A constraint-aware working point selection method for pulsar candidate screening of radio telescope

Abstract

The present application relates to a kind of radio telescope pulsar candidate screening constraint perception working point selection method, belong to pulsar candidate screening technical field.The method includes: obtaining the pulsar candidate data to be screened and constructing structured candidate representation;Structured candidate representation is input into candidate score model, the candidate score of each candidate belonging to pulsar candidate is obtained, as the candidate score of the candidate to be screened;Set candidate threshold set and calculate each candidate threshold corresponding recall rate and deployment constraint index;In the candidate threshold satisfying preset deployment constraint, the candidate threshold of maximum recall rate is selected as deployment working point and is applied to the candidate score of the candidate to be screened, and the candidate screening result is obtained, to output pulsar candidate review queue and budget-aware evaluation result.The present application is aimed at solving the technical problems that the prior art is difficult to adapt to actual review budget, output purity requirement and false alarm burden control.

Description

Technical Field

[0001] This invention relates to a constraint-sensing operating point selection method for screening pulsar candidates for radio telescopes, belonging to the field of pulsar candidate screening technology. Background Technology

[0002] Pulsars are a class of rapidly rotating compact celestial bodies, and their radio pulse signals have significant applications in fundamental physics, testing of gravitational theories, research on the interstellar medium, and detection of gravitational waves. With the development of large radio telescopes and sky survey techniques, pulsar search systems can generate a large amount of candidate object data in a short time. Large radio telescopes, such as the Five-hundred-meter Aperture Spherical radio Telescope (FAST), possess high sensitivity and large-scale survey capabilities, increasing the probability of pulsar discovery while also generating a vast sample of candidate objects.

[0003] In practical pulsar search procedures, candidate pulsars are typically obtained through steps such as periodic search, dispersion reduction, folding, and diagnostic map generation. Each candidate pulsar generally possesses a folded diagnostic image and statistical characteristics related to dispersion, period, and signal-to-noise ratio. Due to the presence of significant radio frequency interference, instrument noise, and non-astronomical signals, the proportion of true pulsars in the candidate set is relatively low, requiring a large number of candidates to undergo both machine screening and manual verification. Given the limited capabilities of manual verification, minimizing false alarms while retaining a larger number of true pulsar candidates is a key challenge in large-scale radio survey candidate selection.

[0004] Existing pulsar candidate screening methods typically employ machine learning or deep learning models to score candidates, then use a fixed threshold to convert the model's output score into a binary classification result. For example, a common practice is to classify samples with a score greater than or equal to 0.5 as pulsar candidates and those with a score below 0.5 as non-pulsar candidates. While this type of method is convenient for offline evaluation, it has certain shortcomings in practical deployment: First, the fixed threshold is difficult to adapt to different observation batches, different radio frequency interference environments, and different manual review budgets; second, reporting accuracy, recall, or F1 score only at a fixed threshold cannot fully reflect the impact of candidate ranking results on the workload of manual review; third, in a large-scale candidate stream, even a small false alarm rate can translate into a large manual review burden.

[0005] Therefore, a method is needed that can adaptively select the screening threshold based on manual review budget, output purity requirements, or tolerable false positive rate. This would allow the candidate screening process to no longer rely on a single fixed threshold, but instead determine an appropriate working point based on deployment constraints. Simultaneously, it is also necessary to incorporate budget-aware metrics such as Top-K, Top-x%, and workload-recall ratios to output candidate screening results that better reflect the actual review process. Summary of the Invention

[0006] The purpose of this invention is to provide a constrained sensing operating point selection method for screening pulsar candidates for radio telescopes, aiming to solve the technical problems that existing technologies are difficult to adapt to actual verification budgets, output purity requirements, and false alarm burden control.

[0007] To achieve the above objectives, the technical solution of the present invention is: a constrained sensing operating point selection method for screening radio telescope pulsar candidates, comprising the following steps:

[0008] Step 1: Obtain pulsar candidate data to be screened; wherein, the pulsar candidate data includes candidate diagnostic images and curve statistical information of the candidates;

[0009] Step 2: Construct a structured candidate representation based on the candidate diagnostic image and the curve statistical information; wherein, the structured candidate representation includes two-dimensional diagnostic image features and one-dimensional curve features;

[0010] Step 3: Input the structured candidate representation into the candidate scoring model to obtain the candidate score of each candidate belonging to the pulsar candidate, which is used as the candidate score of the candidate to be screened;

[0011] Step 4: Set a candidate threshold set, and calculate the corresponding recall rate and deployment constraint metrics for each candidate threshold in the candidate threshold set; wherein, the deployment constraint metrics include precision and false positive rate;

[0012] Step 5: Among the candidate thresholds that meet the preset deployment constraints, select the candidate threshold that maximizes the recall rate as the deployment working point.

[0013] Step 6: Apply the deployment work points to the candidate scores of the candidates to be screened to obtain the candidate screening results;

[0014] Step 7: Output the pulsar candidate review queue and budget-aware evaluation results based on the candidate screening results.

[0015] Optionally, Step 2 specifically includes:

[0016] Step 2.1: Extract multiple diagnostic sub-regions from the candidate diagnostic image; wherein, the diagnostic sub-regions include time-phase sub-regions, frequency-phase sub-regions, dispersion curve sub-regions, and periodic curve sub-regions;

[0017] Step 2.2: Normalize the size of each diagnostic sub-region to ensure that the diagnostic sub-regions of different candidates have a uniform input size;

[0018] Step 2.3: Use the multiple diagnostic sub-regions as two-dimensional diagnostic image features to characterize the performance of candidate signals under different diagnostic views;

[0019] Step 2.4: Standardize the curve statistics and reduce the dimensionality to obtain one-dimensional curve features;

[0020] Step 2.5: Combine the two-dimensional diagnostic image features and the one-dimensional curve features as a structured candidate representation.

[0021] Optionally, the candidate scoring model includes an image coding branch, a numerical coding branch, and a scoring output unit, specifically:

[0022] The image coding branch includes a shared image encoder and an image feature aggregation unit. The shared image encoder is used to extract image features from each diagnostic sub-region, and the image feature aggregation unit is used to stitch or fuse the image features from each diagnostic sub-region to obtain an image feature representation.

[0023] The numerical coding branch includes a numerical encoder, wherein the numerical encoder includes a fully connected layer and a nonlinear activation layer, used to map one-dimensional curve features into numerical feature representations;

[0024] The scoring output unit generates candidate output values ​​based on the image feature representation and the numerical feature representation, converts the candidate output values ​​into candidate scores, and performs score calibration on the candidate scores to obtain the candidate scores of the candidates to be screened.

[0025] Optionally, Step 4 specifically includes:

[0026] Step 4.1: Obtain the candidate scores and true labels for each candidate in the calibration dataset;

[0027] Step 4.2: Set a candidate threshold set, wherein the candidate threshold set includes multiple candidate thresholds that are within a preset score range;

[0028] Step 4.3: For each candidate threshold, candidates with scores not lower than the current candidate threshold are identified as pulsar candidates, and candidates with scores lower than the current candidate threshold are identified as non-pulsar candidates;

[0029] Step 4.4: Based on the true labels of the calibration dataset and the judgment results corresponding to the candidate thresholds, calculate the recall, precision, and false positive rate at the current candidate threshold.

[0030] Optionally, Step 5 specifically includes:

[0031] Step 5.1: Set preset deployment constraints according to the manual review requirements, wherein the preset deployment constraints include minimum accuracy constraints and maximum false alarm rate constraints;

[0032] When the preset deployment constraint is a minimum precision constraint, candidate thresholds with precision not lower than the preset precision threshold are selected from the calibration dataset, and the candidate threshold with the highest recall rate is selected as the deployment working point.

[0033] When the preset deployment constraint is the maximum false positive rate constraint, candidate thresholds with false positive rates not higher than the preset false positive rate threshold are selected from the calibration dataset, and the candidate threshold with the highest recall rate is selected as the deployment working point.

[0034] Among them, when the preset deployment constraint is to set both minimum precision constraint and maximum false alarm rate constraint, the candidate threshold with the largest recall rate is selected as the deployment working point from the candidate thresholds that simultaneously satisfy the requirement that the precision is not lower than the preset precision threshold and the false alarm rate is not higher than the preset false alarm rate threshold.

[0035] Step 5.2: When there are no candidate thresholds that meet the preset deployment constraints, a preset backup strategy is used to determine the deployment working point. The preset backup strategy includes selecting the highest candidate threshold, selecting a fixed threshold, or resetting the constraints according to the review budget.

[0036] Optionally, Step 6 specifically includes:

[0037] Step 6.1: Obtain the candidate scores of the candidates to be filtered;

[0038] Step 6.2: Compare the candidate scores of the candidates to be screened with the deployment work points;

[0039] Specifically, when the candidate score is not lower than the deployment working point, the corresponding candidate will be output as a pulsar candidate to be reviewed;

[0040] When the candidate score is lower than the deployment working point, the corresponding candidate will be output as a non-priority review candidate or a radio frequency interference candidate.

[0041] Step 6.3: Sort the candidates from highest to lowest score and generate a candidate review priority queue.

[0042] Optionally, Step 7 specifically includes:

[0043] Step 7.1: Output the binary classification results of the candidate objects based on the deployment working points;

[0044] Step 7.2: Output the candidate review priority queue, which is sorted according to the candidate scores;

[0045] Step 7.3: Calculate the Top-K recall rate, which represents the proportion of real pulsar candidates recalled from the top K candidates after sorting by candidate score;

[0046] Step 7.4: Calculate the Top-x% recall rate, which represents the proportion of true pulsar candidates recalled from the top x% of candidates after sorting by candidate score;

[0047] Step 7.5: Generate a workload-recall curve, which is used to characterize the relationship between the proportion of manual review and the recall rate of pulsar candidates;

[0048] Step 7.6: Estimate the false alarm burden based on the false alarm rate and the number of non-pulsar candidates. The expression for the false alarm burden is as follows:

[0049]

[0050] in, This indicates an estimated number of false alarms. Indicates the false alarm rate. This indicates the number of non-pulsar candidates.

[0051] The beneficial effects of this invention are:

[0052] 1. This invention transforms the pulsar candidate screening process from a fixed threshold classification process into a constraint-aware operating point selection process, which can adaptively determine the screening threshold based on manual review budget, output purity requirements, or false alarm rate constraints.

[0053] 2. This invention selects the deployment work point on the calibration dataset that satisfies the accuracy constraint or false alarm rate constraint and has the highest recall rate, so that the screening results can retain as many real pulsar candidates as possible while controlling the false alarm burden.

[0054] 3. This invention combines structured two-dimensional diagnostic image features and one-dimensional curve features, and can utilize multi-view information and curve statistical information in candidate diagnostic images to improve the stability and interpretability of candidate scoring.

[0055] 4. This invention uses budget-aware indicators such as Top-K, Top-x%, and workload-recall curves to directly link candidate screening results with the proportion of manual review, which is conducive to forming an operable candidate review priority queue.

[0056] 5. This invention can estimate the false alarm burden based on the false alarm rate, making the screening results more likely to correspond to the actual number of candidates and the cost of manual review, and is applicable to large-scale radio survey candidate screening scenarios. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the overall steps of the present invention;

[0058] Figure 2 This is a schematic diagram of multi-sub-region extraction of candidate body diagnostic images according to the present invention;

[0059] Figure 3 This is a schematic diagram of the candidate scoring model structure of the present invention;

[0060] Figure 4 This is a schematic diagram of the constraint-aware threshold selection process on the calibration dataset of the present invention;

[0061] Figure 5 This is a schematic diagram of the workload-recall curve of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0063] Example 1: As Figure 1 As shown, a constraint-sensing operating point selection method for screening radio telescope pulsar candidates includes the following steps:

[0064] Step 1: Obtain pulsar candidate data to be screened; wherein, the pulsar candidate data includes candidate diagnostic images and curve statistical information of the candidates;

[0065] Step 1.1: Obtain the folded diagnostic image of the pulsar candidate. The folded diagnostic image can be a candidate diagnostic map generated by the pulsar candidate analysis program, which is used to characterize the diagnostic information of the candidate signal in terms of time, frequency, dispersion and period search.

[0066] Step 1.2: Obtain curve statistics related to pulsar candidates, including dispersion response curve, period response curve, signal-to-noise ratio statistics, and statistical features calculated from candidate diagnostic results;

[0067] Step 1.3: Obtain candidate category labels that have been manually verified or confirmed based on the known pulsar catalog. Pulsar candidates are marked as positive, and radio frequency interference candidates are marked as negative. Establish the correspondence between candidate identifier, diagnostic image path, curve statistics and candidate category label for each candidate to form a candidate data record.

[0068] Optionally, during the training or calibration phase, the candidate data may also include candidate labels; during the actual deployment phase, the candidates to be screened may not have manual labels.

[0069] Step 2: As Figure 2 As shown, a structured candidate body representation is constructed based on the candidate body diagnostic image and the curve statistical information; wherein, the structured candidate body representation includes two-dimensional diagnostic image features and one-dimensional curve features;

[0070] Step 2.1: Extract multiple diagnostic sub-regions with physical meaning from the candidate diagnostic image; wherein, the diagnostic sub-regions include time-phase sub-regions, frequency-phase sub-regions, dispersion curve sub-regions, and periodic curve sub-regions;

[0071] It is understood that the time-phase sub-region is used to reflect the stability of the candidate signal in the observation time dimension; the frequency-phase sub-region is used to reflect the stability of the candidate signal in the frequency dimension; the dispersion curve sub-region is used to reflect the response characteristics of the candidate signal related to dispersion; and the period curve sub-region is used to reflect the response characteristics of the candidate signal related to period search.

[0072] Step 2.2: Normalize the size of each diagnostic sub-region to ensure that the diagnostic sub-regions of different candidates have a uniform input size;

[0073] Step 2.3: Use the multiple diagnostic sub-regions as two-dimensional diagnostic image features to characterize the performance of candidate signals under different diagnostic views;

[0074] Step 2.4: Standardize the curve statistics and reduce the dimensionality to obtain one-dimensional curve features;

[0075] Optionally, in this embodiment, principal component analysis is used to reduce the dimensionality of the curve statistics to 32-dimensional one-dimensional curve features.

[0076] Step 2.5: Combine the two-dimensional diagnostic image features and the one-dimensional curve features as a structured candidate representation.

[0077] Step 3: As Figure 3 As shown, the structured candidate representation is input into the candidate scoring model to obtain the candidate score of each candidate belonging to the pulsar candidate, which is used as the candidate score of the candidate to be screened.

[0078] Optionally, the candidate scoring model includes an image coding branch, a numerical coding branch, and a scoring output unit, specifically:

[0079] The image coding branch includes a shared image encoder and an image feature aggregation unit. The shared image encoder is used to extract image features from each diagnostic sub-region, and the image feature aggregation unit is used to stitch or fuse the image features from each diagnostic sub-region to obtain an image feature representation.

[0080] Optionally, the image coding branch can adopt an existing convolutional neural network backbone. In this embodiment, the ConvNeXt Base network is used to encode the diagnostic region with shared weights.

[0081] The numerical encoding branch includes a numerical encoder, wherein the numerical encoder includes a fully connected layer and a nonlinear activation layer, used to map one-dimensional curve features into numerical feature representations. Optionally, in this embodiment, the numerical encoder includes at least one fully connected layer and a nonlinear activation layer.

[0082] The scoring output unit generates candidate output values ​​based on the image feature representation and the numerical feature representation, converts the candidate output values ​​into candidate scores, and performs score calibration on the candidate scores to obtain the candidate scores of the candidates to be screened.

[0083] Specifically, in this embodiment, multiple diagnostic sub-regions are input into the image encoding branch to obtain image features corresponding to each diagnostic sub-region. The image features corresponding to each diagnostic sub-region are spliced ​​or fused to obtain the image feature representation of the candidate. The one-dimensional curve features are input into the numerical encoding branch to obtain the numerical feature representation of the candidate. The candidate output value is generated based on the image feature representation and the numerical feature representation. The candidate output value is converted into a candidate score, which is used to represent the probability that the candidate belongs to the pulsar candidate. The candidate score is calibrated. Optionally, in this embodiment, the score calibration includes one or more of Platt scaling, isometric regression calibration, or temperature scaling. The score calibration is fitted on the calibration dataset to adjust the scale of the output score of the candidate scoring model.

[0084] Optionally, in this embodiment, the scoring output unit can generate candidate scores through image single-branch, output averaging, feature concatenation, or sample adaptive gating methods, specifically:

[0085] Image branch output values ​​are generated solely based on image feature representations, and these image branch output values ​​are used as candidate output values. :

[0086]

[0087] in, This represents the image feature representation of the i-th candidate. Represents an image classification header;

[0088] Image branch output values ​​are generated based on the image feature representation and numerical feature representation, respectively. Sum of numerical branch output values The image branch output value and the numerical branch output value are then averaged and fused to obtain the candidate body output value. :

[0089]

[0090]

[0091]

[0092] in, Numerical feature representation, This indicates the numerical classification header.

[0093] The image feature representation and numerical feature representation are concatenated, and candidate output values ​​are generated by fusing the classification head to obtain the candidate output values. :

[0094]

[0095] in, Indicates feature splicing, This indicates a fusion classification header.

[0096] Based on the image feature representation and numerical feature representation, sample adaptive gating weights are generated, and the image branch output value and the numerical branch output value are weighted and fused using the sample adaptive gating weights.

[0097]

[0098]

[0099] in, Indicates a gating network. Indicates the first The gating weights corresponding to each candidate, and 1.

[0100] Furthermore, in this embodiment, the candidate scoring model outputs the candidate output value. and will Convert to candidate body scores, the expression is:

[0101]

[0102] in, Indicates the first The candidate score of each candidate body belongs to the pulsar candidate body. This represents the Sigmoid function.

[0103] Step 4: Set a candidate threshold set, and calculate the corresponding recall rate and deployment constraint metric for each candidate threshold in the candidate threshold set; wherein, the deployment constraint metric includes precision and false positive rate;

[0104] Step 4.1: Obtain the candidate scores and true labels for each candidate in the calibration dataset;

[0105] Step 4.2: Set a candidate threshold set, wherein the candidate threshold set includes multiple candidate thresholds that are within a preset score range;

[0106] Step 4.3: For each candidate threshold, candidates with scores not lower than the current candidate threshold are identified as pulsar candidates, and candidates with scores lower than the current candidate threshold are identified as non-pulsar candidates;

[0107] Step 4.4: Based on the true labels of the calibration dataset and the judgment results corresponding to the candidate thresholds, calculate the recall, precision, and false positive rate at the current candidate threshold.

[0108] Optionally, the real labels are provided by existing labeled datasets, or obtained through manual verification or confirmation by a known pulsar catalog; wherein, pulsar candidates are marked as positive and radio frequency interference candidates are marked as negative; the candidates to be screened do not need to have candidate category labels in advance during the actual deployment phase.

[0109] Furthermore, constraint-aware threshold selection is performed on the calibration dataset. For example... Figure 4 As shown, a candidate threshold set is first set, for example, multiple candidate thresholds are set in a fixed step size within the range of 0 to 1. For each candidate threshold, candidate bodies with a score not lower than that threshold are determined as pulsar candidates.

[0110] Specifically, for candidate thresholds , No. Prediction results of each candidate Determined in the following manner:

[0111] in, For indicator functions; when hour, This indicates that the candidate has been identified as a pulsar candidate for review; when hour, =0 indicates that the candidate is determined to be a non-priority review candidate or a radio frequency interference candidate.

[0112] Furthermore, on the calibration dataset, for candidate thresholds Statistical true positive False positives True negative and false negatives And calculate the following indicators:

[0113]

[0114]

[0115]

[0116] in, Indicates candidate threshold The percentage of real pulsar candidates that were recalled. This indicates the proportion of actual pulsar candidates among the samples selected as pulsar candidates; This indicates the proportion of non-pulsar candidates that were incorrectly selected as pulsar candidates.

[0117] Step 5: Among the candidate thresholds that meet the preset deployment constraints, select the candidate threshold that maximizes the recall rate as the deployment working point.

[0118] Step 5.1: Set preset deployment constraints according to the manual review requirements, wherein the preset deployment constraints include minimum accuracy constraints and maximum false alarm rate constraints;

[0119] Specifically, when the deployment requirement is to ensure the purity of the output candidates, and the preset deployment constraint is set to a minimum precision constraint, candidate thresholds with precision not lower than the preset precision threshold are selected from the calibration dataset, and the candidate threshold with the highest recall rate is selected as the deployment working point. Specifically:

[0120]

[0121] The constraints are:

[0122]

[0123] in, Represents the set of candidate thresholds. This indicates the preset minimum precision threshold. and These represent the candidate thresholds on the calibration dataset. The corresponding recall and precision.

[0124] Specifically, when the deployment requirement is to control the false alarm burden, and the preset deployment constraint is set as the maximum false alarm rate constraint, candidate thresholds with false alarm rates not exceeding the preset false alarm rate threshold are selected from the calibration dataset, and the candidate threshold with the highest recall rate is selected as the deployment working point. Specifically:

[0125]

[0126] The constraints are:

[0127]

[0128] in, This indicates the preset maximum tolerable false alarm rate. Indicates candidate thresholds on the calibration dataset The corresponding false alarm rate.

[0129] Among them, when the preset deployment constraint is to set both minimum precision constraint and maximum false alarm rate constraint, the candidate threshold with the largest recall rate is selected as the deployment working point from the candidate thresholds that simultaneously satisfy the requirement that the precision is not lower than the preset precision threshold and the false alarm rate is not higher than the preset false alarm rate threshold.

[0130] Step 5.2: When there are no candidate thresholds that meet the preset deployment constraints, a preset backup strategy is used to determine the deployment working point. The preset backup strategy includes selecting the highest candidate threshold, selecting a fixed threshold, or resetting the constraints according to the review budget.

[0131] It should be noted that in this embodiment, the calibration dataset and the final test dataset are independent of each other. The deployment working points are determined only based on the calibration dataset and applied to the final test dataset or the actual candidate flow, so as to avoid mixing the threshold selection process with the final evaluation process.

[0132] Step 6: Apply the deployment work points to the candidate scores of the candidates to be screened to obtain the candidate screening results;

[0133] Step 6.1: Obtain the candidate scores of the candidates to be filtered;

[0134] Step 6.2: Compare the candidate scores of the candidates to be screened with the deployment work points;

[0135] Specifically, when the candidate score is not lower than the deployment working point, the corresponding candidate will be output as a pulsar candidate to be reviewed;

[0136] When the candidate score is lower than the deployment working point, the corresponding candidate will be output as a non-priority review candidate or a radio frequency interference candidate.

[0137] Step 6.3: Sort the candidates from highest to lowest score and generate a candidate review priority queue.

[0138] Step 7: Output the pulsar candidate review queue and budget-aware evaluation results based on the candidate screening results.

[0139] It should be understood that this embodiment ultimately obtains budget-aware screening results, which include candidate binary classification results based on deployment work points, candidate review priority queue, Top-K recall, Top-x% recall, workload-recall curve, and false positive burden estimation results, specifically:

[0140] Step 7.1: Output the binary classification results of the candidate objects based on the deployment working points;

[0141] Step 7.2: Output the candidate review priority queue, which is sorted according to the candidate scores;

[0142] Step 7.3: Calculate the Top-K recall rate, which represents the proportion of real pulsar candidates recalled from the top K candidates after sorting by candidate score;

[0143] Step 7.4: Calculate the Top-x% recall rate, which represents the proportion of true pulsar candidates recalled from the top x% of candidates after sorting by candidate score;

[0144] Step 7.5: As Figure 5 As shown, a workload-recall curve is generated, which is used to characterize the relationship between the proportion of manual review and the recall rate of pulsar candidates.

[0145] Step 7.6: Estimate the false alarm burden based on the false alarm rate and the number of non-pulsar candidates. The expression for the false alarm burden is as follows:

[0146]

[0147] in, This indicates an estimated number of false alarms. Indicates the false alarm rate. This indicates the number of non-pulsar candidates.

[0148] Understandably, in radio telescope pulsar candidate selection, the number of candidates is large, the types of radio frequency interference are complex, and the ability for manual verification is limited. Traditional selection methods typically convert model output scores into binary classification results using a fixed threshold, making it difficult to adjust according to actual verification budgets or false alarm tolerance. To address these issues, this embodiment combines the candidate selection threshold selection process with manual verification constraints through Steps 1-7. Given a manual verification budget, output purity requirements, or false alarm rate constraints, it can automatically determine the candidate selection threshold and output a candidate verification priority queue and budget-aware evaluation results. Compared to traditional selection methods, this invention can better adapt to the selection needs of different deployment scenarios, reduce the verification burden caused by false alarms, and maintain the recall of true pulsar candidates.

[0149] In summary, this invention first acquires diagnostic images and curve statistics of pulsar candidates, extracts multiple physically meaningful candidate sub-regions from the diagnostic images, and constructs a structured candidate representation by combining them with a one-dimensional feature vector obtained by dimensionality reduction from curve statistical features. Then, a candidate scoring model is used to output the probability score of each candidate belonging to a pulsar. Candidate thresholds are scanned on a calibration dataset, and under the condition of satisfying preset accuracy constraints or false alarm rate constraints, the threshold with the highest recall rate is selected as the deployment working point. Finally, the deployment working point is applied to the candidates to be screened, and the candidate screening results, review priority queue, and budget-aware evaluation results are output. This invention transforms fixed-threshold classification into a working point selection process oriented towards review constraints. It can adaptively determine the screening threshold based on manual review budget, output purity requirements, or tolerable false alarm burden, maintaining the pulsar candidate recall rate while controlling the number of false alarms, thus improving the interpretability and deployment applicability of large-scale radio survey candidate screening.

[0150] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A constraint-sensing operating point selection method for screening radio telescope pulsar candidates, characterized in that, The method includes the following steps: Step 1: Obtain pulsar candidate data to be screened; wherein, the pulsar candidate data includes candidate diagnostic images and curve statistical information of the candidates; Step 2: Construct a structured candidate representation based on the candidate diagnostic image and the curve statistical information; wherein, the structured candidate representation includes two-dimensional diagnostic image features and one-dimensional curve features; Step 3: Input the structured candidate representation into the candidate scoring model to obtain the candidate score of each candidate belonging to the pulsar candidate, which is used as the candidate score of the candidate to be screened; Step 4: Set a candidate threshold set, and calculate the corresponding recall rate and deployment constraint metric for each candidate threshold in the candidate threshold set; wherein, the deployment constraint metric includes precision and false positive rate; Step 5: Among the candidate thresholds that meet the preset deployment constraints, select the candidate threshold that maximizes the recall rate as the deployment working point. Step 6: Apply the deployment work points to the candidate scores of the candidates to be screened to obtain the candidate screening results; Step 7: Output the pulsar candidate review queue and budget-aware evaluation results based on the candidate screening results.

2. The constraint-sensing operating point selection method for screening radio telescope pulsar candidates according to claim 1, characterized in that, Step 2 specifically refers to: Step 2.1: Extract multiple diagnostic sub-regions from the candidate diagnostic image; wherein, the diagnostic sub-regions include time-phase sub-regions, frequency-phase sub-regions, dispersion curve sub-regions, and periodic curve sub-regions; Step 2.2: Normalize the size of each diagnostic sub-region to ensure that the diagnostic sub-regions of different candidates have a uniform input size; Step 2.3: Use the multiple diagnostic sub-regions as two-dimensional diagnostic image features to characterize the performance of candidate signals under different diagnostic views; Step 2.4: Standardize the curve statistics and reduce the dimensionality to obtain one-dimensional curve features; Step 2.5: Combine the two-dimensional diagnostic image features and the one-dimensional curve features as a structured candidate representation.

3. The constraint-sensing operating point selection method for screening radio telescope pulsar candidates according to claim 1, characterized in that, The candidate scoring model includes an image coding branch, a numerical coding branch, and a scoring output unit, specifically: The image coding branch includes a shared image encoder and an image feature aggregation unit. The shared image encoder is used to extract image features from each diagnostic sub-region, and the image feature aggregation unit is used to stitch or fuse the image features from each diagnostic sub-region to obtain an image feature representation. The numerical coding branch includes a numerical encoder, wherein the numerical encoder includes a fully connected layer and a nonlinear activation layer, used to map one-dimensional curve features into numerical feature representations; The scoring output unit generates candidate output values ​​based on the image feature representation and the numerical feature representation, converts the candidate output values ​​into candidate scores, and performs score calibration on the candidate scores to obtain the candidate scores of the candidates to be screened.

4. The constraint-sensing operating point selection method for screening radio telescope pulsar candidates according to claim 3, characterized in that, Step 4 specifically refers to: Step 4.1: Obtain the candidate scores and true labels for each candidate in the calibration dataset; Step 4.2: Set a candidate threshold set, wherein the candidate threshold set includes multiple candidate thresholds that are within a preset score range; Step 4.3: For each candidate threshold, candidates with scores not lower than the current candidate threshold are identified as pulsar candidates, and candidates with scores lower than the current candidate threshold are identified as non-pulsar candidates; Step 4.4: Based on the true labels of the calibration dataset and the judgment results corresponding to the candidate thresholds, calculate the recall, precision, and false positive rate at the current candidate threshold.

5. The constraint-sensing operating point selection method for screening radio telescope pulsar candidates according to claim 1, characterized in that, Step 5 specifically includes: Step 5.1: Set preset deployment constraints according to the manual review requirements, wherein the preset deployment constraints include minimum accuracy constraints and maximum false alarm rate constraints; When the preset deployment constraint is a minimum precision constraint, candidate thresholds with precision not lower than the preset precision threshold are selected from the calibration dataset, and the candidate threshold with the highest recall rate is selected as the deployment working point. When the preset deployment constraint is the maximum false positive rate constraint, candidate thresholds with false positive rates not higher than the preset false positive rate threshold are selected from the calibration dataset, and the candidate threshold with the highest recall rate is selected as the deployment working point. Among them, when the preset deployment constraint is to set both minimum precision constraint and maximum false alarm rate constraint, the candidate threshold with the highest recall rate is selected as the deployment working point from the candidate thresholds that simultaneously satisfy the requirement that the precision is not lower than the preset precision threshold and the false alarm rate is not higher than the preset false alarm rate threshold. Step 5.2: When there are no candidate thresholds that meet the preset deployment constraints, a preset backup strategy is used to determine the deployment working point. The preset backup strategy includes selecting the highest candidate threshold, selecting a fixed threshold, or resetting the constraints according to the review budget.

6. The constraint-sensing operating point selection method for radio telescope pulsar candidate screening according to claim 5, characterized in that, Step 6 specifically refers to: Step 6.1: Obtain the candidate scores of the candidates to be filtered; Step 6.2: Compare the candidate scores of the candidates to be screened with the deployment work points; Specifically, when the candidate score is not lower than the deployment working point, the corresponding candidate will be output as a pulsar candidate to be reviewed; When the candidate score is lower than the deployment working point, the corresponding candidate will be output as a non-priority review candidate or a radio frequency interference candidate. Step 6.3: Sort the candidates from highest to lowest score and generate a candidate review priority queue.

7. The constraint-sensing operating point selection method for screening radio telescope pulsar candidates according to claim 6, characterized in that, Step 7 specifically refers to: Step 7.1: Output the binary classification results of the candidate objects based on the deployment working points; Step 7.2: Output the candidate review priority queue, which is sorted according to the candidate scores; Step 7.3: Calculate the Top-K recall rate, which represents the proportion of real pulsar candidates recalled from the top K candidates after sorting by candidate score; Step 7.4: Calculate the Top-x% recall rate, which represents the proportion of true pulsar candidates recalled from the top x% of candidates after sorting by candidate score; Step 7.5: Generate a workload-recall curve, which is used to characterize the relationship between the proportion of manual review and the recall rate of pulsar candidates; Step 7.6: Estimate the false alarm burden based on the false alarm rate and the number of non-pulsar candidates. The expression for the false alarm burden is as follows: ； in, This indicates an estimated number of false alarms. Indicates the false alarm rate. This indicates the number of non-pulsar candidates.

Images