Incremental pulsar candidate body identification method and system based on low-rank Bayesian neural network
By using an incremental learning method based on a low-rank Bayesian neural network, the problems of model adaptation to dynamic data distribution changes and high computational complexity were solved, achieving efficient pulsar candidate identification while maintaining identification accuracy and improving the robustness and interpretability of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-20
- Publication Date
- 2026-05-15
AI Technical Summary
Existing deep learning models cannot adapt to the dynamic data distribution changes of different radio telescopes in pulsar candidate identification, and suffer from catastrophic forgetting problems. At the same time, Bayesian neural networks have too high computational complexity and memory consumption, making them difficult to deploy on resource-constrained devices.
A low-rank Bayesian neural network is adopted, and the weight parameters are modeled by the low-rank plus diagonal Gaussian approximation method. Combined with multimodal fusion and attention mechanism, an incremental learning system is constructed. The Bayesian posterior update mechanism is used to fine-tune the model, reduce computational complexity and maintain recognition ability.
It significantly reduces computational and storage overhead, improves the model's incremental learning capability, maintains the recognition accuracy of old data, enhances the ability to capture weak signals, and provides uncertainty assessment to support manual review.
Smart Images
Figure CN122045998A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of astronomical big data processing, deep learning, and pattern recognition technology, and in particular to an incremental pulsar candidate identification method and system based on a low-rank Bayesian neural network. Specifically, it relates to an incremental pulsar candidate identification method and system based on a low-rank Bayesian neural network that can adapt to the differences in data distribution from different observation equipment, has low computational complexity, and is resistant to forgetting. Background Technology
[0002] Pulsars are highly magnetized rotating neutron stars, crucial for gravitational wave detection, interstellar medium research, and deep space navigation. With the commissioning of next-generation high-sensitivity radio telescopes such as the Five-hundred-meter Aperture Spherical Radio Telescope (FAST) and the Square Kilometre Array (SKA), the volume of astronomical observation data has grown exponentially, entering the "exascale" era. This massive amount of data contains significant amounts of radio frequency interference (RFI) and noise, making traditional manual screening methods insufficient.
[0003] In recent years, deep learning methods based on convolutional neural networks (CNNs) have achieved significant results in pulsar candidate identification. However, existing static deep learning models face two major challenges. First, different radio telescopes (such as Parkes and FAST) or the same telescope at different times exhibit significant differences in observation configurations and RFI environments, leading to inconsistencies in data distribution. When traditional models are fine-tuned for new datasets (such as FAST data), they often quickly forget the knowledge learned on older datasets (such as HTRU data). This catastrophic forgetting phenomenon prevents the model from continuously accumulating capabilities during long-term sky surveys. Second, Bayesian neural networks (BNNs) can theoretically address the catastrophic forgetting problem by probabilistically modeling the weights and providing uncertainty estimation, but in standard variational inference, their computational complexity is typically O(d). 2 (d is the parameter dimension) and has a huge memory footprint. This makes BNN difficult to deploy on astronomical data processing terminals that need to handle high-throughput data or on resource-constrained edge computing devices. Therefore, there is an urgent need for a pulsar identification system that can achieve incremental learning to overcome catastrophic forgetting, while also having low computational complexity and low memory footprint to ensure engineering feasibility. Summary of the Invention
[0004] The purpose of this invention is to solve the technical problems in the prior art where the model cannot adapt to changes in dynamic data distribution and the Bayesian network consumes too much computing resources. It proposes an incremental pulsar candidate identification method and system based on a low-rank Bayesian neural network.
[0005] This invention is achieved through the following technical solution: This invention proposes an incremental pulsar candidate identification method based on a low-rank Bayesian neural network, the method comprising: S1: Acquire the observation data stream from the radio telescope, preprocess and fold the data to generate multimodal diagnostic map data of pulsar candidates; S2: Construct a multimodal fusion recognition model based on a Bayesian neural network, the model including a one-dimensional Bayesian convolution branch for processing one-dimensional diagnostic maps and a two-dimensional Bayesian convolution branch for processing two-dimensional diagnostic maps; S3: In the variational inference process of the Bayesian neural network, the low-rank plus diagonal Gaussian approximation (LRDGA) method is used to model the posterior distribution of the network weight parameters, and the high-dimensional weight matrix is decomposed into a combination of low-rank components and diagonal components. S4: Construct a hybrid loss function that includes classification loss, KL divergence loss and sparse regularization term, and train or incrementally update the model based on the observation data of the current batch. S5: Input the multimodal data of the candidate to be identified into the trained model, output the identification result and its uncertainty confidence, and fine-tune the model parameters online using the Bayesian posterior update mechanism when a new data domain arrives.
[0006] Furthermore, in S1, the generated multimodal diagnostic map data includes: pulse profile curve, dispersion curve, sub-band map, and sub-integral map.
[0007] Furthermore, in S2, the one-dimensional Bayesian convolution branch is used to extract features of the pulse profile curve and the dispersion curve; the two-dimensional Bayesian convolution branch is used to extract features of the sub-band map and the sub-integral map, and a channel attention mechanism is embedded in this branch to enhance the attention to the features of the real pulsar signal.
[0008] Furthermore, in S3, the low-rank plus diagonal Gaussian approximation method specifically refers to: The weight matrix w in a Bayesian neural network can be approximated as the mean matrix. Low-rank factor matrices U and V, and diagonal covariance matrix The combination of , the formula is as follows:
[0009] in, Let U be the mean vector of the parameters, and V be low-rank matrices of rank r, and σ be the diagonal variance vector that captures the uncertainty of the parameters; this approximation reduces the computational complexity of variational inference from O(d) to O(n). 2 The complexity is reduced to O(dr), where d is the parameter dimension and r < 0. <d。
[0010] Furthermore, when calculating the KL divergence between the variational distribution q(w|σ) and the prior distribution P(w|D), the dependencies between parameters are calculated using the aforementioned LRDGA decomposition structure. The formula for calculating the KL divergence includes a trace term, a mean offset term, and a logarithmic determinant ratio term, specifically in the following form:
[0011] in, The posterior covariance matrix is... The prior covariance matrix, is the prior mean, and k is the dimension of the variable.
[0012] Furthermore, in S4, the hybrid loss function L total Defined as:
[0013] in, Focal Loss is used to address class imbalance. The KL divergence loss includes L1 regularization, specifically:
[0014] In the formula These are the weighting coefficients. This represents the L1 norm, used to induce structured sparsity in network weights.
[0015] Furthermore, the incremental update in S5 refers to the following: when new data is received from different observation devices or different distributions, the model does not reset the parameters, but uses the posterior distribution learned from the old task as the prior distribution of the new task, and integrates the features of the new data through the Bayesian update mechanism, so as to retain the recognition ability of the old data while learning the new data distribution.
[0016] This invention also proposes an incremental pulsar candidate identification system based on a low-rank Bayesian neural network, the system comprising: Data stream preprocessing module: configured to receive the raw data stream generated by the radio telescope, perform achromatic, periodic search and folding operations, and generate multimodal diagnostic maps in real time; Multimodal feature extraction module: contains parallel one-dimensional convolutional units and two-dimensional convolutional units, used to extract deep features of curve-type data and image-type data respectively; Low-rank Bayesian inference engine: Configured to load a Bayesian neural network model, using low-rank plus diagonal Gaussian approximation (LRDGA) technology to perform forward inference and backward propagation, and output classification probability and uncertainty assessment; Incremental learning controller: Configured to monitor changes in data distribution. When new domain data is detected, it triggers the online update process of the model and uses the new data to fine-tune the Bayesian posterior parameters. The results output and interaction module is used to output the determination results of whether the candidate is a pulsar, and to mark low-confidence samples according to the uncertainty assessment value for manual review.
[0017] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the incremental pulsar candidate identification method based on a low-rank Bayesian neural network.
[0018] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the incremental pulsar candidate identification method based on a low-rank Bayesian neural network.
[0019] The beneficial effects of this invention are: 1. Significantly reduce computation and storage overhead: By introducing LRDGA technology, this invention performs low-rank compression of the parameter space of Bayesian networks. Experiments have shown that it can reduce memory usage by about 72% and improve training efficiency by 60%, enabling complex Bayesian networks to be "engineered" and implemented.
[0020] 2. Excellent incremental learning capability: By utilizing the Bayesian posterior update mechanism, this invention can effectively avoid catastrophic forgetting and maintain the recognition accuracy of old data in continuous learning scenarios across devices (such as from FAST to HTRU) (experimental data shows that it remains above 97%), while traditional methods will experience a significant decline.
[0021] 3. High robustness and interpretability: Combining multimodal fusion and attention mechanisms improves the ability to capture weak signals; at the same time, the uncertainty index of the model output provides astronomers with a reliable basis for manual verification. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0023] Figure 1 This is a flowchart of an incremental pulsar candidate identification method based on a low-rank Bayesian neural network according to the present invention.
[0024] Figure 2This is a schematic diagram of the overall architecture of the multimodal fusion Bayesian neural network in an embodiment of the present invention. Detailed Implementation
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] Specifically, in combination Figures 1-2 To address the catastrophic forgetting problem of existing pulsar identification models under varying data distributions across devices, and the excessive computational cost of standard Bayesian networks, this invention proposes an incremental pulsar candidate identification method based on a low-rank Bayesian neural network. The method includes: S1: Acquire the observation data stream from the radio telescope, preprocess and fold the data to generate multimodal diagnostic map data of pulsar candidates; S2: Construct a multimodal fusion recognition model based on a Bayesian neural network, the model including a one-dimensional Bayesian convolution branch for processing one-dimensional diagnostic maps and a two-dimensional Bayesian convolution branch for processing two-dimensional diagnostic maps; S3: In the variational inference process of the Bayesian neural network, the Low-Rank plus Diagonal Gaussian Approximation (LRDGA) method is used to model the posterior distribution of the network weight parameters, decomposing the high-dimensional weight matrix into a combination of low-rank components and diagonal components. S4: Construct a hybrid loss function that includes classification loss, KL divergence loss and sparse regularization term, and train or incrementally update the model based on the observation data of the current batch. S5: Input the multimodal data of the candidate objects to be identified into the trained model, output the identification result and its uncertainty confidence level, and fine-tune the model parameters online using a Bayesian posterior update mechanism when new data domains arrive. Specifically, the model is trained using the hybrid loss function, and incremental learning is performed using a Bayesian posterior update mechanism when new data domains are received, outputting the identification result of the pulsar candidate and its uncertainty confidence level.
[0027] Furthermore, in S1, the generated multimodal diagnostic map data includes: pulse profile curve, dispersion curve (DM curve), sub-bands, and sub-integrations.
[0028] Furthermore, in S2, the one-dimensional Bayesian convolution branch is used to extract features of the pulse profile curve and the dispersion curve; the two-dimensional Bayesian convolution branch is used to extract features of the sub-band map and the sub-integral map, and a channel attention mechanism is embedded in this branch to enhance the attention to the features of the real pulsar signal, which is used to adaptively adjust the weights of the feature channels.
[0029] Furthermore, in S3, the low-rank plus diagonal Gaussian approximation method specifically refers to: The weight matrix w in a Bayesian neural network can be approximated as the mean matrix. Low-rank factor matrices U and V, and diagonal covariance matrix The combination of , the formula is as follows:
[0030] in, Let U be the mean vector of the parameters, and V be low-rank matrices of rank r, and σ be the diagonal variance vector that captures the uncertainty of the parameters; this approximation reduces the computational complexity of variational inference from O(d) to O(n). 2 The complexity is reduced to O(dr), where d is the parameter dimension and r < 0. <d。
[0031] Furthermore, when calculating the KL divergence between the variational distribution q(w|σ) and the prior distribution P(w|D), the dependencies between parameters are calculated using the aforementioned LRDGA decomposition structure. The formula for calculating the KL divergence includes a trace term, a mean offset term, and a logarithmic determinant ratio term, specifically in the following form:
[0032] in, The posterior covariance matrix is... The prior covariance matrix, is the prior mean, and k is the dimension of the variable.
[0033] Further, in step S4, the hybrid loss function includes Focal Loss to address class imbalance, KL divergence loss to measure distributional differences, and an L1 regularization term to induce structured sparsity. The hybrid loss function L... total Defined as:
[0034] in, Focal Loss is used to address class imbalance. The KL divergence loss includes L1 regularization, specifically:
[0035] In the formula These are the weighting coefficients. This represents the L1 norm, used to induce structured sparsity in network weights.
[0036] Furthermore, the incremental update in S5 refers to the following: when new data is received from different observation devices or different distributions, the model does not reset the parameters, but uses the posterior distribution learned from the old task as the prior distribution of the new task, and integrates the features of the new data through the Bayesian update mechanism, so as to retain the recognition ability of the old data while learning the new data distribution.
[0037] This invention also proposes an incremental pulsar candidate identification system based on a low-rank Bayesian neural network, the system comprising: Data stream preprocessing module: configured to receive the raw data stream generated by the radio telescope, perform achromatic, periodic search and folding operations, and generate multimodal diagnostic maps in real time; Multimodal feature extraction module: It is equipped with a parallel Bayesian convolutional neural network branch, which contains parallel one-dimensional convolutional units and two-dimensional convolutional units, used to extract deep features of curve-type data and image-type data respectively; Low-rank Bayesian inference engine: Configured to load a Bayesian neural network model, using low-rank plus diagonal Gaussian approximation (LRDGA) technology to perform forward inference and backward propagation, and output classification probability and uncertainty assessment; Incremental learning controller: Configured to monitor changes in data distribution. When new domain data is detected, it triggers the online update process of the model and uses the new data to fine-tune the Bayesian posterior parameters. The results output and interaction module is used to output the determination results of whether the candidate is a pulsar, and to mark low-confidence samples according to the uncertainty assessment value for manual review.
[0038] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0039] Example 1 like Figure 1 As shown, this embodiment provides an incremental pulsar candidate identification method based on a low-rank Bayesian neural network, which specifically includes the following steps: S1: Data Acquisition and Multimodal Generation Drift scan data from radio telescopes (such as FAST) are preprocessed using the PRESTO toolkit. The preprocessing includes radio frequency interference (RFI) removal, dedispersion, and period search. For each candidate signal found, a diagnostic map of four modes is generated by folding. (1) Pulse Profile: A one-dimensional vector that shows the integral waveform of the pulse; (2) DM Curve: A one-dimensional vector that shows the change in signal-to-noise ratio with the amount of dispersion; (3) Sub-bands: A two-dimensional matrix that shows the distribution of signal strength with frequency and phase; (4) Sub-integrations: Two-dimensional matrices that show the distribution of signal strength over time and phase.
[0040] S2: Construct a multimodal Bayesian network, such as Figure 2 As shown, the constructed network model contains two parallel branches: One-dimensional branch: The input is Profile and DM Curve, and three one-dimensional Bayesian Conv1D layers are used to extract temporal features.
[0041] Two-dimensional branch: The input consists of sub-bands and sub-integrations, using a 3-layer two-dimensional Bayesian Conv2D layer. To enhance sensitivity to weak pulse signals, a channel attention module (CAM) is embedded after the two-dimensional convolutional layers. The CAM calculates the weight coefficients of each channel through global average pooling and fully connected layers to reconstruct the feature map using weights.
[0042] Fusion layer: The feature vectors output from the two branches above are concatenated and input into the Bayesian fully connected layer for final classification.
[0043] S3: Low-rank variational inference (LRDGA) This is the core step of the invention. Traditional Bayesian networks assume that the weights w follow a Gaussian distribution N(μ,Σ). Directly solving for the covariance matrix Σ results in a huge number of parameters.
[0044] This embodiment employs a low-rank plus diagonal Gaussian approximation strategy. The variational distribution of the weights is defined as q(w|θ), where the parameter θ is no longer the complete covariance matrix, but rather consists of the mean μ, two low-rank matrices U ∈ R^(d×r) and V ∈ R^(d×r), and a diagonal vector σ ∈ R^d.
[0045] The weight sampling process is represented as follows: Where ε1 and ε2 are standard Gaussian noise.
[0046] When calculating the KL divergence (Kullback-Leibler Divergence), the matrix determinant lemma and Woodbury matrix identity are used to calculate it directly based on U, V, and σ, avoiding the storage and inversion operations of a large d×d matrix.
[0047] S4: Hybrid Loss Function Optimization To address the problem of extremely scarce pulsar samples (a significant disparity between positive and negative samples) and to achieve sparsity, a total loss function L is defined. total : L total = L focal + λ·L KL Where L focal = -α t (1-p t ) γ log (p t We set α=0.25 and γ=4, focusing on samples that are difficult to classify. KL It not only includes the KL divergence of the variational and prior distributions, but also incorporates the L1 regularization term λ. reg (|W μ |1+ |U|1+ |V|1) promotes the sparsity of network weights, further compressing the model size.
[0048] S5: Incremental Learning and Inference Initial training: The model is trained using source domain data (such as FAST data) to obtain the posterior distribution parameters θ. source .
[0049] Incremental update: When target domain data (such as HTRU data) arrives, the model is not reinitialized; instead, θ is updated. source As the prior distribution, a new posterior distribution θ is calculated by combining it with the target domain data. target Due to the characteristics of Bayesian inference, the model preserves the distribution information of the source domain.
[0050] Inference output: For the test sample, the model performs multiple Monte Carlo samplings, outputs the mean of the predicted probabilities as the classification result, and outputs the variance as the uncertainty confidence level.
[0051] Example 2 This embodiment provides a system based on the above method, which runs on a high-performance server configured with an NVIDIA GPU.
[0052] The system includes: Data stream interface module: Reads FITS files stored on the back end of the telescope in real time.
[0053] The LRBNN inference engine is implemented using the PyTorch framework. Internally, it encapsulates custom LRLinear and LRConv operators, which implement the low-rank sampling logic in step S3. Compared to standard operators, this engine consumes only 4GB of GPU memory while processing 1.6 million candidates per day.
[0054] Model library management module: Stores the distribution of model parameters at different observation stages. When the system detects a drift in the signal-to-noise ratio characteristics of the input data, it automatically invokes the incremental learning process to fine-tune the current model.
[0055] Experimental verification In the mixed test of FAST and HTRU datasets, the method of this embodiment still maintained an F1-score of 97.05% for the recognition of historical FAST data after completing the training on HTRU data, demonstrating its anti-forgetting ability; at the same time, the training speed was improved by 60% compared with the standard Bayesian network.
[0056] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the incremental pulsar candidate identification method based on a low-rank Bayesian neural network.
[0057] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the incremental pulsar candidate identification method based on a low-rank Bayesian neural network.
[0058] The memory in this application embodiment can be volatile memory or non-volatile memory, or it can include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory used in the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.
[0059] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).
[0060] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.
[0061] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above methods.
[0062] The above provides a detailed description of the incremental pulsar candidate identification method and system based on a low-rank Bayesian neural network proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
Claims
1. An incremental pulsar candidate identification method based on a low-rank Bayesian neural network, characterized in that, The method includes: S1: Acquire the observation data stream from the radio telescope, preprocess and fold the data to generate multimodal diagnostic map data of pulsar candidates; S2: Construct a multimodal fusion recognition model based on a Bayesian neural network, the model including a one-dimensional Bayesian convolution branch for processing one-dimensional diagnostic maps and a two-dimensional Bayesian convolution branch for processing two-dimensional diagnostic maps; S3: In the variational inference process of the Bayesian neural network, the low-rank plus diagonal Gaussian approximation (LRDGA) method is used to model the posterior distribution of the network weight parameters, and the high-dimensional weight matrix is decomposed into a combination of low-rank components and diagonal components. S4: Construct a hybrid loss function that includes classification loss, KL divergence loss and sparse regularization term, and train or incrementally update the model based on the observation data of the current batch. S5: Input the multimodal data of the candidate to be identified into the trained model, output the identification result and its uncertainty confidence, and fine-tune the model parameters online using the Bayesian posterior update mechanism when a new data domain arrives.
2. The method according to claim 1, characterized in that, In step S1, the generated multimodal diagnostic map data includes: pulse profile curve, dispersion curve, sub-band map, and sub-integral map.
3. The method according to claim 2, characterized in that, In S2, the one-dimensional Bayesian convolution branch is used to extract features of the pulse profile curve and the dispersion curve; the two-dimensional Bayesian convolution branch is used to extract features of the sub-band map and the sub-integral map, and a channel attention mechanism is embedded in this branch to enhance the attention to the features of the real pulsar signal.
4. The method according to claim 1, characterized in that, In S3, the low-rank plus diagonal Gaussian approximation method is specifically as follows: The weight matrix w in a Bayesian neural network can be approximated as the mean matrix. Low-rank factor matrices U and V, and diagonal covariance matrix The combination of , the formula is as follows: in, Let U be the mean vector of the parameters, and V be low-rank matrices of rank r, and σ be the diagonal variance vector that captures the uncertainty of the parameters; this approximation reduces the computational complexity of variational inference from O(d) to O(n). 2 The complexity is reduced to O(dr), where d is the parameter dimension and r < 0. <d。 5. The method according to claim 4, characterized in that, When calculating the KL divergence between the variational distribution q(w|σ) and the prior distribution P(w|D), the dependency relationship between the parameters is calculated using the aforementioned LRDGA decomposition structure. The formula for calculating the KL divergence includes a trace term, a mean offset term, and a logarithmic determinant ratio term, specifically in the following form: in, The posterior covariance matrix is... The prior covariance matrix, is the prior mean, and k is the dimension of the variable.
6. The method according to claim 1, characterized in that, In S4, the hybrid loss function L total Defined as: in, Focal Loss is used to address class imbalance. The KL divergence loss includes L1 regularization, specifically: In the formula These are the weighting coefficients. This represents the L1 norm, used to induce structured sparsity in network weights.
7. The method according to claim 1, characterized in that, The incremental update in S5 refers to the following: when new data is received from different observation devices or different distributions, the model does not reset the parameters, but uses the posterior distribution learned from the old task as the prior distribution of the new task, and integrates the features of the new data through the Bayesian update mechanism, so as to retain the recognition ability of the old data while learning the new data distribution.
8. An incremental pulsar candidate identification system based on a low-rank Bayesian neural network, characterized in that, The system includes: Data stream preprocessing module: configured to receive the raw data stream generated by the radio telescope, perform achromatic, periodic search and folding operations, and generate multimodal diagnostic maps in real time; Multimodal feature extraction module: contains parallel one-dimensional convolutional units and two-dimensional convolutional units, used to extract deep features of curve-type data and image-type data respectively; Low-rank Bayesian inference engine: Configured to load a Bayesian neural network model, using low-rank plus diagonal Gaussian approximation (LRDGA) technology to perform forward inference and backward propagation, and output classification probability and uncertainty assessment; Incremental learning controller: Configured to monitor changes in data distribution. When new domain data is detected, it triggers the online update process of the model and uses the new data to fine-tune the Bayesian posterior parameters. The results output and interaction module is used to output the determination results of whether the candidate is a pulsar, and to mark low-confidence samples according to the uncertainty assessment value for manual review.
9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-7.
10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-7.