Monopulse candidate body identification method based on multi-modal feature fusion

Abstract

The invention relates to the technical field of radio astronomical signal processing, and discloses a monopulse candidate body identification method based on multi-modal feature fusion, and the method comprises the steps: firstly carrying out the preprocessing of Fits format data observed by a FAST telescope, and generating two types of multi-modal input of a dispersion-time (DM-Time) graph and a frequency-time (Freq-Time) graph; then, constructing a double-branch feature extraction network, capturing global correlation features of the DM-Time graph through a ConvNeXt network embedded with a CBAM attention mechanism, and extracting local discrimination features of the Freq-Time graph through a ResNet18 network embedded with the CBAM attention mechanism; deep interactive fusion of bimodal features is realized by using a cross attention mechanism, and feature expression is optimized in combination with learnable weight and residual connection; and finally, binary classification is completed through a full connection layer, and a real monopulse candidate recognition result is output. Compared with an existing FETCH method, the method has the advantages that the false positive rate is reduced by 89.54% or above, the F1-score is improved by 14.79% or above, and large-scale searching of transient radio phenomena such as pulsar and rapid radio storm can be efficiently supported.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radio astronomy signal processing, in particular to a single pulse candidate body recognition method based on multi-modal feature fusion. BACKGROUND

[0002] As a high-speed rotating neutron star, pulsars emit periodic pulse signals, which have important scientific value in the fields of stellar evolution, gravitational wave detection and cosmic navigation. Single pulses, as a single radio pulse emitted by a pulsar within one rotation period, contain rich physical information and are the key carriers for studying pulsar radiation mechanisms, discovering new pulsars and identifying fast radio bursts.

[0003] Currently, single pulse candidate body recognition relies on the traditional process of "data collection-interference removal-dispersion-pulse search-artificial diagnosis". With the use of FAST and other devices, the volume of observation data is growing exponentially, and the existing technology faces the following core problems: Single modal feature limitation: traditional methods mostly rely on single modal data (such as only using dispersion-time (DM-Time) or frequency-time (Freq-Time) diagrams), which cannot fully utilize the complementary information of different modal data, leading to signal omission or high false positive rate; Low feature extraction efficiency: some methods use hand-designed statistical features (such as pulse profile mean, standard deviation), which heavily rely on human experience and are difficult to capture nonlinear features in complex interference environments; Coarse fusion strategy: existing multi-modal fusion methods mostly use simple concatenation or weighted summation, without establishing deep interaction between modalities, and cannot exploit feature complementarity; Insufficient robustness: in the face of imbalanced positive and negative sample distribution (real signals are much less than interference signals), the model has poor generalization ability, and a large number of false positive candidates need to be manually screened, which restricts the efficiency of large-scale sky surveys.

[0004] To solve the above problems, it is urgent to develop a single pulse candidate body recognition method that can efficiently extract multi-modal features, deeply fuse modal information and has strong robustness, to improve recognition accuracy and automation level. SUMMARY

[0005] (I) Technical problems solved In view of the deficiencies of the prior art, the present application provides a single pulse candidate body recognition method based on multi-modal feature fusion, which extracts multi-modal features through a double-branch network, enhances key feature perception with an attention mechanism, realizes deep fusion of modalities using cross-attention, and finally improves the accuracy, recall rate and reduces the false positive rate of single pulse candidate body recognition.

[0006] (II) Technical solutions To achieve the above object, the application provides the following technical scheme: A single pulse candidate body recognition method based on multi-modal feature fusion, comprising the following steps: (1) Data preprocessing and multi-modal data generation: convert the Fits data observed by FAST into Filterbank format, screen single pulse candidates and generate DM-Time graph and Freq-Time graph; (2) Double-branch feature extraction: CBAM-ConvNeXt branch is used to extract DM-Time graph features, and CBAM-ResNet18 branch is used to extract Freq-Time graph features, and the CBAM attention module is embedded in the feature stage or residual block of each branch; (3) Cross-attention feature fusion: the double-branch features are fused through a bidirectional cross-attention mechanism, combined with learnable weights and residual connection to obtain the final fusion features; (4) Candidate body classification: the fusion features are input into a fully connected layer, and the classification results are output through a Sigmoid function to determine whether the candidate body is a real single pulse signal.

[0007] Further, in the step (1), the resolution of the DM-Time graph and the Freq-Time graph is 256x256, which is obtained by analyzing the Cand file generated by Your tool through Heimdall.

[0008] Further, in the step (2), the CBAM-ConvNeXt branch contains 4 feature stages, and the number of channels is 96, 192, 384 and 768 in turn, and a CBAM attention module is embedded after the output of each stage; the CBAM-ResNet18 branch contains 4 residual layers, and a CBAM attention module is embedded at the end of each residual block.

[0009] Further, in the step (3), the cross-attention fusion is realized through a query-key-value mechanism, and the double-branch features are used to generate fusion features fc2r and fr2c respectively, then the fusion features are weighted and fused through the learnable weights of Softmax normalization, and residual connection is added.

[0010] Further, in the step (4), the model is trained using weighted binary cross-entropy loss, the weight coefficient is dynamically adjusted according to the positive and negative sample ratio, the optimizer is AdamW, and the initial learning rate is 1e-4.

[0011] (Three) Beneficial technical effects The application extracts DM-Time graph and Freq-Time graph features through a double-branch network respectively, combines CBAM attention to enhance key area perception, avoids signal omission caused by missing of single modal information, adopts a cross-attention mechanism to establish deep interaction between modes, and adaptively learns fusion weights, so that F1-score is improved by 1.3% to 2.5% compared with simple splicing. In the unbalanced test set (positive and negative samples 1:10), the false positive rate is controlled below 0.0251, and the accuracy is 0.9765, which significantly reduces the cost of manual screening; the end-to-end implementation realizes the whole process from data preprocessing to candidate body classification, without the need for manual design of features, and adapts to the demand of large-scale radio astronomy survey. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 It is a single pulse candidate body recognition process flowchart based on multi-modal feature fusion. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical scheme and advantages of the embodiments of the application more clear, the technical scheme in the embodiments of the application is described clearly and completely. Obviously, the described embodiments are part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the application.

[0014] Embodiment 1 1.1 Experimental environment construction Hardware configuration: GPU is NVIDIA A100 (40GB video memory), CPU is Intel Xeon 8375C (32 cores 64 threads), memory is 128GB, storage is 4TB SSD (used for storing Fits format original data and intermediate products); Software environment: operating system Ubuntu 22.04LTS, deep learning framework PyTorch 2.1.0, Python 3.9.18, data processing tools Sigproc v3.0, Heimdall v1.4.2, Your v2.1.1, CUDA 12.1; Model parameter initialization: CBAM-ConvNeXt branch: adopt ImageNet pre-training weight, input channel 3, output feature dimension 768, DropPath probability 0.1, channel number of each stage 96→192→384→768; CBAM-ResNet18 branch: random initialization weight, input channel 3, output feature dimension 512, then mapped to 768 dimensions through linear projection layer, BatchNorm momentum 0.9; Cross-attention fusion module: query / key / value projection dimension 768, number of attention heads 8, dropout probability 0.1; Classification head: fully connected layer input 768 dimensions, output 1 dimension, activation function Sigmoid.

[0015] 1.2 Dataset preparation Original data source: FAST telescope 2023 August-October galactic plane survey observation data, covering the sky area RA 10°-30°, Dec -20°-10°, the original data is in Fits format, the single file size is about 2GB, a total of 120 observation files, the total data amount is 240GB; Data preprocessing process: Use Sigproc tool to remove interference on Fits file (use median filter to remove narrowband RFI, threshold set to 3σ), convert to Filterbank format (sampling frequency 1GHz, frequency range 1.0-2.0GHz, time resolution 1ms); Use Heimdall tool to disperse (DM search range 0-1000 pc / cm 3 , step 1 pc / cm 3 ), filter candidate bodies with signal-to-noise ratio (S / N) ≥6, generate Cand file (containing candidate body DM value, signal-to-noise ratio, arrival time, etc.); Use Your tool to parse Cand file, generate 256x256 resolution DM-Time and Freq-Time maps (pixel value normalized to [0,1]), and label true labels (determine positive and negative samples by manual review + known pulsar catalog cross-validation); Dataset division: Training set: 27000 positive and negative samples each, a total of 54000 samples, divided into training subset and validation subset in the ratio of 8:2; Balanced test set: 4120 positive and negative samples each, a total of 8240 samples (no overlap with training set sky area); Unbalanced test set: 4120 positive samples and 41200 negative samples (expand negative samples by random sampling interference signals), a total of 45320 samples.

[0016] 1.3 Training process control Training strategy: end-to-end training, batch size (BatchSize) 32, training rounds (Epoch) 100, early stopping strategy (Patience=10, using validation set F1-score as monitoring indicator); Optimizer and Loss Function: The optimizer is AdamW, with an initial learning rate of 1e-4 and a weight decay of 1e-5. The learning rate scheduler is CosineAnnealingLR (T_max=50, eta_min=1e-6). The loss function is weighted binary cross-entropy, with the positive and negative sample weight ratio dynamically adjusted (initially 1:1, updated every 10 rounds based on the misclassification rate of positive and negative samples on the validation set, with a maximum weight ratio not exceeding 1:5). Training monitoring: Record the accuracy, recall, FPR, and F1-score of the training / validation set in each round. Save the model weights every 20 rounds. Finally, select the weight with the highest F1-score on the validation set for testing.

[0017] 1.4 Test Procedure and Results Test method: Perform a single forward inference on both the balanced and unbalanced test sets without model fine-tuning, record the predicted probability (Sigmoid output) for each sample, and determine the category with a threshold of 0.5; Repeated verification: To eliminate random errors, the test process was repeated 5 times, and the mean ± standard deviation of each indicator was taken. Final test results:

[0018] 1.5 Stability Analysis of Results In five repeated tests, the coefficients of variation (standard deviation / mean) of each indicator were all less than 0.15%, indicating that the model has extremely strong stability under the optimal parameter configuration. Analysis of prediction error samples: In the balanced test set, the misclassified positive samples were all candidates with low signal-to-noise ratio (S / N=6-8), and the misclassified negative samples were mostly broadband RFI with a "bow tie" structure; in the unbalanced test set, the misclassified samples were mainly concentrated in DM>800pc / cm 3 The high dispersion candidates accounted for approximately 0.3%.

[0019] Example 2 (Robustness test under different observation conditions) 2.1 Variable settings (simulating differences observed in real-world situations) Variable 1: Observed signal-to-noise ratio (S / N): By adding Gaussian white noise, the candidate's signal-to-noise ratio is divided into 3 groups: low signal-to-noise ratio group (S / N=4-6), medium signal-to-noise ratio group (S / N=6-10), and high signal-to-noise ratio group (S / N=10-20). Variable 2: Dispersion (DM): Divided into 3 groups according to DM value: Low DM group (0-200 pc / cm) 3 ), medium DM group (200-600pc / cm) 3 High DM group (600-1000pc / cm) 3 ); Variable 3: RFI intensity: Set RFI intensity as weak (RFI ratio <5%), medium (RFI ratio 5%-15%), strong (RFI ratio 15%-30%) by superimposing real RFI samples (extracted from FAST historical data); Fixed conditions: Keep model parameters, input image resolution (256x256), and the number of test set samples (1000 positive samples and 1000 negative samples per group) consistent.

[0020] 2.2 Group test results

[0021] 2.3 Robustness analysis Under the conditions of medium-high signal-to-noise ratio (S / N≥6), medium-low DM (DM≤600), and weak-medium RFI intensity (RFI ratio≤15%), the F1-score of the model is all ≥0.98, showing stable performance; Even under the extreme combination conditions of strong RFI, high DM, and low signal-to-noise ratio (S / N=4-6, DM=800-1000, RFI ratio 25%-30%), the model accuracy is still above 0.90, and the recall rate is 0.87, which is significantly better than existing methods (FETCH model accuracy is only about 0.75 under similar conditions).

[0022] Example 3: Different input resolution and computational efficiency trade-off test 3.1 Test variables and control conditions Input resolution settings: 3 groups for comparison: 128x128, 256x256 (baseline), and 512x512. Image scaling uses bilinear interpolation, and the pixel value normalization standard is kept consistent; Control conditions: Training parameters are completely consistent with Example 1 (learning rate, batch size, number of training rounds, etc.), and the test set is a balanced test set (8240 samples). Computational efficiency is measured by "single sample inference time" and "training 1 round time consumption".

[0023] 3.2 Performance and efficiency comparison results

[0024] 3.3 Result optimization suggestions 128x128 resolution: least parameters, fastest speed (inference speed is 12.7% faster than the baseline), but higher false positive rate (0.0415), suitable for real-time fast screening scenarios (such as preliminary filtering of patrol data); 256x256 resolution: optimal balance of performance and efficiency, false positive rate as low as 0.0170, inference speed meets large-scale data processing requirements (single day can process 1 million+ candidates), and is the preferred configuration; 512x512 resolution: slightly improved performance (F1-score is only 0.08% higher than the baseline), but significantly increased training and inference time (training time is 2.8 times that of the baseline), suitable only for high-precision verification scenarios (such as known candidate body review).

[0025] Comparative Example 1 Full-dimensional comparison with existing mainstream models 1.1 Comparison model details configuration FETCH model (2020, Devansh et al.), input is 256x256 DM-Time and Freq-Time map, backbone is ResNet50, feature fusion method is simple concatenation, training parameters are consistent with the present invention (learning rate 1e-4, batch size 32, training rounds 100); CNN-Dual model (2022, Liu et al.): based on two-input CNN (double ResNet18 branch), feature fusion is element addition, open source code is adapted and trained on the same dataset; DRAFTS model (2025, Zhang et al.): target detection + binary classification architecture, input is time series after color dispersion, detection network is YOLOv8-nano, classification network is MLP, trained according to the official recommended parameters; The model of the present invention (CCCRC): same as the benchmark configuration of Example 1.

[0026] 1.2 Unified test standard Test set: use the balanced test set (8240 samples) and unbalanced test set (45320 samples) of the present invention to ensure consistent data distribution; Evaluation indicators: accuracy, recall rate, FPR, F1-score, inference speed (samples / second), total training time (hours); Test procedure: all models are independently trained and tested under the same hardware environment, each model is repeated 3 times, and the average value is taken.

[0027] 1.3 Comparison results

[0028] 1.4 Advantage analysis Compared with the DRAFTS comparison model with the best performance: the present invention reduces FPR by 48.17%, increases F1-score by 0.99%, increases inference speed by 65.17%, and reduces training time by 34.01% on the balanced test set. Compared with the widely used FETCH model: the accuracy on the unbalanced test set is improved by 28.03%, the FPR is reduced by 90.12%, and the F1-score is improved by 112.62%, solving the core problem of the performance collapse of the FETCH model in the class imbalance scene.

[0029] It should be noted that in this paper, the term "comprise", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including a" does not exclude the existence of other identical elements in the process, method, article or equipment including the element.

[0030] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

[0031] Those skilled in the art should understand that the above only describes some specific embodiments of the present application, not all embodiments.

Claims

1. A single-pulse candidate body recognition method based on multi-modal feature fusion, characterized in that, The method comprises the following steps: (1) Data preprocessing and multi-modal data generation: converting the Fits data of FAST observation into Filterbank format, screening single-pulse candidates and generating DM-Time and Freq-Time graphs; (2) Double-branch feature extraction: CBAM-ConvNeXt branch is used to extract DM-Time graph features, and CBAM-ResNet18 branch is used to extract Freq-Time graph features, and the CBAM attention module is embedded in the feature stage or residual block of each branch; (3) Cross-attention feature fusion: the double-branch features are fused through a bidirectional cross-attention mechanism, combined with learnable weights and residual connections to obtain the final fused features; (4) Candidate classification: the fused features are input into a fully connected layer, and the classification results are output through a Sigmoid function to determine whether the candidate is a real single-pulse signal.

2. The single-pulse candidate body recognition method based on multi-modal feature fusion according to claim 1, characterized in that, In step (1), the resolution of the DM-Time graph and the Freq-Time graph is 256x256, which is obtained by analyzing the Cand file generated by Your tool and Heimdall. 3.The single-pulse candidate body recognition method based on multi-modal feature fusion according to claim 1, characterized in that, In step (2), the CBAM-ConvNeXt branch contains 4 feature stages, and the number of channels is 96, 192, 384, and 768 in turn, and a CBAM attention module is embedded after each stage output; the CBAM-ResNet18 branch contains 4 residual layers, and a CBAM attention module is embedded at the end of each residual block.

4. The single-pulse candidate body recognition method based on multi-modal feature fusion according to claim 1, characterized in that, In step (3), cross-attention fusion is achieved through a query-key-value mechanism, and the double-branch features are used to generate fused features fc2r and fr2c, respectively, which are then weighted and fused through learnable weights normalized by Softmax, and residual connections are added.

5. The single-pulse candidate body recognition method based on multi-modal feature fusion according to claim 1, characterized in that, In step (4), the model is trained using weighted binary cross-entropy loss, the weight coefficient is dynamically adjusted according to the positive and negative sample ratio, the optimizer is AdamW, and the initial learning rate is 1e-4.

Images