Pulsar searching method and system based on phase characteristics
Through the pulsar search method based on phase characteristics, combined with autocorrelation truncation processing and ResNet deep residual neural network, the problems of large computing resource requirements and difficulty in screening candidate signals in the existing technology are solved, and efficient and accurate pulsar search is achieved, reducing false positive signals and improving the progress of the sky inspection project.
Patent Information
- Application Number
- CN202510447012.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-10
- Publication Date
- 2025-08-15
AI Technical Summary
The existing pulsar search methods have large computing resources requirements, difficult screening of candidate signals, and are prone to missed low signal-to-noise ratio pulsars, resulting in slow progress in data accumulation and sky survey projects.
The phase characteristics-based pulsar search method is adopted, including data preprocessing, channel dispersion matrix and dispersion frequency vector calculation, broadband interference weakening and machine learning prediction candidates. Through autocorrelation truncation processing, constant component deinterference, etc., real pulsar candidates are screened out by combining ResNet deep residual neural network.
Effectively suppress noise and interference, improve signal-to-noise ratio, reduce computing resource requirements, improve search efficiency and accuracy, reduce the difficulty of screening candidate signals, quickly identify real signals, reduce false positives, and improve the progress of the sky inspection project.
Smart Images
Figure CN120492908A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of astronomical observation technology, and in particular to a pulsar search method and system based on phase characteristics. Background Art
[0002] Pulsars are a special type of neutron star that were discovered by humans because they continuously emit electromagnetic pulse signals that sweep across the earth. These pulse signals can be detected in multiple bands. The rotation period of pulsars is short and stable, and they are regarded as accurate "clocks" in the universe. Since the first discovery of pulsars in 1967, astronomers have discovered thousands of pulsars. my country's FAST (500-meter Aperture Spherical Radio Telescope) has achieved remarkable results in pulsar searches. As of November 2024, the number of pulsars discovered has exceeded 1,000.
[0003] Currently, pulsar searches primarily employ two methods: periodic search and single-pulse search. Periodic search relies on the periodic nature of pulsar signals, using Fast Fourier Transform (FFT) technology to convert time-domain signals into frequency-domain signals to identify and extract potential pulse periods. After determining the period, the original time series data is folded to improve the signal-to-noise ratio. Single-pulse search focuses on finding strong, non-periodic pulse signals, making it suitable for capturing isolated burst events. However, both search methods face several challenges:
[0004] Huge computing resource requirements: For example, the 19-beam FAST receiver collects up to 1.6 TB of astronomical observation data per hour, with an annual data volume of 10-20 PB. The mismatch between data generation speed and search speed leads to a serious data accumulation, which impacts the progress of the survey project.
[0005] Generates a large number of candidate signals: most of which are false positives, and only a small number of true signals are real. Currently, candidate signal screening mainly relies on machine learning or manual work;
[0006] Pulsars with low signal-to-noise ratio may be missed: because the pulsar signal may be mixed with radio frequency interference (RFI) when reaching the receiver, the interference signal may mask the true characteristics of the pulsar; although there are some improved methods, such as the FDD method proposed by CGBassa et al. in 2022, which realizes time domain de-dispersion by rotating the phase in the Fourier domain; BoPeng et al. directly search for pulsars in the Fourier domain based on FDD; but the existing pulsar search methods still have shortcomings. Therefore, a pulsar search method and system based on phase characteristics are proposed to solve the above-mentioned problems. Summary of the Invention
[0007] (1) Technical problems solved
[0008] In response to the shortcomings of the existing technology, the present invention provides a pulsar search method and system based on phase characteristics, which has the advantages of improving search efficiency and reducing the difficulty of screening candidate signals. It solves the problems of existing pulsar search methods such as large computing resource requirements, difficulty in screening candidate signals, and easy miss of low signal-to-noise ratio pulsars.
[0009] (2) Technical solution
[0010] To achieve the above-mentioned purpose of improving search efficiency and reducing the difficulty of candidate signal screening, the present invention provides the following technical solution: a pulsar search method based on phase characteristics, comprising the following steps:
[0011] Step 1: Data preprocessing: Perform autocorrelation truncation on the received pulsar time domain signal. First, convert the time domain signal into a frequency domain signal through fast Fourier transform. Then, multiply the frequency domain signal by its own modulus to enhance the signal-to-noise ratio of the periodic signal. Then, perform standardization and truncate the complex vectors whose modulus is greater than the set threshold, retaining the original phase to obtain the preprocessed frequency domain signal.
[0012] Step 2: Calculation of channel dispersion matrix and dispersion frequency vector: For each channel frequency domain signal after preprocessing, calculate the angular frequency ω i , dispersion value kcm - 3pc is taken as an example, for each channel at the angular frequency ω i The frequency domain signal is processed to eliminate dispersion and the angular frequency ω is obtained. i At the dispersion value kcm -3 The frequency domain signal of pc, set the search range of dispersion value according to the requirements, assuming the initial value is 0cm - 3 pc, the maximum value to search is Mcm -3 pc, step size dcm -3 pc, we can get the matrix shown on the left side of the formula, which is called ω i The channel dispersion matrix C D M(ω i ), summing each column, we can get the vector shown on the right side of the formula, which is called ω i The dispersion frequency vector DM_f(ω i ), the formula is:
[0013]
[0014] Step 3: Broadband interference reduction step: Observe the characteristic diagram of the channel dispersion matrix and the dispersion frequency vector. For broadband interference, when no de-dispersion is performed, the phase difference between each channel is small and the energy is strong. The mode of each channel is truncated to the set threshold. It can be considered that F k (ω i ) dm＝0, k∈[1,C] is a constant complex number; subtract a constant complex number from the real frequency domain signal, that is, F k (ω i ) dm＝0 =μ,k∈[1,C], where μ is the average of the complex sum of each channel, which weakens the interference effect; for some broadband interference existing in most channels, the second-order constant component interference removal is adopted, and the receiving channel is divided into two groups, the first half as one group and the second half as another group, and the constant component interference removal method is performed within the group.
[0015] Step 4: Search step: Add the constant component interference removal method in the data preprocessing step to calculate the Array, Search The amplitude spectrum of the pulsar is determined to determine whether the peak shape at its maximum value matches the characteristic peak shape of the pulsar signal; for vector, generating its dispersion channel matrix C D M(ω i ) and its phase spectrum; by screening the appropriate To draw its C D M(ω i ) phase spectrum to determine whether it is a pulsar signal characteristic.
[0016] Step 5: Machine learning prediction of candidate body step: Use training machine learning model instead of manual judgment, use ResNet deep residual neural network, and transform C D M(ω i ) matrix and The amplitude spectrum and phase spectrum of the spliced matrix are used as the input of the network, and the neural network is used to determine whether it meets the characteristics of the pulsar signal; the network is trained using the stochastic gradient descent algorithm and the cross-entropy loss function. The training effect is evaluated based on the change curve of the loss function, the accuracy change curve, the positive rate and the false positive rate during the training process to screen out the true pulsar candidates.
[0017] Preferably, in the data preprocessing step, the set threshold is determined through multiple experiments in combination with the noise level of the received signal and the expected signal enhancement effect, so as to strike a balance between effectively suppressing noise and retaining useful signals.
[0018] Preferably, in the channel dispersion matrix and dispersion frequency vector calculation steps, the search range of the dispersion value is dynamically adjusted according to the common dispersion value distribution of pulsars in historical observation data and the interstellar medium characteristics of the current observation area to improve the search efficiency and accuracy.
[0019] Preferably, in the broadband interference reduction step, for the second-order constant component interference removal method, the division ratio of the two groups is optimized according to the distribution of actual interference signals in different channels to achieve the best suppression effect on different types of broadband interference.
[0020] Preferably, in the machine learning candidate prediction step, before training the ResNet deep residual neural network, the input amplitude spectrum and phase spectrum data are normalized so that data with different features are at the same order of magnitude, so as to accelerate the convergence of the model and improve the training effect.
[0021] Preferably, in the search step, the DM f (ω i ) maximum value matches the characteristic peak shape of the pulsar signal, a template matching algorithm is used to pre-construct characteristic peak shape templates of multiple typical pulsar signals, and the similarity scores between the peak shape to be detected and each template are calculated. When the similarity score exceeds the set matching threshold, the peak shape is determined to be matched, thereby improving the accuracy and objectivity of peak shape matching.
[0022] A pulsar search system based on phase characteristics includes a data preprocessing module, a matrix and vector calculation module, an interference reduction module, a search module and a machine learning prediction module, wherein:
[0023] Data preprocessing module: used to perform autocorrelation truncation processing on the received pulsar time domain signal, convert the time domain signal into a frequency domain signal, enhance the signal-to-noise ratio, perform normalization and truncation processing, retain the original phase, and output the preprocessed frequency domain signal;
[0024] Matrix and vector calculation module: Calculate the channel dispersion matrix C based on the preprocessed frequency domain signal D M(ω i ) and the dispersion frequency vector DM f (ω i ), calculate according to the set dispersion value search range and step size;
[0025] Interference reduction module: By observing the characteristic diagram, it identifies broadband interference and uses constant component interference removal and second-order constant component interference removal methods to reduce the impact of broadband interference on the signal, allowing the pulsar signal characteristics to be revealed;
[0026] Search module: Calculate the DM at each frequency for the data after interference reduction processing f (ω i ) array, search its amplitude spectrum, and screen DM that meets the characteristic peak shape of pulsar signal f (ω i ) vector, generating the corresponding dispersion channel matrix C D M(ωi ) and its phase spectrum to determine whether it is a pulsar signal feature;
[0027] Machine learning prediction module: Using ResNet deep residual neural network, C D M(ω i ) matrix and DM f (ω i ) The amplitude spectrum and phase spectrum of the spliced matrix are taken as input, and the stochastic gradient descent algorithm and the cross entropy loss function are used for training. The real pulsar candidates are screened out according to the training effect evaluation index.
[0028] Preferably, the data preprocessing module is provided with an adaptive threshold adjustment unit, which can dynamically adjust the set threshold during truncation processing according to the noise level and signal strength changes of the pulsar time domain signal received in real time; when the noise level is high, the threshold is increased to enhance the noise suppression capability, and when the signal is weak, the threshold is lowered to retain useful signals as much as possible, thereby further optimizing the signal preprocessing effect.
[0029] Preferably, the matrix and vector calculation module is equipped with a data storage and analysis submodule, which is used to store the dispersion value distribution of pulsars in historical observation data and the interstellar medium characteristic data of the current observation area; when calculating the channel dispersion matrix and dispersion frequency vector, the submodule can dynamically adjust the search range and step size of the dispersion value in real time according to these data to improve the search efficiency and accuracy, and can perform preliminary analysis and screening of the calculation results to reduce the processing burden of subsequent modules.
[0030] (3) Beneficial effects
[0031] Compared with the prior art, the present invention provides a pulsar search method and system based on phase characteristics, which has the following beneficial effects:
[0032] 1. This phase-based pulsar search method and system effectively suppresses the effects of noise and interference, enhances the signal-to-noise ratio of periodic signals, reduces the amount of unnecessary data in the calculation process, and improves the accuracy and efficiency of the search through a series of data preprocessing and interference reduction measures such as autocorrelation truncation and constant component interference reduction. At the same time, the dispersion value search range and step size are dynamically adjusted according to historical observation data and the characteristics of the current observation area, avoiding unnecessary calculations, reducing the demand for computing resources, alleviating the problem of data accumulation, and accelerating the progress of sky survey projects.
[0033] 2. This phase-feature-based pulsar search method and system uses a template matching algorithm to determine peak shape matching, improving the accuracy and objectivity of screening peak shape vectors that meet the characteristics of pulsar signals. It also uses the ResNet deep residual neural network for machine learning to predict candidates. Combined with multiple evaluation indicators during the training process, it can more accurately screen out true pulsar candidates. When searching for real pulsar files, it greatly reduces the number of candidates, efficiently distinguishes real signals from a large number of false positive signals, and reduces the difficulty and workload of candidate signal screening. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 The present invention is Vector and C D Characteristic diagram of the M(ω) matrix;
[0035] Figure 2 The present invention is a constant plural and C D M(ω i ) Feature diagram. DETAILED DESCRIPTION
[0036] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0037] See also Figure 1 and Figure 2 , a pulsar search method based on phase characteristics, comprising the following steps:
[0038] Step 1: Data preprocessing: Perform autocorrelation truncation on the received pulsar time domain signal. First, convert the time domain signal into a frequency domain signal through fast Fourier transform. Then, multiply the frequency domain signal by its own modulus to enhance the signal-to-noise ratio of the periodic signal. Then, perform standardization and truncate the complex vectors whose modulus is greater than the set threshold, retaining the original phase to obtain the preprocessed frequency domain signal.
[0039] Step 2: Calculation of channel dispersion matrix and dispersion frequency vector: For each channel frequency domain signal after preprocessing, calculate the angular frequency ω i Dispersion value kcm - 3pc is taken as an example, for each channel at the angular frequency ω i The frequency domain signal is processed to eliminate dispersion and the angular frequency ω is obtained. i At the dispersion value kcm -3The frequency domain signal of pc, set the search range of dispersion value according to the requirements, assuming the initial value is 0cm - 3 pc, the maximum value to search is Mcm -3 pc, step size dc m -3 pc, we can get the matrix shown on the left side of the formula, which is called ω i The channel dispersion matrix C D M(ω i )Sum each column and we can get the vector shown on the right side of the formula, which is called ω i The dispersion frequency vector DM_f(ω i ), the formula is:
[0040]
[0041] Step 3: Broadband interference reduction step: Observe the characteristic diagram of the channel dispersion matrix and the dispersion frequency vector. For broadband interference, when no de-dispersion is performed, the phase difference between each channel is small and the energy is strong. The mode of each channel is truncated to the set threshold. It can be considered that F k (ω i ) dm＝0 , k∈[1,C] is a constant complex number; subtract a constant complex number from the real frequency domain signal, that is, F k (ω i ) dm＝0 =μ,k∈[1,C], where μ is the average of the complex sum of each channel, which weakens the interference effect; for some broadband interference existing in most channels, the second-order constant component interference removal is adopted, and the receiving channel is divided into two groups, the first half as one group and the second half as another group, and the constant component interference removal method is performed within the group.
[0042] Step 4: Search step: Add the constant component interference removal method in the data preprocessing step to calculate the Array, Search The amplitude spectrum of the pulsar is determined to determine whether the peak shape at its maximum value matches the characteristic peak shape of the pulsar signal; for vector, generating its dispersion channel matrix C D M(ω i ) and its phase spectrum; by screening the appropriate To draw its C D M(ω i ) phase spectrum to determine whether it is a pulsar signal characteristic.
[0043] Step 5: Machine learning prediction of candidate body step: Use training machine learning model instead of manual judgment, use ResNet deep residual neural network, and transform C D M(ω i ) matrix and The amplitude spectrum and phase spectrum of the spliced matrix are used as the input of the network, and the neural network is used to determine whether it meets the characteristics of the pulsar signal; the network is trained using the stochastic gradient descent algorithm and the cross-entropy loss function. The training effect is evaluated based on the change curve of the loss function, the accuracy change curve, the positive rate and the false positive rate during the training process to screen out the true pulsar candidates.
[0044] Specifically, data preprocessing includes:
[0045] Receive pulsar signals and obtain raw time domain data Where C represents the total number of channels;
[0046] Perform fast Fourier transform (FFT) on the original time domain data to obtain the frequency domain signal
[0047] Let the frequency domain signal point multiply its own modulus, that is, F new (ω)=F(ω)·|F(ω)|, so as to enhance the signal-to-noise ratio of the periodic signal.
[0048] The enhanced frequency domain signal is normalized, that is Where std(|F(ω)|) represents the standard deviation of the frequency domain signal modulus; the standardized frequency domain signal is truncated, and a threshold is set. For complex vectors with moduli greater than the threshold, their moduli are truncated to the threshold, and the phase remains unchanged, thereby obtaining the preprocessed frequency domain signal.
[0049] Calculation of channel dispersion matrix C_DM and dispersion frequency vector DM_f
[0050] At angular frequency ω i and dispersion value kcm -3 pc as an example, the pre-processed channels are i The frequency domain signal is de-dispersed. According to the frequency domain de-dispersion principle, the shift of the time domain signal corresponds to the rotation of the complex vector in the frequency domain. For each channel k, the frequency domain signal F k (ω i ) The signal after de-dispersion is where d k is the time required for time shifting of the channel calculated based on the dispersion value k;
[0051] Set the search range of the dispersion value, assuming the initial dispersion value is 0cm -3 pc, the maximum value to be searched is M cm -3 pc, step size is dcm -3 pc, for each dispersion value, calculate the angular frequency ω of each channel iThe frequency domain signal after de-dispersion is obtained by matrix
[0052] This matrix is ω i The channel dispersion matrix C D M(ω i ).
[0053] C D M(ω i ) is summed up, [Fc(ω i ) dm＝0 ,Fc(ω i ) dm＝d ,…,Fc(ω i ) dm＝M ], the vector is ω i The dispersion frequency vector
[0054] Characteristic graph of channel dispersion matrix C_DM and dispersion frequency vector DM_f
[0055] By observing C D M(ω i ) matrix and The characteristic diagram of the array is used to visualize the phase characteristics of the pulsar signal, C D M(ω i ) matrix and The elements of a vector are all complex vectors, so their amplitude spectrum and phase spectrum can be plotted to show their characteristics. Through experiments, plotting The magnitude spectrum of the vector and C D M(ω i ) matrix can best capture the characteristics of pulsar signals. The magnitude spectrum of the vector shows the value of i The energy of each channel under different dispersion values is accumulated. In the time domain signal, the pulse peaks of the pulsar pulse will be aligned after de-dispersion under the correct dispersion value. Similarly, in the frequency domain signal, the phase of the pulsar complex vector will be aligned after de-dispersion under the correct dispersion value. After the accumulation of each channel, a peak will be formed, and the channel dispersion matrix C D M(ω i ) shows the phase distribution of each channel under different dispersion values.
[0056] Figure 1 Shows the most common types of C in the experiment D M(ω i ) matrix and DM f (ω i ) vector feature map, each sub-map Figure 1 Representative DM f (ω i) vector amplitude spectrum, and the sub Figure 2 Representative C D M(ω i ) matrix, Figure 1 (a) depicts the characteristic diagram without interference and pulsar signal characteristics, Figure 1 (b) is the characteristic diagram when there is broadband interference but no pulsar signal characteristics. Figure 1 (c) is the characteristic diagram of the pulsar signal, where the true dispersion value of the pulsar is 78.951 cm - 3 pc, the real frequency is 4Hz. In the figure, we can see that DM f (ω i ) vector reaches its maximum value near the true dispersion value. Similarly, C D M(ω i ) matrix converges at a point close to the true dispersion value. As the phase of each channel progresses with the de-dispersion, the angle of rotation of its complex vector is different. The angle of rotation of the complex vector of the low-frequency channel is larger, and the angle of rotation of the complex vector of the high-frequency channel is smaller. Therefore, C D M(ω i The phase spectrum of the ) matrix exhibits a cone shape at the true dispersion value. Figure 1 (d) is the characteristic diagram of the coexistence of broadband interference and pulsar signal at the same frequency. D M(ω i ) matrix will tend to have a shape similar to that of broadband interference. If the broadband interference is significantly stronger than the pulsar signal, the characteristic diagram will completely tend to be Figure 1 (b) Shape.
[0057] Broadband interference mitigation
[0058] The autocorrelation truncation process has a relatively good effect on weakening narrowband RFI, but it has a general effect on broadband RFI. Further processing is needed to remove the influence of broadband interference. Figure 1 (b) Characteristic diagram of broadband interference, when no dedispersion is performed, i.e., dm = 0 cm -3 At pc, the phase difference of each channel is not big. Since the autocorrelation truncation processing and broadband interference generally have strong energy, the mode of each channel will be truncated to the set threshold. Therefore, it can be considered that F k (ω i ) dm＝0 ,k∈[1,C] is a constant complex number.
[0059] If we artificially create an initial frequency domain signal where each channel is a constant complex number, that is, F k (ω i ) dm＝0,k∈[1,C], perform de-dispersion processing on it to generate the channel dispersion matrix, C_DM and dispersion frequency vector DM_f, as shown Figure 2 is a characteristic graph of a constant complex number, and its characteristics are similar to those of broadband interference. Based on this, a constant complex number is subtracted from the real frequency domain signal, that is, F k (ω i ) dm＝0 =F k (ω i ) dm＝0 -μ, can effectively weaken the influence of interference, allowing the pulsar signal characteristics to reappear; as shown in the following formula, at the angular frequency, ωi, a constant complex number is subtracted from the frequency-domain complex vector of each channel. This constant complex number is the average of the sum of the complex numbers of each channel. This method is called constant component interference removal.
[0060] The formula is:
[0061] Search Process
[0062] In the data preprocessing process, a constant component interference removal method is added, that is, the frequency domain signal after autocorrelation truncation is subjected to constant component interference removal processing; the constant component interference removal method is added to the frequency domain signal after autocorrelation truncation ... Array, Search The amplitude spectrum of the pulsar signal is obtained by determining whether the peak shape at its maximum value matches the characteristic peak shape of the pulsar signal. After the pulsar signal is de-dispersed at the correct dispersion value, The amplitude spectrum of the vector will reach its maximum value when it is close to the true dispersion value, and the peak shape has certain characteristics. D M(ω i ) vector, generating its dispersion channel matrix C D M(ω i ) and its phase spectrum, by observing C D M(ω i ) matrix, the phase of each channel will converge near the true dispersion value and present a cone shape, which can be used to further determine whether it is a pulsar signal sign.
[0063] Machine learning predicts candidates:
[0064] Selection of training samples: In the process of searching for the maximum value, a section of data near the maximum value is selected to determine whether the peak where the maximum value is located is a single peak, thereby screening out the first batch of candidate frequencies and dispersion values; C is calculated based on the screened frequencies and dispersion values. D M(ω i ) matrix, C D M(ω i ) matrix and The amplitude spectrum and phase spectrum of the concatenated matrix are used as the input of the network.
[0065] ResNet is used for training. The input of the network is a feature matrix of (2, 129, 256), where the first dimension 2 is the number of channels, representing the amplitude spectrum and phase spectrum of the feature matrix; (129, 256) is represented by the C matrix of (128, 256). D M(ωi) characteristic matrix and (1,256) Feature matrix composition.
[0066] The network model adopts binary classification, that is, judging true samples or false samples, outputting the probability of true and false, and screening candidate samples according to the threshold. The network adopts the stochastic gradient descent (SGD) algorithm, and the update formula is w:=w-η·ablaJ(w;x (i) ;y (i) ), where is the model parameter, is the learning rate, is the gradient of the loss function with respect to the sample, and the loss function of the network uses the cross entropy loss function, the formula is
[0067] The training effect is evaluated based on the change curves of the loss function, accuracy, true positive rate and false positive rate during the training process. As the number of iterations increases, the training loss and validation loss should show a downward trend, indicating that the model is gradually learning and optimizing its prediction ability. The accuracy can reach up to 97.60%, the true positive rate can reach up to 95.45%, and the false positive rate can reach down to 0.8%. When searching for real pulsar files, only about 20 candidates are finally produced. The real pulsar candidates can be easily screened out by folding and observing their outlines.
[0068] A pulsar search system based on phase characteristics includes a data preprocessing module, a matrix and vector calculation module, an interference reduction module, a search module and a machine learning prediction module, wherein:
[0069] Data preprocessing module: used to perform autocorrelation truncation processing on the received pulsar time domain signal, convert the time domain signal into a frequency domain signal, enhance the signal-to-noise ratio, perform normalization and truncation processing, retain the original phase, and output the preprocessed frequency domain signal;
[0070] The data preprocessing module includes a fast Fourier transform unit, an enhanced signal-to-noise ratio unit, a normalization unit and a truncation unit; the fast Fourier transform unit converts the received original time domain data into a frequency domain signal; the enhanced signal-to-noise ratio unit multiplies the frequency domain signal by its own modulus to enhance the signal-to-noise ratio of the periodic signal; the normalization unit normalizes the enhanced frequency domain signal; the truncation unit truncates the normalized frequency domain signal according to the set threshold, retains the original phase, and outputs the preprocessed frequency domain signal.
[0071] Matrix and vector calculation module: Calculate the channel dispersion matrix C based on the preprocessed frequency domain signal D M(ω i ) and the dispersion frequency vector DM f (ω i ), calculate according to the set dispersion value search range and step size;
[0072] The matrix and vector calculation module includes a dispersion elimination calculation unit and a matrix vector generation unit. The dispersion elimination calculation unit calculates the frequency domain signal of each channel after dispersion elimination according to the input pre-processed frequency domain signal, angular frequency and dispersion value according to the frequency domain dispersion principle; the matrix vector generation unit generates the channel dispersion matrix C according to the result of the dispersion elimination calculation unit and the set dispersion value search range and step size. D M(ω i ), and C D M(ω i ) and sum each column of DM f (ω i ), and obtain the dispersion frequency vector.
[0073] Interference reduction module: By observing the characteristic diagram, it identifies broadband interference and uses constant component interference removal and second-order constant component interference removal methods to reduce the impact of broadband interference on the signal, allowing the pulsar signal characteristics to be revealed;
[0074] The interference reduction module includes a feature graph observation unit, a constant component calculation unit and a second-order constant component calculation unit. The feature graph observation unit is used to observe the channel dispersion matrix C D M(ω i ) and the DM of the dispersion frequency vector f (ω i ) feature map to identify broadband interference; the constant component calculation unit calculates a constant complex number based on the characteristics of the broadband interference and performs constant component interference removal on the frequency domain signal; the second-order constant component calculation unit groups the receiving channels for the broadband interference that exists in most channels, calculates different constant complex numbers, and performs second-order constant component interference removal.
[0075] Search module: Calculate the DM at each frequency for the data after interference reduction processing f (ωi ) array, search its amplitude spectrum, and screen DM that meets the characteristic peak shape of pulsar signal f (ω i ) vector, generating the corresponding dispersion channel matrix C D M(ω i ) and its phase spectrum to determine whether it is a pulsar signal feature;
[0076] The search module includes a feature map observation unit, a constant component calculation unit and a second-order constant component calculation unit; the feature map observation unit is used to observe the channel dispersion matrix C D M(ω i ) and the DM of the dispersion frequency vector f (ω i ) feature map to identify broadband interference; the constant component calculation unit calculates a constant complex number based on the characteristics of the broadband interference and performs constant component interference removal on the frequency domain signal; the second-order constant component calculation unit groups the receiving channels for the broadband interference that exists in most channels, calculates different constant complex numbers, and performs second-order constant component interference removal.
[0077] Machine learning prediction module: Using ResNet deep residual neural network, C D M(ω i ) matrix and DM f (ω i ) The amplitude spectrum and phase spectrum of the spliced matrix are taken as input, and the stochastic gradient descent algorithm and the cross entropy loss function are used for training. The real pulsar candidates are screened out according to the training effect evaluation index.
[0078] In summary, this phase-feature-based pulsar search method and system effectively suppresses the influence of noise and interference, enhances the signal-to-noise ratio of periodic signals, reduces the amount of unnecessary data in the calculation process, and improves the accuracy and efficiency of the search through a series of data preprocessing and interference reduction measures such as autocorrelation truncation processing and constant component interference reduction. At the same time, it dynamically adjusts the dispersion value search range and step size according to historical observation data and the characteristics of the current observation area, avoiding unnecessary calculations, reducing the demand for computing resources, alleviating the data accumulation problem, and accelerating the progress of sky survey projects.
[0079] In addition, a template matching algorithm is used to determine peak shape matching, which improves the accuracy and objectivity of screening peak shape vectors that meet the characteristics of pulsar signals; and the ResNet deep residual neural network is used for machine learning to predict candidates. Combined with multiple evaluation indicators in the training process, it can more accurately screen out true pulsar candidates; when searching for real pulsar files, the number of candidates is greatly reduced, and real signals are efficiently identified from a large number of false positive signals, reducing the difficulty and workload of candidate signal screening.
[0080] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.
[0081] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. A pulsar search method based on phase characteristics, characterized in that: The following steps are involved: Step 1: Data preprocessing: Perform autocorrelation truncation on the received pulsar time domain signal. First, convert the time domain signal into a frequency domain signal through fast Fourier transform. Then, multiply the frequency domain signal by its own modulus to enhance the signal-to-noise ratio of the periodic signal. Then, perform standardization and truncate the complex vectors whose modulus is greater than the set threshold, retaining the original phase to obtain the preprocessed frequency domain signal. Step 2: Calculation of channel dispersion matrix and dispersion frequency vector: For each channel frequency domain signal after preprocessing, calculate the angular frequency ω i , dispersion value kcm - 3pc is taken as an example, for each channel at the angular frequency ω i The frequency domain signal is processed to eliminate dispersion and the angular frequency ω is obtained. i At the dispersion value kcm -3 The frequency domain signal of pc, set the search range of dispersion value according to the requirements, assuming the initial value is 0cm -3 pc, the maximum value to search is Mcm -3 pc, step size dcm -3 pc, we can get the matrix shown on the left side of the formula, which is called ω i The channel dispersion matrix C D M(ω i ), summing each column, we can get the vector shown on the right side of the formula, which is called ω i The dispersion frequency vector DM_f(ω i ), the formula is: Step 3: Broadband interference reduction step: Observe the characteristic diagram of the channel dispersion matrix and the dispersion frequency vector. For broadband interference, when no de-dispersion is performed, the phase difference between each channel is small and the energy is strong. The mode of each channel is truncated to the set threshold. It can be considered that F k (ω i ) dm＝0 , k∈[1,C] is a constant complex number; subtract a constant complex number from the real frequency domain signal, that is, F k (ω i ) dm＝0 =μ,k∈[1,C], where μ is the average of the complex sum of each channel, which weakens the interference effect; for some broadband interference existing in most channels, the second-order constant component interference removal is adopted, and the receiving channel is divided into two groups, the first half as one group and the second half as another group, and the constant component interference removal method is performed within the group. Step 4: Search step: Add the constant component interference removal method to the data preprocessing step and calculate the DM at each frequency f (ω i ) array, search DM f (ω i ) amplitude spectrum, determine whether the peak shape at its maximum value matches the characteristic peak shape of the pulsar signal; for DM f (ω i ) vector, generating its dispersion channel matrix C D M(ω i ) and its phase spectrum; by screening the appropriate DM f (ω i ) to draw its C D M(ω i ) phase spectrum to determine whether it is a pulsar signal characteristic. Step 5: Machine learning prediction of candidate body step: Use training machine learning model instead of manual judgment, use ResNet deep residual neural network, and transform C D M(ω i ) matrix and DM f (ω i ) The amplitude spectrum and phase spectrum of the spliced matrix are used as the input of the network, and the neural network is used to determine whether it meets the characteristics of the pulsar signal; the network is trained using the stochastic gradient descent algorithm and the cross-entropy loss function. The training effect is evaluated according to the change curve of the loss function, the accuracy change curve, the positive rate and the false positive rate during the training process to screen out the real pulsar candidates.
2. The pulsar search method based on phase characteristics according to claim 1, characterized in that: In the data preprocessing step, the threshold is determined through multiple experiments in combination with the noise level of the received signal and the expected signal enhancement effect, so as to strike a balance between effectively suppressing noise and retaining useful signals.
3. The pulsar search method based on phase characteristics according to claim 1, characterized in that: In the channel dispersion matrix and dispersion frequency vector calculation step, the search range of the dispersion value is dynamically adjusted according to the common dispersion value distribution of pulsars in historical observation data and the interstellar medium characteristics of the current observation area to improve the search efficiency and accuracy.
4. The pulsar search method based on phase characteristics according to claim 1, characterized in that: In the broadband interference reduction step, for the second-order constant component interference removal method, the division ratio of the two groups is optimized according to the distribution of the actual interference signal in different channels to achieve the best suppression effect on different types of broadband interference.
5. The pulsar search method based on phase characteristics according to claim 1, characterized in that: In the machine learning candidate prediction step, before training the ResNet deep residual neural network, the input amplitude spectrum and phase spectrum data are normalized so that data with different features are at the same order of magnitude, so as to accelerate the convergence of the model and improve the training effect.
6. The pulsar search method based on phase characteristics according to claim 1, characterized in that: In the search step, the DM f (ω i ) maximum value matches the characteristic peak shape of the pulsar signal, a template matching algorithm is used to pre-construct characteristic peak shape templates of multiple typical pulsar signals, and the similarity scores between the peak shape to be detected and each template are calculated. When the similarity score exceeds the set matching threshold, the peak shape is determined to be matched, thereby improving the accuracy and objectivity of peak shape matching.
7. A system for searching pulsars based on phase characteristics, characterized by: A pulsar search method based on phase characteristics applied to any one of claims 1-6, comprising a data preprocessing module, a matrix and vector calculation module, an interference reduction module, a search module, and a machine learning prediction module, wherein: Data preprocessing module: used to perform autocorrelation truncation processing on the received pulsar time domain signal, convert the time domain signal into a frequency domain signal, enhance the signal-to-noise ratio, perform normalization and truncation processing, retain the original phase, and output the preprocessed frequency domain signal; Matrix and vector calculation module: Calculate the channel dispersion matrix C based on the preprocessed frequency domain signal D M(ω i ) and the dispersion frequency vector DM f (ω i ), calculate according to the set dispersion value search range and step size; Interference reduction module: By observing the characteristic diagram, it identifies broadband interference and uses constant component interference removal and second-order constant component interference removal methods to reduce the impact of broadband interference on the signal, allowing the pulsar signal characteristics to be revealed; Search module: Calculate the DM at each frequency for the data after interference reduction processing f (ω i ) array, search its amplitude spectrum, and screen DM that meets the characteristic peak shape of pulsar signal f (ω i ) vector, generating the corresponding dispersion channel matrix C D M(ω i ) and its phase spectrum to determine whether it is a pulsar signal feature; Machine learning prediction module: Using ResNet deep residual neural network, C D M(ω i ) matrix and DM f (ω i ) The amplitude spectrum and phase spectrum of the spliced matrix are taken as input, and the stochastic gradient descent algorithm and the cross entropy loss function are used for training. The real pulsar candidates are screened out according to the training effect evaluation index.
8. The system for searching pulsars based on phase characteristics according to claim 7, characterized in that: The data preprocessing module is provided with an adaptive threshold adjustment unit, which can dynamically adjust the set threshold during truncation processing according to the noise level and signal strength changes of the pulsar time domain signal received in real time; When the noise level is high, the threshold is increased to enhance the noise suppression capability, and when the signal is weak, the threshold is lowered to retain the useful signal as much as possible, further optimizing the signal preprocessing effect.
9. The system for searching pulsars based on phase characteristics according to claim 7, characterized in that: The matrix and vector calculation module is equipped with a data storage and analysis submodule, which is used to store the dispersion value distribution of pulsars in historical observation data and the interstellar medium characteristic data of the current observation area. When calculating the channel dispersion matrix and dispersion frequency vector, this submodule can dynamically adjust the search range and step size of the dispersion value in real time based on these data to improve search efficiency and accuracy, and can also perform preliminary analysis and screening of the calculation results to reduce the processing burden of subsequent modules.