairs in the candidate pairing list, outputting a fine similarity score and a time delay estimate; And the global decision module (5) is used for constructing a weighted graph based on the refined similarity score, executing a maximum weight matching algorithm and outputting a final gravitational lens gravitational wave image pairing result.

Description

Gravity lens gravitational wave pairing method and system based on deep learning Technical Field The invention relates to the technical field of gravitational wave data processing, in particular to a gravitational lens gravitational wave pairing method and a gravitational wave pairing system based on deep learning. Background The gravitational lens effect causes signals from the same gravitational wave source to produce multiple images that arrive at the earth at different times and with different magnifications and phase offsets (e.g., morse phase). In gravitational wave astronomy, it is important to identify multiple lens images (i.e. "pairs") belonging to the same source. This not only avoids repeated counting of the same coincidence events, but also allows studying the nature and cosmic parameters of the lens celestial body by measuring the time delay and magnification ratio. Existing pairing methods typically rely on bayesian parameter estimation, i.e. comparing the degree of overlap of posterior parameter distributions (e.g. mass, spin, sky position) of different events. However, this approach is computationally expensive and the degeneracy of the parameters can lead to false positives with high red-shift or low signal-to-noise ratio. In addition, machine learning methods based on spectrograms (Spectrogram) tend to lose phase information, making it difficult to accurately estimate time delays on the order of milliseconds. Thus, there is a need for an automated pairing scheme that can directly utilize time domain waveform information, efficiently process large-scale catalogs, and is insensitive to amplitude differences. Disclosure of Invention The invention aims to overcome the defects of the prior art and provides a gravitational lens gravitational wave pairing method and a gravitational lens gravitational wave pairing system based on deep learning. In order to achieve the above purpose, the present invention adopts the following technical scheme: a gravitational lens gravitational wave pairing method based on deep learning comprises the following steps: Step S1, time series strain data of a gravitational wave detector are obtained, the data are subjected to frequency domain whitening treatment and then are converted back to a time domain, and a whitened strain segment with a fixed length is obtained by interception; S2, inputting the whitened strain segments into a pre-trained twin encoder and mapping the whitened strain segments into low-dimensional feature vectors; Step S3, selecting a plurality of other events with highest similarity for each gravitational wave event based on cosine similarity among feature vectors to form a candidate pairing set; step S4, for each pair of events in the candidate pairing set, calculating waveform cross-correlation by adopting a generalized cross-correlation-phase transformation algorithm to obtain a refined similarity score and a relative time delay estimation between the event pairs; And S5, constructing a weighted graph by taking the gravitational wave event as a node and the refined similarity score as an edge weight, and obtaining a final pairing result by solving the maximum weight matching of the graph. Further, in step S2, the twin encoder includes two one-dimensional convolutional neural network branches with identical structures and shared weights, each branch performs feature extraction on input time domain data through a plurality of convolutional blocks and a pooling layer, and finally outputs a unit norm feature vector of 128 dimensions. Further, in step S2, the pretraining process of the twin encoder adopts a contrast learning mode, uses a normalized temperature scaling cross entropy loss function, uses different images generated by the same gravitational wave source through the gravitational lens effect as positive sample pairs, and uses signals of different gravitational wave sources as negative sample pairs for training. Further, the expression of the loss function is: in the formula, AndRespectively outputting characteristic vectors of the sample i and the sample j through a projection head of the twin encoder; Representing cosine similarity; is a temperature parameter; Is the positive sample pair number; in order to indicate the function, Index for all samples in the batch; Is the first in the batch The feature vectors of the samples are output by a projection head of the twin encoder; by minimizing the loss function, the positive pairs of samples are brought closer together in feature space and the negative pairs are brought farther apart. Further, in step S4, the generalized cross-correlation-phase transformation algorithm is implemented by performing amplitude normalization processing on cross-power spectrums of two signals in a frequency domain, calculating a cross-correlation sequence based on normalized phase information, taking a peak value of the sequence as a refined similarity score, and taking a peak value position as a relative ti