An interpretable pulsar candidate classification method and system

Abstract

The present application relates to the technical field of pulsar candidate classification, and provides an interpretable pulsar candidate classification method and system, wherein a PRESTO diagnosis graph obtained is input into a physical feature bottleneck network and an OCR numerical extraction model; the physical feature bottleneck network comprises a visual encoder, a physical feature prediction head and a multilayer perception machine, and is used for outputting a pulsar candidate classification label and a confidence; the OCR numerical extraction model comprises a normalized coordinate clipping and a large visual language model, and is used for outputting physical values of the pulsar candidate; based on the pulsar candidate classification label and the confidence, a deterministic template engine selects a matching template from a predefined explanation template, and fills the physical values into placeholders of the matching template to generate a natural language explanation text. The present application improves the accuracy of pulsar candidate classification and the reliability of explanation through cross-telescope transfer learning, a physical feature prediction head and a deterministic template engine.

Description

Technical Field

[0001] This invention relates to the field of pulsar candidate classification technology, and in particular to an interpretable pulsar candidate classification method and system. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Pulsar candidate classification is a core screening step in radio astronomy survey data processing, directly determining whether astronomers can efficiently identify real pulsars with scientific research value from massive signals. Existing pulsar candidate classification techniques mainly include fully supervised deep learning classification, partially labeled semi-supervised learning classification, and incremental learning classification.

[0004] In terms of fully supervised deep learning classification, although the Multimodal Fusion Pulsar Recognition Model (MFPIM) involves transfer learning, it is an intra-distribution transfer of data from the same telescope or of the same type. For example, the Five Hundred-meter Aperture Spherical Radio Telescope (FAST) and the High Time Resolution Universe Research Unit (HTRU) both use PRESTO to generate diagnostic maps in the same format, without addressing the cross-domain distribution shift caused by differences in instrument characteristics between different telescopes. Moreover, the Multimodal Fusion Pulsar Recognition Model only outputs classification labels, making the model a complete black box, and astronomers cannot understand the classification basis.

[0005] In the area of ​​semi-supervised learning with few labels, the pulsar candidate identification method based on semi-supervised generative adversarial networks (SGAN) changes the standard discriminator output from binary classification to tri-class classification {Pulsar, Non-Pulsar, Fake} and uses unlabeled samples from the same domain to train the classifier, which improves the recognition accuracy to some extent in scenarios with scarce labels. However, it relies on tens of thousands of unlabeled data from the same telescope to support semi-supervised training and cannot utilize the historical labeling resources of other telescopes. Moreover, SGAN only outputs classification labels and does not provide any mechanism to explain the classification decision. In addition, the discriminator adopts a convolutional neural network (CNN) architecture, which has insufficient global modeling capabilities.

[0006] In incremental learning, incremental pulsar candidate identification based on low-rank Bayesian neural networks employs a Bayesian neural network with one-dimensional and two-dimensional convolutional branches. During variational inference, a low-rank plus diagonal Gaussian approximation technique is introduced, decomposing the weight matrix into the sum of a low-rank matrix and a diagonal matrix to reduce parameter space complexity. However, this incremental pulsar candidate identification focuses on incremental learning within the same telescope's data stream, leading to catastrophic forgetting and neglecting cross-domain transfer learning between different telescopes. In real-world scenarios where target telescope annotations are extremely scarce while source telescopes have rich historical annotations, incremental learning cannot benefit from existing annotations from other telescopes. Furthermore, incremental learning methods only output classification labels, lacking any interpretability mechanism; and the CNN architecture used results in insufficient global modeling capabilities.

[0007] Existing technologies, such as Explainable Table Learning (TabNet) and Transformer for tabular data (FT-Transformer), provide explanations for classification results, but they use ex post-interpretation tools to analyze feature importance. However, ex post-interpretation methods only provide approximate values ​​of feature importance, which have a semantic gap with the actual reasoning process of the model. Moreover, the inputs of TabNet and FT-Transformer are manually extracted tabular features rather than diagnostic images, resulting in the loss of spatial structure information.

[0008] In summary, the existing technology has three structural defects: lack of cross-telescope transfer capability, uninterpretable classification decisions and lack of physical semantic alignment, and the risk of illusion in generative interpretation.

[0009] (1) Lack of cross-telescope transfer capability. Existing methods such as MFPIM's same-dataset transfer, SGAN's same-telescope semi-supervised learning, and same-domain incremental learning all operate within the same telescope or homogeneous data distribution. When the target telescope annotations are extremely scarce while the source telescope has a massive amount of historical annotations, existing methods cannot effectively utilize cross-telescope annotation resources. The differences in data distribution in cross-telescope scenarios stem from the inherent differences in telescope system characteristics (sensitivity, frequency coverage, radio frequency interference environment). Zero-sample direct transfer accuracy is only about 63.5%, and existing technologies have not yet systematically answered the three key questions: "Is cross-telescope transfer feasible? How many target annotations are needed? What is the optimal transfer strategy?"

[0010] (2) Classification decisions are uninterpretable and lack physical semantic alignment. Existing methods only output classification labels, making the model a black box. In scientific discovery scenarios, when astronomers need to decide whether to allocate valuable follow-up observation time, the classification criteria ("why it is classified as a pulsar") are often more important than the classification result itself. Post-hoc interpretability methods (SHAP / LIME) provide approximate values ​​of feature importance and cannot provide traceable explanations at the physical semantic level, such as "the TPP plot has periodic vertical stripes, the FPP plot shows broadband emission, and the DM curve has a significant single peak at non-zero values."

[0011] (3) Generative interpretations carry the risk of illusion and are unsuitable for scientific decision-making scenarios. If large visual language models (VLMs) are used directly to generate classification interpretations, the models may fabricate seemingly reasonable but inconsistent descriptions, such as "the DM curve shows significant dispersion features," which do not match the actual images. In scientific scenarios, an incorrect but seemingly reasonable interpretation is more dangerous than no interpretation at all, as it may lead astronomers to make incorrect follow-up observation decisions based on false information. Existing technologies do not provide an interpretation generation scheme that eliminates illusions from a mechanistic perspective while maintaining the readability of natural language. Summary of the Invention

[0012] To overcome the shortcomings of the prior art, the present invention provides an interpretable pulsar candidate classification method and system, which aims to solve the problems of scarce target telescope annotation data, lack of traceable physical basis for classification results, and the risk of illusion in the generated interpretation in the prior art.

[0013] To achieve the above objectives, the technical solution provided by the present invention is as follows: In a first aspect, the present invention provides an interpretable pulsar candidate classification method, comprising: Obtain the PRESTO diagnostic graph; The PRESTO diagnostic map is input into the physical feature bottleneck network and the OCR numerical extraction model, respectively. The physical feature bottleneck network includes a visual encoder, a physical feature prediction head, and a multilayer perceptron, which are used to output the classification labels and confidence scores of pulsar candidates. The OCR numerical extraction model includes normalized coordinate clipping and a large visual language model, which are used to output the physical values ​​of pulsar candidates. Based on the classification labels and confidence scores of pulsar candidates, the deterministic template engine selects a matching template from predefined interpretation templates and fills the physical values ​​into the placeholders of the matching template to generate natural language interpretation text.

[0014] As a further technical solution, before inputting the PRESTO diagnostic map into the physical feature bottleneck network, the PRESTO diagnostic map is preprocessed to obtain a preprocessed PRESTO diagnostic map. The preprocessing includes scaling and normalization.

[0015] As a further technical solution, the visual encoder is trained using a cross-telescope transfer learning method, specifically: A visual language encoder was trained from scratch using large-scale labeled data from the source telescope. The visual language encoder was fine-tuned using a small amount of labeled data from the target telescope; During fine-tuning, only the parameters of the last K layer encoder are unfrozen, while the parameters of the previous NK layer encoders are frozen.

[0016] As a further technical solution, the visual encoder is a ViT-Base encoder.

[0017] As a further technical solution, the physical feature bottleneck network includes a visual encoder, a physical feature prediction head, and a multilayer perceptron, specifically: The visual encoder extracts global visual features based on the PRESTO diagnostic map; The physics feature prediction head extracts astrophysical semantic features based on global visual features; The multilayer perceptron outputs pulsar candidate classification labels and confidence scores based on the fusion of global visual features and astrophysical semantic features.

[0018] As a further technical solution, the OCR numerical extraction model includes normalized coordinate cropping and a large visual language model, specifically: Normalized coordinate clipping is used to cut out the search information area from the PRESTO diagnostic map; A large-scale visual language model extracts physical values ​​from the search information region based on prompts. These physical values ​​include dispersion, pulse period, signal-to-noise ratio, and reduced chi-square value.

[0019] Secondly, the present invention provides an interpretable pulsar candidate classification system, comprising: Acquisition module: Used to acquire PRESTO diagnostic diagrams; The classification module is used to input the PRESTO diagnostic map into the physical feature bottleneck network and the OCR numerical extraction model, respectively. The physical feature bottleneck network includes a visual encoder, a physical feature prediction head, and a multilayer perceptron, which is used to output the classification labels and confidence scores of pulsar candidates. The OCR numerical extraction model includes normalized coordinate clipping and a large visual language model, which is used to output the physical values ​​of pulsar candidates. Explanation module: Based on the classification labels and confidence scores of pulsar candidates, the deterministic template engine selects a matching template from predefined explanation templates and fills the physical values ​​into the placeholders of the matching template to generate natural language explanation text.

[0020] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the interpretable pulsar candidate classification method described in the first aspect above.

[0021] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the interpretable pulsar candidate classification method described in the first aspect above.

[0022] Fifthly, the present invention provides a computer program product, including software code, wherein the program in the software code performs the steps in the interpretable pulsar candidate classification method described in the first aspect above.

[0023] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the issue of insufficient classification accuracy caused by sparse target telescope samples through cross-telescope transfer learning. Furthermore, it introduces a physical feature prediction head, improving the interpretability of pulsar candidate classification. Compared to existing black-box classifiers that only output labels, this invention's classification results are accompanied by verifiable physical evidence. Astronomers can examine intermediate parameters such as TPP periodicity score, FPP broadband score, and DM peak significance score to independently determine whether the network output for each physical dimension matches the actual image features, ensuring traceability. Additionally, the deterministic template engine fundamentally eliminates the risk of misleading explanatory text generated by the visual language model. Classification serves as guiding information input to the visual language model, improving the consistency between the explanatory text and the classification results. Moreover, the separation of classification and explanation generation balances the accuracy of classification with the readability of the explanation. Attached Figure Description

[0024] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the structure of an interpretable pulsar candidate classification method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the performance variation curves under different target telescope annotations provided in an embodiment of the present invention; Figure 3 A schematic diagram illustrating the cross-telescope transfer training and inference process provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the physical feature bottleneck network provided in an embodiment of the present invention. Detailed Implementation

[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. The terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0029] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0030] Terminology Explanation: Vision Transformer (ViT) is a neural network model that adapts the Transformer architecture from the field of natural language processing to image classification tasks. ViT divides the input image into fixed-size patches, linearly projects each patch into a fixed-dimensional vector, and then uses this vector as a token sequence to input into the standard Transformer encoder. Unlike convolutional neural networks that expand the receptive field layer by layer through local convolutional kernels, ViT establishes global dependencies between all patch images from the first layer using a self-attention mechanism, making it suitable for structured image classification tasks that require integrating information from multiple regions. This invention uses ViT-Base (12 Transformer Blocks, approximately 87M parameters, input size 448×448) as the image encoder and builds a physical feature prediction head on top of it.

[0031] The Physical Feature Bottleneck Network (PINN) is a design paradigm that embeds domain knowledge into a structured neural network architecture. Multiple prediction heads with explicit physical semantics are appended in parallel on top of the global feature vector output by the ViT encoder. Each physical feature prediction head corresponds to a feature dimension with astrophysical meaning (such as TPP periodicity, FPP bandwidth, and DM curve peak significance). All physical feature vectors are concatenated with the ViT encoded features and then input into the classification head, forming a bottleneck structure of "ViT encoding → physical feature prediction → feature fusion classification". The core idea of ​​this design is that classification decisions must pass through a semantically interpretable intermediate layer of physical features, allowing any classification result to be decomposed into the specific contributions of each physical dimension—for example, a model classifying a pulsar as a pulsar implies high TPP periodicity, strong FPP bandwidth, a significant single peak in the DM curve, and high pulse profile SNR, etc. Astronomers can verify whether the judgment of each physical quantity matches the actual image.

[0032] Cross-Telescope Transfer Learning (CTRL) is a cross-domain adaptation method that utilizes a source telescope with a large amount of historical labeled data (such as the Parkes Telescope in Australia, with approximately 84,000 labeled diagnostic images) to pre-train a physical feature bottleneck network, and then transfers the pre-trained weights to a target telescope with extremely scarce labeled data (such as the FAST telescope in China, with only about 2,224 labeled diagnostic images) for few-sample fine-tuning. Unlike standard transfer learning, which pre-trains on general natural images such as ImageNet, the source and target domains of CTRL are both astronomical diagnostic images. The differences in data distribution stem from the system characteristics of different telescopes, such as sensitivity, frequency coverage, and radio frequency interference environments. This invention systematically studies three key factors affecting the efficiency of CTRL: the number of Transformer Blocks unfrozen, the number of pre-training rounds in the source domain, and the model size, and establishes selectively unfreezing the last 12 Transformer Blocks as the optimal transfer strategy.

[0033] Normalized Coordinate Cropping (RCC) utilizes the fixed layout of PRESTO diagnostic maps to locate specific information regions using proportional coordinates (rather than absolute pixel coordinates) relative to the original image size. For a 2200×1700 pixel PRESTO diagnostic map, the search information region is always located at the relative position of the top left corner of the image (left boundary 28.6%, top boundary 6.9%, right boundary 93.7%, bottom boundary 33.2%). Using RCC eliminates pixel shifts caused by scaling or slight format differences between diagnostic maps from different sources, ensuring that the cropped area accurately covers text areas containing physical values ​​such as dispersion (DM), pulse period, signal-to-noise ratio (SNR), and reduced chi-square value (χ²).

[0034] A Deterministic Template Engine (DTE) is a zero-illusion natural language interpretation generation mechanism based on predefined text templates and structured numerical imputation. Multiple predefined interpretation templates are indexed by combinations of (classification category × confidence interval), such as: pulsar + high confidence, pulsar + medium confidence, non-pulsar + high confidence, and non-pulsar + medium confidence. After receiving the classification label and confidence level, the template engine selects the corresponding template and fills the template placeholders with physical values ​​(DM, Period, SNR, χ²) extracted from a large visual language model and feature values ​​output from a physical feature prediction head. Unlike generative methods using large language models, the DTE does not involve any random sampling or generative inference. Every assertion in each interpretation can be traced back to a specific model output value or a value extracted by the large visual language model, fundamentally eliminating the "false but seemingly reasonable" illusionary descriptions that generative methods may produce. This mechanism is suitable for scientific observation data processing scenarios with strict requirements for output determinism and reproducibility.

[0035] Selective unfreezing refers to updating only the parameters of the last K Transformer Blocks of the pre-trained ViT encoder during the fine-tuning phase of transfer learning, while keeping the parameters of the first NK Transformer Blocks frozen. Freezing all parameters (K=0) only allows for fine-tuning the classifier head and is insufficient to bridge the data distribution differences between different telescopes; unfreezing all parameters (K=N) is highly prone to overfitting in scenarios with few samples. Selective unfreezing achieves an optimal balance between domain adaptability and overfitting risk by controlling the number of trainable parameters. This invention experimentally compared three unfreezing strategies (K=0, K=4, and K=12) and established K=12 as the optimal configuration for cross-telescope transfer learning.

[0036] PRESTO Diagnostic Plots are standardized candidate object visualizations generated by the PRESTO (Pulsar Exploration and Search Toolkit) software package. They are typically composed of multiple subplots of 2200×1700 pixels. Each diagnostic plot contains four sub-regions: the upper left is the pulse profile subplot (presenting a one-dimensional folded curve), the upper right is the time-phase subplot (TPP, presenting a two-dimensional grayscale texture), the lower left is the frequency-phase subplot (FPP, presenting a two-dimensional grayscale texture), and the lower right is the DM curve subplot (presenting a one-dimensional signal-to-noise ratio-dispersion curve). The upper left corner also includes a text area containing search information recording physical values. Different radio telescopes (such as Parkes and FAST) use PRESTO to generate diagnostic plots in the same format, providing a data format consistency foundation for cross-telescope knowledge transfer.

[0037] ORC: Optical Character Recognition.

[0038] Example 1 like Figure 1 As shown, this embodiment provides an interpretable pulsar candidate classification method, including: Step 1: Obtain the PRESTO diagnostic chart; Step 2: Input the PRESTO diagnostic map into the physical feature bottleneck network and the OCR numerical extraction model respectively; the physical feature bottleneck network includes a visual encoder, a physical feature prediction head and a multilayer perceptron, which are used to output the classification labels and confidence scores of pulsar candidates; the OCR numerical extraction model includes normalized coordinate clipping and a large visual language model, which are used to output the physical values ​​of pulsar candidates. Step 3: Based on the pulsar candidate classification labels and confidence scores, the deterministic template engine selects a matching template from predefined interpretation templates and fills the physical values ​​into the placeholders of the matching template to generate natural language interpretation text.

[0039] The following content provides a more detailed description of the interpretable pulsar candidate classification method provided in this embodiment.

[0040] In step 1, obtain the PRESTO diagnostic map.

[0041] In this embodiment, the PRESTO diagnostic map is a multi-sub-map stitched together with a size of 2200×1700 pixels. Each diagnostic map contains four sub-regions: the upper left is the pulse profile sub-map (presenting a one-dimensional folded curve), the upper right is the time-phase sub-map (TPP, presenting a two-dimensional grayscale texture), the lower left is the frequency-phase sub-map (FPP, presenting a two-dimensional grayscale texture), the lower right is the DM curve sub-map (presenting a one-dimensional signal-to-noise ratio-dispersion curve), and the upper left corner also contains a search information text area that records physical values.

[0042] In step 2, the PRESTO diagnostic map is input into the physical feature bottleneck network and the OCR numerical extraction model, respectively. The physical feature bottleneck network includes a visual encoder, a physical feature prediction head, and a multilayer perceptron, which is used to output the classification labels and confidence scores of pulsar candidates. The OCR numerical extraction model includes normalized coordinate clipping and a large visual language model, which is used to output the physical values ​​of pulsar candidates.

[0043] In one specific implementation, before inputting the PRESTO diagnostic map into the physical feature bottleneck network, the PRESTO diagnostic map is preprocessed to obtain a preprocessed PRESTO diagnostic map, wherein the preprocessing includes scaling and normalization.

[0044] Furthermore, the PRESTO diagnostic image is preprocessed to obtain a preprocessed PRESTO diagnostic image. Specifically, the PRESTO diagnostic image is first scaled down from 2200×1700 pixels to 448×448 pixels. Then, the 448×448 pixel PRESTO diagnostic image is normalized using ImageNet standards (mean [0.485, 0.456, 0.406], standard deviation [0.229, 0.224, 0.225]) to obtain the preprocessed PRESTO diagnostic image.

[0045] It should be noted that the sub-regions of the PRESTO diagnostic map have a fixed spatial layout (search information and pulse profile sub-maps are in the upper left, TPP in the upper right, FPP in the lower left, and DM curves in the lower right). Data augmentation operations that may destroy spatial semantics, such as random flipping or color dithering, are not used in the preprocessing stage.

[0046] In one specific implementation, the visual encoder is trained using a cross-telescope transfer learning method, specifically: A visual language encoder was trained from scratch using large-scale labeled data from the source telescope. The visual language encoder was fine-tuned using a small amount of labeled data from the target telescope; During fine-tuning, only the parameters of the last K layer encoder are unfrozen, while the parameters of the previous NK layer encoders are frozen.

[0047] In this embodiment, the large-scale labeled data of the source telescope is Psrkes, approximately 70,000 PRESTO diagnostic maps; the small-scale labeled data of the target telescope is FAST, approximately 1,000 PRESTO diagnostic maps. Fine-tuning is performed using 100-1,000 labeled PRESTO diagnostic maps. The value of K determines the number of trainable parameters of the visual language encoder: K=0 indicates all parameters are frozen (only the classification head is fine-tuned, resulting in the fewest trainable parameters), K=12 indicates all 12 Transformer Blocks are unfrozen (maximum number of trainable parameters, but prone to overfitting with few samples). The AdamW optimizer is used in the fine-tuning stage (initial learning rate is...). The model was trained for 5 epochs with a weight decay of 0.01, and the checkpoint with the best F1 score on the validation set was used as the final deployment model. All FAST annotation subsets maintained a pulsar and non-pulsar class balance (1:1).

[0048] In this invention, the training method achieves an optimal balance between "domain adaptability" and "overfitting risk" by controlling K. Comparative experiments reveal three key principles: First, the unfreezing strategy is the primary factor in transfer learning effectiveness, from freezing all Transformers (K=0, accuracy 93.5%) to unfreezing all four Transformers. From Block (K=4, 96.5%) to unfreezing 12 Transformer Blocks (K=12, 97.6%), the accuracy improved by 4.1 percentage points, and the false positive rate decreased from 2.78% to 1.14%, indicating that fine-tuning the classification head alone is insufficient to bridge the differences in cross-telescope data distribution, and the shallow feature extraction layer also needs to participate in domain adaptation. Secondly, the quality of pre-training affects the downstream transfer effect. Extending the training rounds from 50 to 100 rounds, under the same K=12 condition, the accuracy improved from 97.6% to 98.2%, and the number of false negatives decreased significantly from 22 to 11. Thirdly, larger models are not necessarily better. ViT-Large (304M, 24 layers, K=12, unfreezing 12 Transformer Blocks) only achieved an accuracy of 94.3% with 500 FAST annotations, and the false positive rate was as high as 4.41%, indicating that large models are more prone to overfitting in scenarios with few samples for fine-tuning, while medium-sized models have better generalization ability.

[0049] like Figure 2The figure shows the performance curves of the target telescope with 0 to 950 labeled images. Zero-sample direct transfer (0 FAST labels, only Parkes pre-trained weights) achieved an accuracy of only 63.5% (F1=0.426), indicating a significant difference in the data distribution between the two telescopes. After fine-tuning with only 100 FAST labels, the accuracy jumped to 96.1% (F1=0.960), an improvement of 32.6 percentage points compared to zero samples, meaning that 100 labels are sufficient for the classifier to reach a practical level. Further increasing the label count to 500 images resulted in 98.2%, 800 images 98.4%, and 950 images 98.8%, representing the optimal configuration (K=12 + 100 rounds of pre-training + 950 FAST labels). On a 1224 FAST test set, it achieved 98.8% accuracy, an F1 score of 0.988, and a false positive rate of 1.14% (corresponding to 7 false positives and 8 false negatives).

[0050] Specifically, the complete process of cross-telescope transfer training and inference is as follows: Figure 3 As shown, it includes three stages: Phase 1 source telescope pre-training, Phase 2 target telescope selective unfreezing and fine-tuning, and inference deployment.

[0051] In one specific implementation, the visual encoder is a ViT-Base encoder.

[0052] In one specific implementation, such as Figure 4 As shown, the physical feature bottleneck network includes a visual encoder, a physical feature prediction head, and a multilayer perceptron, specifically: The visual encoder extracts global visual features based on the PRESTO diagnostic map; The physics feature prediction head extracts astrophysical semantic features based on global visual features; The multilayer perceptron outputs pulsar candidate classification labels and confidence scores based on the fusion of global visual features and astrophysical semantic features.

[0053] To address the issue that existing technologies directly jump from ViT encoded features to classification labels, lacking explicit physical semantic features, this invention introduces physical feature prediction heads. In this embodiment, there are eight physical feature prediction heads, each corresponding to one astronomical physical feature dimension, as follows: (1) TPP periodicity The physical feature prediction head consists of a linear layer (768→128) → a ReLU activation layer → a linear layer (128→1) → a Sigmoid activation layer, predicting the regularity of vertical stripes in the time-phase graph. A value close to 1 indicates that the periodic vertical stripes (a characteristic of strong pulsars) are clearly visible in the TPP. A value close to 0 indicates random "salt and pepper" noise.

[0054] (2) Number of harmonic components of TPP The physical feature prediction head consists of a linear layer (768→128) → a ReLU activation layer → a linear layer (128→1) → a soft-plus function, which predicts the number of harmonics in a periodic signal and reflects the complexity of the pulse profile.

[0055] (3) FPP broadband The physical feature prediction head consists of a linear layer (768→128) → a ReLU activation layer → a linear layer (128→1) → a Sigmoid activation layer, and predicts whether the signal in the frequency-phase diagram covers the entire observed frequency band. A value close to 1 indicates that the signals in each frequency channel are in the same phase alignment (wideband transmission, pulsar characteristic). A value close to 0 indicates that only a few frequency channels respond (narrowband radio frequency interference characteristic).

[0056] (4) Significance of the peak value of the DM curve The physical feature prediction head consists of a linear layer (768→128) → ReLU activation layer → linear layer (128→1) → Sigmoid activation layer, and predicts whether the peak value at a non-zero DM value on the dispersion curve is significant.

[0057] (5) Classification of DM curve shapes The physical feature prediction head outputs three types of logits from a linear layer (768→3): single-peak, multi-peak, and no-peak. The single-peak structure is a typical dispersive feature of pulsars.

[0058] (6) Pulse profile SNR ≥0: The physical feature prediction head consists of a linear layer (768→128) → ReLU activation layer → linear layer (128→1) → Softplus function, which predicts the intensity of the signal peak relative to the noise baseline in the pulse profile.

[0059] (7) Number of pulse profile peaks ≥0: The physical feature prediction head consists of a linear layer (768→128) → ReLU activation layer → linear layer (128→1) → Softplus function, which predicts the number of identifiable peaks in the pulse profile.

[0060] (8) RFI level ∈[0,1]: The physical feature prediction head consists of a linear layer (768→128) → ReLU activation layer → linear layer (128→1) → Sigmoid activation layer, which predicts the severity of radio frequency interference in the diagnostic diagram. A value close to 1 indicates a high RFI level.

[0061] All physical features are concatenated into a vector p (7 scalars + 3-dimensional DM curve shape = 10 dimensions). The vector P is then concatenated and fused with the global visual features and input into a multilayer perceptron for classification.

[0062] In this embodiment, there are two multilayer perceptrons.

[0063] The design of the aforementioned physical feature prediction head is based on the fact that the input of the multilayer perceptron not only includes global visual features, but also eight astrophysical semantic features with clear physical meanings. During the training phase, the physical feature prediction head can be supplemented by physical feature annotations providing auxiliary supervision signals. MSE loss is used for continuous value physical feature prediction heads, while cross-entropy loss is used for physical feature prediction heads classifying DM curve shapes. The auxiliary loss is weighted... Adding total losses: ; in, Indicates the total loss. Indicates the weighting coefficient. Indicates the first The physical feature predicts the head loss.

[0064] Preferably, =0.1.

[0065] During the inference phase, the physical feature prediction head directly outputs the predicted value. This architecture allows any classification result to be "interpreted," for example, the model might classify it as "pulsar + high confidence." (TPP cyclical) high, (FPP bandwidth) high, (DM peak) high, (Contour SNR) High, (RFI level) Low. Astronomers can open the numerical values ​​of each physical quantity one by one and cross-validate them with the actual visual features of the diagnostic chart. This traceability is a capability that existing black-box classifiers do not possess at all.

[0066] In one specific implementation, the OCR numerical extraction model includes normalized coordinate cropping and a large visual language model, specifically: Normalized coordinate clipping is used to cut out the search information area from the PRESTO diagnostic map; The large-scale visual language model extracts physical values ​​from the search information area based on prompts. These physical values ​​include dispersion (DM), pulse period, signal-to-noise ratio (SNR), and reduced chi-square value (χ²).

[0067] Specifically, since the search information region is always located in the upper left corner of the PRESTO diagnostic image, normalized coordinates (0.286, 0.069, 0.937, 0.332) are used to define the cropping region. On a 2200×1700 PRESTO diagnostic image, this corresponds to pixel coordinates (629, 117, 2061, 564), with an area of ​​approximately 1432×447 pixels (accounting for approximately 17% of the total image area). Normalized coordinates eliminate pixel offsets caused by scaling between diagnostic images from different sources, ensuring that the cropping region always accurately covers the target text. The cropped sub-image is fed into a pre-trained large-scale visual-language model, which extracts four physical values ​​using a fixed-format prompt: "Extract DM, Period (ms), SNR (sigma), and chi2 values ​​from this PRESTO search information panel. Output format: DM=..., Period=..., SNR=..., chi2=...". Experiments show that this method can successfully extract all four values ​​from 92% of the diagnostic charts in Parkes data and 98% of the charts in FAST data.

[0068] The large-scale visual language model can be Qwen3-VL-8B-Instruct.

[0069] In step 3, based on the pulsar candidate classification labels and confidence levels, the deterministic template engine selects a matching template from predefined interpretation templates and fills the physical values ​​into the placeholders of the matching template to generate natural language interpretation text.

[0070] The deterministic template engine predefines multiple explanatory templates, indexed by combining "category × confidence interval". Let the high confidence threshold be... The confidence threshold for the middle is The template selection logic is as follows: ; in, This represents a high-confidence interpretation template for pulsars. This represents the confidence interpretation template in pulsars. This represents a high-confidence interpretation template for non-pulsars. This represents the confidence interpretation template for non-pulsars. This indicates the category label (1 = pulsar, 0 = non-pulsar). This represents the classification confidence score, with a value range of [0,1]. This indicates the high confidence threshold. This represents the confidence threshold.

[0071] Preferably, the high confidence threshold The confidence threshold is 0.8. It is 0.5.

[0072] After selecting a template, fill the template placeholders with the physical values ​​and feature values ​​output by the physical feature prediction head: ; Here, E represents the output after template filling. The Fill operation is a purely deterministic string replacement, replacing placeholders in the template with their corresponding values. This represents a placeholder for the dispersion quantity. This represents a placeholder for the pulse period. This is a placeholder for the signal-to-noise ratio. This represents a placeholder for the reduced chi-square value.

[0073] The four templates cover different categories and confidence levels: (a) The pulse + high confidence template describes periodic vertical stripes in the TPP, broadband emission characteristics in the FPP, and a single-peak structure in the DM curve (peak position and SNR embedded physical values), concluding "high-priority timing follow-up observation is recommended"; (b) The pulse + medium confidence template describes weak but discernible periodic features, concluding "astronomers are advised to visually confirm manually"; (c) The non-pulse + high confidence template describes random noise in the TPP, narrowband RFI dominance in the FPP, and no significant dispersion features in the DM curve, concluding "safe to filter"; (d) The non-pulse + medium confidence template describes a high level of RFI contamination but retains uncertainty, concluding "low-priority review is recommended". Each template is approximately 200 to 300 Chinese characters long.

[0074] Every assertion in each explanation of this invention can be traced back to a deterministic data source, with zero illusions; for example, the assertion "the DM curve has a significant single peak at 138.4 pc / cm³" is derived from the comparison between the OCR extracted value and the preset reference range, "138.4 pc / cm³" comes directly from the original string extracted by the OCR, and the judgment of "significant single peak" comes from the physical feature prediction head. (DM peak significance) and The output value of (DM shape classification) is used; there is no room for the model to play freely in the entire interpretation generation process; and the inference time is less than 1 millisecond (pure string replacement), which is suitable for the high-throughput real-time processing requirements of the FAST production environment.

[0075] Example 2 This embodiment provides an interpretable pulsar candidate classification system, including: Acquisition module: Used to acquire PRESTO diagnostic diagrams; The classification module is used to input the PRESTO diagnostic map into the physical feature bottleneck network and the OCR numerical extraction model, respectively. The physical feature bottleneck network includes a visual encoder, a physical feature prediction head, and a multilayer perceptron, which is used to output the classification labels and confidence scores of pulsar candidates. The OCR numerical extraction model includes normalized coordinate clipping and a large visual language model, which is used to output the physical values ​​of pulsar candidates. Explanation module: Based on the classification labels and confidence scores of pulsar candidates, the deterministic template engine selects a matching template from predefined explanation templates and fills the physical values ​​into the placeholders of the matching template to generate natural language explanation text.

[0076] It should be noted that the acquisition module, classification module, and interpretation module described above correspond to steps 1 to 3 in Embodiment 1. The examples and application scenarios implemented by these modules and their corresponding steps are the same, but they are not limited to the content disclosed in Embodiment 1. It should also be noted that these modules, as part of the system, can be executed in a computer system, such as a set of computer-executable instructions.

[0077] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0078] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0079] Example 3 This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the interpretable pulsar candidate classification method described in Embodiment 1 above.

[0080] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0081] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0082] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0083] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0084] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0085] Example 4 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the interpretable pulsar candidate classification method described in Embodiment 1 above.

[0086] Example 5 This embodiment provides a computer program product, including software code, wherein the program in the software code executes the steps in the interpretable pulsar candidate classification method described in Embodiment 1 above.

[0087] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An interpretable pulsar candidate classification method, characterized in that, include: Obtain the PRESTO diagnostic graph; The PRESTO diagnostic map is input into the physical feature bottleneck network and the OCR numerical extraction model, respectively. The physical feature bottleneck network includes a visual encoder, a physical feature prediction head, and a multilayer perceptron, which are used to output the classification labels and confidence scores of pulsar candidates. The OCR numerical extraction model includes normalized coordinate clipping and a large visual language model, which are used to output the physical values ​​of pulsar candidates. Based on the classification labels and confidence scores of pulsar candidates, the deterministic template engine selects a matching template from predefined interpretation templates and fills the physical values ​​into the placeholders of the matching template to generate natural language interpretation text.

2. The interpretable pulsar candidate classification method as described in claim 1, characterized in that, Before inputting the PRESTO diagnostic map into the physical feature bottleneck network, the PRESTO diagnostic map is preprocessed to obtain a preprocessed PRESTO diagnostic map. The preprocessing includes scaling and normalization.

3. The interpretable pulsar candidate classification method as described in claim 1, characterized in that, The visual encoder is trained using a cross-telescope transfer learning method, specifically: A visual language encoder was trained from scratch using large-scale labeled data from the source telescope. The visual language encoder was fine-tuned using a small amount of labeled data from the target telescope; During fine-tuning, only the parameters of the last K layer encoder are unfrozen, while the parameters of the previous NK layer encoders are frozen.

4. The interpretable pulsar candidate classification method as described in claim 1, characterized in that, The visual encoder is a ViT-Base encoder.

5. The interpretable pulsar candidate classification method as described in claim 1, characterized in that, The physical feature bottleneck network includes a visual encoder, a physical feature prediction head, and a multilayer perceptron, specifically: The visual encoder extracts global visual features based on the PRESTO diagnostic map; The physics feature prediction head extracts astrophysical semantic features based on global visual features; The multilayer perceptron outputs pulsar candidate classification labels and confidence scores based on the fusion of global visual features and astrophysical semantic features.

6. The interpretable pulsar candidate classification method as described in claim 1, characterized in that, The OCR numerical extraction model includes normalized coordinate cropping and a large-scale visual language model, specifically: Normalized coordinate clipping is used to cut out the search information area from the PRESTO diagnostic map; A large-scale visual language model extracts physical values ​​from the search information region based on prompts. These physical values ​​include dispersion, pulse period, signal-to-noise ratio, and reduced chi-square value.

7. An interpretable pulsar candidate classification system, characterized in that, include: Acquisition module: Used to acquire PRESTO diagnostic diagrams; The classification module is used to input the PRESTO diagnostic map into the physical feature bottleneck network and the OCR numerical extraction model, respectively. The physical feature bottleneck network includes a visual encoder, a physical feature prediction head, and a multilayer perceptron, which is used to output the classification labels and confidence scores of pulsar candidates. The OCR numerical extraction model includes normalized coordinate clipping and a large visual language model, which is used to output the physical values ​​of pulsar candidates. Explanation module: Based on the classification labels and confidence scores of pulsar candidates, the deterministic template engine selects a matching template from predefined explanation templates and fills the physical values ​​into the placeholders of the matching template to generate natural language explanation text.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the interpretable pulsar candidate classification method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps in an interpretable pulsar candidate classification method as described in any one of claims 1-6.

10. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps in the interpretable pulsar candidate classification method according to any one of claims 1-6.

Images