Multi-frequency radio power source modeling method based on hybrid expert model and multi-Gaussian fitting

Abstract

The invention relates to the technical field of radio astronomy data cube processing, and aims to solve the problems that the traditional two-dimensional Gaussian fitting only fits a radio source in a single-frequency two-dimensional image, and rich information in multi-frequency data is not fully utilized, so that the information of the frequency dimension is split, and the processing efficiency is poor. The invention provides a multi-frequency radio power source modeling method based on a hybrid expert model and multi-Gaussian fitting, aiming at solving the problem that the overall distribution characteristics of a radio power source in space and frequency cannot be comprehensively considered during fitting, the invention provides a multi-frequency radio power source modeling method based on a hybrid expert model and multi-Gaussian fitting, and on the basis of a single-target detection method, a plurality of target detection methods are combined to form the hybrid expert model. A multi-frequency data cube obtained by a radio telescope array is effectively detected, a comprehensive three-dimensional model is established for a three-dimensional radio power source based on a detection result, the relationship among frequencies in the radio power source can be analyzed more correlatively, and the method is used for three-dimensional modeling and feature analysis of the multi-frequency radio power source in the radio astronomical data cube.

Description

Technical Field

[0001] The invention relates to the technical field of radio astronomy data cube processing, and in particular to a multi-frequency radio source modeling method based on a mixed expert model and multi-Gaussian fitting. Background Art

[0002] The observation frequency of radio telescopes covers a wide frequency band from hundreds of hertz to tens of gigahertz, enabling them to observe rich radio information of celestial bodies. These data are usually stored in the form of three-dimensional data cubes, where two dimensions correspond to spatial positions and the other dimension corresponds to the observation frequency. The foreground source (radio source) of the data cube contains important astrophysical information. Different settings of radio telescope equipment parameters produce different numbers of frequencies in the data cube, and the information between different frequencies of radio sources in the data cube is highly correlated, such as spectral energy distribution or morphological characteristics under frequency changes. However, existing technologies have significant limitations in processing such data. Traditional target detection methods only search for targets (radio sources) in single-frequency two-dimensional images. For multi-frequency data cubes, only each frequency can be detected, which is inefficient and consumes a lot of computing resources. At the same time, the correlation between adjacent frequencies is ignored and the detection index needs to be improved. Existing target detection methods show great limitations when processing complex scenes (such as dense star fields or low signal-to-noise ratio backgrounds) and are easily affected by noise interference or background complexity. Moreover, traditional target detection methods usually rely on a single detection model, which is difficult to cope with the diverse characteristics of radio sources, resulting in a significant decrease in performance in complex scenarios such as high density and extreme brightness (too bright or too dark). Traditional two-dimensional Gaussian fitting only fits radio sources in single-frequency two-dimensional images, fails to fully utilize the rich information in multi-frequency data, and divides the three-dimensional data cube into multiple two-dimensional images for processing, resulting in the fragmentation of information in the frequency dimension. It is impossible to fully consider the overall distribution characteristics of radio sources in space and frequency during fitting, making it difficult to accurately reveal the physical properties of radio sources, such as the evolution of spectrum distribution or structure with frequency. Summary of the invention

[0003] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a multi-frequency radio source modeling method based on a hybrid expert model and multi-Gaussian fitting. The invention can combine multiple target detection methods to form a hybrid expert model on the basis of a single target detection method, effectively detect the multi-frequency data cube obtained by the radio telescope array, and finally establish a comprehensive three-dimensional model of the three-dimensional radio source based on the detection results, which can analyze the relationship between the frequencies in the radio source in a more relevant manner, thereby facilitating subsequent scientific research and analysis.

[0004] To achieve the above object, the present invention provides the following technical solution, comprising the following steps:

[0005] A multi-frequency radio source modeling method based on a hybrid expert model and multi-Gaussian fitting, wherein a multi-frequency data cube is divided into n parts according to a training strategy, and each part is sent to n target detection methods for independent training. The optimal model of each target detection method is combined to form a hybrid expert model, and the hybrid expert model is used to train all data to generate the optimal model of the hybrid expert model, and perform target detection on all radio data, output the three-dimensional coordinate information of the radio source, fit the multi-frequency data cube using a multi-Gaussian fitting method, and finally reconstruct the output three-dimensional model. Specifically, the method comprises the following steps:

[0006] Step 1. Construct a radio image annotation dataset: Based on the equipment parameters of the radio telescope array and the observation data, the radio image annotation dataset is generated using the Oskar simulation software. The radio image annotation dataset contains a simulated multi-frequency observation data cube, the spatial position of the target in the multi-frequency observation data cube, and the frequency distribution position;

[0007] Step 2. Generate the best model of the target detection method: divide the radio image annotation dataset into n sub-datasets, train them using n target detection methods respectively, and obtain the best model of each target detection method;

[0008] Step 3. Generate the best model of the hybrid expert model: train n sub-datasets respectively through the best models of n target detection methods. During the training process, the hybrid expert model convolves and downsamples the outputs of the best models of the n target detection methods obtained in step 2, combines the initialized gating network, obtains dynamic weights and fuses the output results, and obtains the best model of the hybrid expert model based on the verification loss comparison, which is used to detect the multi-frequency observation data cube and obtain the category and location information of the multi-frequency observation data cube. The gating network dynamically adjusts the weight vector of the target detection method to improve the average precision, F1 score and recall rate of the hybrid expert model.

[0009] Step 4. Reconstruct the three-dimensional Gaussian model of the radio source: Based on the category and location information of the multi-frequency observation data cube obtained in step 3, cut the radio source into separate data cubes, build an initial multi-Gaussian model, minimize the difference between the initial multi-Gaussian model and the data cube through the loss function and minimize function, and solve the optimal parameters of the multi-Gaussian model in the data cube; reconstruct the three-dimensional model of the radio source based on the optimal parameters of the multi-Gaussian model, and perform visualization and residual evaluation.

[0010] Furthermore, in step 1, based on the characteristics that different radio sources have different observable frequency ranges, a deep convolutional neural network model with added frequency labels is established to directly predict the three-dimensional coordinates of the output radio source, the feature extraction backbone network of the deep convolutional neural network is optimized, and the channel attention mechanism is added. The deep convolutional neural network outputs the spatial position and frequency information of the target.

[0011] Furthermore, in step 2, the radio source is divided into sub-datasets with distinct characteristics according to its distribution characteristics, brightness characteristics and category characteristics, the number n of sub-datasets and target detection methods is determined, and according to the characteristics of each sub-dataset, the n sub-datasets are trained using n target detection methods respectively. The network parameters are optimized through repeated iterations to generate the best model of the n target detection methods to obtain different feature information in different sub-datasets.

[0012] Furthermore, in step 3, the hybrid expert model includes a gating network and convolution and downsampling of the outputs of each target detection method to obtain dynamic weights and fusion output results; the gating network includes four fully connected layers and activation functions and dropout functions between layers. The gating network inputs the feature vector of the data cube, dynamically adjusts and outputs the normalized weight of each target detection method, and realizes weighted fusion of the outputs of different target detection methods.

[0013] Furthermore, in step 4, the data cube is processed with 3 sigma to remove outliers and noise, and then normalized to ensure the convergence speed; the multi-Gaussian fitting method traverses different optimization method parameters to solve the optimal multi-Gaussian model parameters.

[0014] In summary, the invention has the following beneficial effects:

[0015] Based on the traditional target detection neural network, the present invention combines multiple target detection methods to form a hybrid expert model, fully utilizing the advantages of different target detection methods to achieve accurate detection of multi-frequency radio sources; the present invention can simultaneously consider the correlation between different frequencies and provide the spatial information and frequency distribution information of the target.

[0016] The present invention uses a multi-Gaussian fitting method to fit the three-dimensional target detected by the hybrid expert model and find the optimal multi-Gaussian parameters. These multi-Gaussian parameters combine spatial information and the correlation between multiple frequencies, and can directly construct a three-dimensional model of the emission source. Compared with the traditional two-dimensional space model and one-dimensional frequency model, the three-dimensional model of the present invention is more direct and comprehensive, more in line with physical interpretation, and has a smaller fitting residual, providing more comprehensive data information for subsequent astronomical research.

[0017] The present invention performs real-time end-to-end radio source detection and modeling analysis on a multi-frequency radio data cube, outputs reliable three-dimensional coordinates of the radio source and important astronomical parameters of multi-Gaussian parameters, and is used for three-dimensional modeling and feature analysis of multi-frequency radio sources in radio astronomy data cubes, and is particularly suitable for radio source target detection, frequency correlation extraction, and accurate multi-Gaussian fitting modeling in large-field-of-view radio observation data. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the present invention.

[0019] Figure 2 The figure shows the mean absolute error (MAE) analysis result of the multi-Gaussian fitting of the present invention for the multi-frequency radio source fitting with a frequency of 96. The horizontal axis is the serial number of the multi-frequency radio source, and there are 480 radio sources in total. The vertical axis is the mean absolute error (MAE) of the multi-Gaussian fitting of each radio source. DETAILED DESCRIPTION

[0020] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0021] It should be noted that, for the sake of convenience, the directions described below are consistent with the directions of the drawings themselves, but do not limit the structure of the present invention.

[0022] like Figure 1 As shown, the present invention discloses a multi-frequency radio source modeling method based on a hybrid expert model and multi-Gaussian fitting. On the basis of a single target detection method, multiple target detection methods are combined to form a hybrid expert model, and the multi-frequency data cube obtained by the radio telescope array is effectively detected. Finally, based on the detection results, a comprehensive three-dimensional model is established for the three-dimensional radio source, which can analyze the relationship between the frequencies in the radio source with more relevance, thereby facilitating subsequent scientific research and analysis, such as the evolution of the radio source spectrum distribution or the radio source structure with frequency. Specifically, the following steps are included:

[0023] Step 1. Construct radio image annotation dataset: According to the equipment parameters of the radio telescope array and the observation data, the radio image annotation dataset is generated using the Oskar simulation software. The radio image annotation dataset contains simulated multi-frequency observation data cubes, the spatial position and frequency distribution position of the target in the multi-frequency observation data cube. According to the observation principle, observation parameter setting and model of the radio array, the GLEAM (GaLactic and Extragalactic All-sky MWA) star catalog is first imported by using the Oskar simulation software. Then, a sky model is constructed based on the RA (right ascension), DEC (declination) and Ref_freq (reference frequency) variables. The starting frequency and frequency increment of the radio source observation are set to generate a data cube containing the target spatial position and frequency information. These data cubes constitute the annotation dataset. Since the radio source of radio observation can only be observed in a specific frequency range, the observable frequency range of different radio sources also varies. Therefore, in view of this feature, a deep convolutional neural network model was established. The deep convolutional neural network established has added frequency labels on the basis of the original input labels of only coordinates x and coordinates y, so that the deep convolutional neural network can directly predict the three-dimensional coordinates of the output radio source. In order to better obtain the characteristic information of radio data at different frequencies and the correlation between frequencies, the feature extraction backbone network of the deep convolutional neural network was optimized as follows: First, the number of input channels of the first layer of convolution kernel is set to the number of frequencies that can automatically match the input data cube, that is, the number of channels. Then, the number of output channels is adjusted according to the original input-output ratio coefficient. In order to avoid the problem of non-convergence during training due to too many network parameters, the threshold of the number of input channels of the convolution kernel is set. On the basis of the original feature extraction backbone network, a channel attention mechanism is added to extract the frequency distribution information of the data and the frequency information of each radio source. The specific adjustment is: the dimension of the input feature map is C×H×W, where C is the number of frequencies, and H and W are height and width respectively. First, the input feature map of the deep convolutional neural network is subjected to global average pooling, and average pooling is performed in the spatial dimension to obtain a global average pooling feature vector with a dimension of C. Then, the global average pooling feature vector is input into the first fully connected layer to generate an intermediate feature vector, and the dimension of the intermediate feature vector is still C. Subsequently, the intermediate feature vector is input into the second fully connected layer and the third fully connected layer, so that the deep convolutional network can learn to automatically match the complex relationship between position and frequency in the input data cube. Finally, two weight vectors with a dimension of C are obtained through the second fully connected layer and the third fully connected layer, which represent the frequency importance weight vector and the frequency suppression weight vector respectively.The frequency importance weight is multiplied by the input feature map to obtain a weighted feature map; the weighted feature map is then input into a sigmoid function to map the frequency suppression weight vector to a probability value within [0,1]; then, the frequency suppression weight vector is multiplied by the weighted feature map to obtain the final channel attention feature map, which is used to further extract feature information. Finally, the deep convolutional neural network output includes the spatial position coordinates and frequency information of the target.

[0024] Step 2. Generate the best model for the target detection method: Divide the radio image annotation dataset into n sub-datasets, train them using n target detection methods respectively, and obtain the best model for each target detection method. According to the distribution characteristics, brightness characteristics and category characteristics of the previously simulated radio sources, it is divided into "dense star field" dataset, "sparse star field" dataset, "bright source dense" dataset, "dark source dense" dataset, "category 1" dataset, and "category 2" dataset characteristics. Since the computing resources required for the subsequent hybrid expert model are linearly related to the number of single target detection methods, the number of datasets and target detection methods is selected according to the actual situation. According to the richness of computing resources, the number of specific sub-datasets and target detection methods n is determined, and then these n sub-datasets are input into n target detection methods for training. The selected single target detection methods include YOLOX and YOLOv5. By matching each sub-dataset with a specific target detection method, the advantages of different target detection methods are fully utilized. For dense star fields, the target detection method needs to accurately distinguish radio sources in complex backgrounds; while for sparse star fields, the target detection method needs to capture the characteristics of a single radio source more sensitively. Similarly, for bright source data, the target detection method needs to avoid the oversaturation effect and accurately identify high-brightness radio sources; while for dim source data, the target detection method needs to have higher sensitivity to capture weak signals. During the training process, the single target detection method optimizes the network parameters through repeated iterations, and finally generates the corresponding optimal target detection method model file. These target detection method model files contain the optimal parameter weights for detecting radio source targets. In this way, each target detection method can focus on its own strengths in feature extraction and target detection during the training process, thereby maximizing the extraction of radio source feature information in each sub-dataset, and providing high-quality target detection method models for subsequent hybrid expert model training. In this process, each sub-dataset has a corresponding training model and method, and the training of each model is independent, and each target detection method generates the optimal target detection method model.

[0025] Step 3. Generate the best model of the hybrid expert model: train n sub-datasets respectively through the best models of n target detection methods. During the training process, the hybrid expert model convolves and downsamples the outputs of the best models of the n target detection methods obtained in step 2, combines the initialized gating network, obtains dynamic weights and fuses the output results, and obtains the best model of the hybrid expert model based on the verification loss comparison, which is used to detect multi-frequency observation data cubes and obtain the category and location information of the multi-frequency observation data cubes. The gating network dynamically adjusts the weight vector of the target detection method to improve the average precision, F1 score and recall rate of the hybrid expert model.

[0026] The gating network includes four fully connected layers, and activation functions (ReLU / LeakyReLU) and dropout functions are added between layers to improve generalization ability. The input of the gating network is the feature vector of the data cube. After multiple layers of fully connected layers and dynamic adjustment of nonlinear transformations, the normalized weights of each expert are output to achieve weighted fusion of the outputs of different target detection methods. The hybrid expert model includes an initialized gating network and convolution and downsampling of the outputs of each expert, i.e., the target detection method, and finally dynamic weight calculation and fusion output results. The n target detection model files and n sub-datasets in step 2 are loaded into the hybrid expert model at one time for training. During the training process, each sub-dataset generates a specific feature output through the corresponding target detection method (such as YOLOX, YOLOv5, etc.). The gating network (GatingNetwork) generates a set of weights through the fully connected layer combined with the feature information of the sub-dataset, and each sub-dataset corresponds to a weight. In this process, the gating network dynamically assigns weights to different experts (target detection methods). Specifically, for different feature distributions in the input data cube, such as the scale, shape complexity or background noise level of the target, some target detection methods may perform better in specific indicators (such as detection accuracy, recall rate or inference time). According to the characteristics of the input data cube and the performance of each expert (target detection method) on these indicators, the gating network assigns higher weights to the experts (target detection methods) that are suitable for the data characteristics, thereby achieving weighted fusion of the outputs of different target detection methods. Through repeated iterative optimization, the hybrid expert model can gradually improve the detection capabilities of different target detection methods at different frequencies in different data cubes. After training, the hybrid expert model generates an optimal model file, which is used by the hybrid expert model to detect the category and location information of the data cube. The gating network in it will dynamically adjust the vector of the expert (target detection method) weight. The size of the weight reflects the degree of adaptation of each expert (target detection method) to the current input, thereby controlling the contribution of the output of each expert (target detection method). This not only detects the two-dimensional spatial position of the target in the radio data (in the form of a detection box), but also analyzes and outputs the frequency number, revealing the correlation between different frequencies. Furthermore, the detection index output can be improved. Specifically, the average precision (AP), F1 score and recall rate (Recall) of the hybrid expert model are effectively improved compared with the single target detection method. Table 1 is a comparison of the detection indicators of the single target detection method and the hybrid expert model method when score_threhold = 0.5.

[0027] Table 1

[0028] score_threhold = 0.5 AP F1 P R yolox 62.14％ 0.72 96.48％ 56.84％ yolov5 62.52％ 0.73 98.57％ 58.00％ Mixed Expert Model MOE 65.90％ 0.74 97.25％ 59.62％

[0029] Step 4. Reconstruct the three-dimensional Gaussian model of the radio source: Based on the category and location information of the multi-frequency observation data cube obtained in step 3, cut the radio source into separate data cubes, build an initial multi-Gaussian model, minimize the difference between the initial multi-Gaussian model and the data cube through the loss function and minimize function, and solve the optimal parameters of the multi-Gaussian model in the data cube; reconstruct the three-dimensional model of the radio source based on the optimal parameters of the multi-Gaussian model, and perform visualization and residual evaluation.

[0030] Based on the two-dimensional coordinate position and frequency of the radio source detected by the hybrid expert model, the radio source is cut into separate data cubes, and the data cubes are processed with 3sigma to eliminate the influence of outliers and noise, and then normalized to ensure that the subsequent optimization method converges quickly. Then the initial multi-Gaussian model is constructed from the data cube. The initial mean and covariance parameters are derived through the statistical characteristics of the data cube. During the fitting process, the minimize function in the scipy library will use the loss function and the built-in parameter optimization method provided by the minimize function to minimize the difference between the model and the data cube, thereby solving the optimal parameters of the multi-Gaussian model in the data cube. In addition, the multi-Gaussian fitting method will also traverse different optimization method parameters to solve the optimal multi-Gaussian model parameters. Then, based on the multi-Gaussian model parameters (multi-Gaussian mean and multi-Gaussian covariance), the three-dimensional model of the radio source is reconstructed, and visualization and residual evaluation are performed. Mean absolute error (MAE) evaluation, Figure 2 The experimental results show that the mean absolute error (MAE) of the multi-Gaussian fitting results of 469 out of 480 radio sources is within 0.01, and the evaluation results are reliable.

[0031] The present invention combines the loss function, the initial multi-Gaussian parameters and the data with the library function calculation to find the multi-Gaussian parameters with the minimum loss, and in this state, the optimal multi-Gaussian parameters can be identified. The traditional two-dimensional fitting and one-dimensional frequency fitting separate the three-dimensional radio source in the data cube as a whole, and do not consider it as a whole. The multi-Gaussian fitting method of the present invention fits the spatial distribution of the radio source more comprehensively and directly than the two-dimensional fitting and one-dimensional frequency fitting, and the multi-Gaussian model parameters provided are also of great significance to the subsequent physical analysis of astronomy.

[0032] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A multi-frequency radio source modeling method based on a hybrid expert model and multi-Gaussian fitting, characterized in that: The multi-frequency data cube is divided into n parts according to the training strategy, and each part is sent to n target detection methods for independent training. The optimal model of each target detection method is combined to form a hybrid expert model. The hybrid expert model is used to train all data, generate the optimal model of the hybrid expert model, and perform target detection on all radio data. The three-dimensional coordinate information of the radio source is output, and the multi-frequency data cube is fitted using the multi-Gaussian fitting method. Finally, the output three-dimensional model is reconstructed. The specific steps include: Step 1. Construct a radio image annotation dataset: Based on the equipment parameters of the radio telescope array and the observation data, the radio image annotation dataset is generated using the Oskar simulation software. The radio image annotation dataset contains a simulated multi-frequency observation data cube, the spatial position of the target in the multi-frequency observation data cube, and the frequency distribution position; Step 2. Generate the best model of the target detection method: divide the radio image annotation dataset into n sub-datasets, train them using n target detection methods respectively, and obtain the best model of each target detection method; Step 3. Generate the best model of the hybrid expert model: train n sub-datasets respectively through the best models of n target detection methods. During the training process, the hybrid expert model convolves and downsamples the outputs of the best models of the n target detection methods obtained in step 2, combines the initialized gating network, obtains dynamic weights and fuses the output results, and obtains the best model of the hybrid expert model based on the verification loss comparison, which is used to detect the multi-frequency observation data cube and obtain the category and location information of the multi-frequency observation data cube. The gating network dynamically adjusts the weight vector of the target detection method to improve the average precision, F1 score and recall rate of the hybrid expert model. Step 4. Reconstruct the three-dimensional Gaussian model of the radio source: Based on the category and location information of the multi-frequency observation data cube obtained in step 3, cut the radio source into separate data cubes, build an initial multi-Gaussian model, minimize the difference between the initial multi-Gaussian model and the data cube through the loss function and minimize function, and solve the optimal parameters of the multi-Gaussian model in the data cube; reconstruct the three-dimensional model of the radio source based on the optimal parameters of the multi-Gaussian model, and perform visualization and residual evaluation.

2. The multi-frequency radio source modeling method based on a hybrid expert model and multi-Gaussian fitting according to claim 1, characterized in that: In the step 1, according to the characteristic that the observable frequency ranges of different radio sources are different, a deep convolutional neural network model with added frequency labels is established to directly predict the three-dimensional coordinates of the output radio source, the feature extraction backbone network of the deep convolutional neural network is optimized, and the channel attention mechanism is added, and the deep convolutional neural network outputs the spatial position and frequency information of the target.

3. The multi-frequency radio source modeling method based on a hybrid expert model and multi-Gaussian fitting according to claim 1, characterized in that: In step 2, the radio source is divided into sub-datasets with distinct characteristics according to its distribution characteristics, brightness characteristics and category characteristics, the number n of sub-datasets and target detection methods is determined, and the n sub-datasets are trained using n target detection methods according to the characteristics of each sub-dataset. The network parameters are optimized through repeated iterations to generate the best model of the n target detection methods to obtain different feature information in different sub-datasets.

4. The multi-frequency radio source modeling method based on a hybrid expert model and multi-Gaussian fitting according to claim 1, characterized in that: In step 3, the hybrid expert model includes a gated network and convolution and downsampling of the outputs of each target detection method to obtain dynamic weights and fusion output results; The gating network consists of four fully connected layers and activation functions and dropout functions between layers. The gating network inputs the feature vector of the data cube, dynamically adjusts and outputs the normalized weight of each target detection method, and realizes the weighted fusion of the outputs of different target detection methods.

5. The multi-frequency radio source modeling method based on a hybrid expert model and multi-Gaussian fitting according to claim 1, characterized in that: In step 4, the data cube is processed by 3sigma to remove outliers and noise, and then normalized to ensure the convergence speed; The multi-Gaussian fitting method traverses different optimization method parameters to solve the best multi-Gaussian model parameters.

Images