RaSPDAM

Introduction

RaSPDAM (Radio Single-Pulse Detection Algorithm Based on Visual Morphological Features) is a novel machine learning algorithm specifically designed for the detection of Fast Radio Bursts (FRBs). Developed to address the computational and time challenges associated with traditional FRB search methods, RaSPDAM offers significant improvements in both efficiency and accuracy.

Background

Fast Radio Bursts (FRBs) are brief, intense pulses of radio energy originating from distant galaxies. Their discovery and study are crucial for understanding the distribution and evolution of matter in the universe. However, the detection of FRBs is a challenging task due to the vast amounts of data produced by radio telescopes and the computational complexity of existing search methods. Traditional techniques often struggle with detecting weak signals and are time-consuming.

Methodology

Pre-processing

Figure 1. Signal Preprocessing & Enhancement

The preprocessing pipeline begins by converting the original signal sequence into standardized images without de-dispersion, where the sequence is divided into smaller slices using a 2-second sliding window with a 1-second overlap. These image slices are then resized to 512 × 512 pixels to standardize the model input. To enhance signal characteristics, particularly the typical curve of the FRB signal amidst noise, convolution is applied using multiple kernels generated from fixed curve slopes, transforming the image by processing each pixel and its local neighbors. After this, morphological dilation further enhances the signal. Finally, the enhanced images are combined with the original image to form an RGB image, which serves as the model input, as illustrated in Figure 1.

Pulse searching

Figure 2. Pulse Search Procedure

RaSPDAM utilizes a U-Net architecture, a convolutional neural network (CNN) designed for image segmentation tasks. This U-Net model comprises an encoder that extracts slope-based features from the input image and a decoder that reconstructs the image using these features. Despite this process, some noise may still remain in the output, so we apply additional filtering steps to identify potential FRB signals. Using the "regionprops" function, we analyze connected regions in the segmented image and calculate each candidate's projections on the x- and y-axes. Candidates exceeding defined thresholds on both axes are flagged as potential FRB signals.

We have upgraded the model to use nnUNet, aligning it with NMS (Non-Maximum Suppression) and overlapping suppression techniques, in order to enhance performance.

For more details, please refer to the paper.

Dataset

The testing of RaSPDAM is facilitated by the FAST-FREX dataset, which is built upon observations obtained by the Five-hundred-meter Aperture Spherical radio Telescope (FAST). The dataset consists of:

• 600 positive samples: Observed FRB signals from three sources (FRB20121102, FRB20180301, and FRB20201124).
• 1000 negative samples: Noise and Radio Frequency Interference (RFI).

Key Features

Efficiency and Accuracy

• High Precision: RaSPDAM achieves a precision of 97.28%, significantly outperforming traditional methods like PRESTO and Heimdall.
• High Recall: With a recall rate of 83.50%, RaSPDAM effectively identifies a large proportion of true FRB signals.
• F1 Score: An F1 score of 0.9253 indicates a well-balanced trade-off between precision and recall.

Versatility

• ToA and DM: While RaSPDAM currently provides Time of Arrival (ToA) as a result, future enhancements aim to include Dispersion Measure (DM) for more comprehensive signal verification.

Performance Benchmarks

Comparison with Traditional Methods

Software	TN	TP	FN	FP	Recall	Precision	F1 Score
PRESTO	3	472	0	26963700	0.7867	1.7505E-05	3.5009E-05
Heimdall	218	489	36	5854	0.8150	0.0771	0.1409
RaSPDAMv1(UNet)	989	466	128	6	0.7767	0.9873	0.8694
RaSPDAMv2(nnUNet)	994	501	67	14	0.8350	0.9728	0.9253

Discoveries

Since its deployment, RaSPDAM has been instrumental in identifying:

• 2 new FRBs: FRB20211103A and FRB20230104.
• 80 pulsars: Including 13 previously undiscovered pulsars, highlighting the algorithm's efficacy in uncovering new celestial objects.

---

Usage:

Training

Generate Training Data

The following command can be used to generate training data in the running directory. Replace {RaSPDAM_PATH} with the actual path to the RaSPDAM directory. You can change the training data generation parameters in {RaSPDAM_PATH}/utils/traning_data_gen.py.

python {RaSPDAM_PATH}/utils/traning_data_gen.py

Train the Model

Follow the instructions of nnUNet to train the model, refer to nnUNet.

• Create the nnUNet_raw/{DatasetXXX_Name} directory, place the generated training data and {RaSPDAM_PATH}/models/dataset.json in it, change the numTraining in json file.
• Run nnUNetv2_plan_and_preprocess to generate the necessary preprocessing files.
• Run nnUNetv2_train to train the model. For FOLD in [0, 1, 2, 3, 4], run:

nnUNetv2_train {DATASET_NAME_OR_ID} 2d {FOLD} --npz

• Copy the corresponding trained model nnUNet_results/{DatasetXXX_Name}/nnUNetTrainer__nnUNetPlans__2d/ to the {RaSPDAM_PATH}/models/ directory.

Install Requirements

pip install -r requirements.txt

Run the Script

python raspdam.py <fits_file> [-m <model_path>] [-o <output_path>] [-iou_theshold <value>] [-overlap_threshold <value>] [-box_fill_threshold <value>] [-box_projection_threshold <value>]

Arguments：

• fits_file: The path to the FITS file to be processed (required).
• -m, -model_path: The path to the model file (optional, default: models).
• -o, -output_path: The path to save the output file (optional, default: current directory).
• -iou_threshold: Minimum IoU threshold for bounding box filtering (optional, default: 0.5).
• -overlap_threshold: Minimum overlap threshold for bounding box filtering (optional, default: 0.5).
• -box_fill_threshold: Minimum percentage threshold for the area filled within each bounding box (optional, default: 0.25).
• -box_projection_threshold: Minimum projection percentage threshold for bounding box filtering (optional, default: 0.15).