Radiation source identification model training method and identification method based on random connection network

Abstract

The invention discloses a radiation source identification model training method and identification method based on a random connection network, and relates to the technical field of target identification. The training method comprises the steps of performing feature extraction and feature dimension reduction on an electromagnetic sample signal to obtain a sample feature vector, and performing normalization and quantization processing on the sample feature vector; performing density coding on the normalized and quantized sample feature vectors to obtain bit vectors; performing super-dimensional coding on the bit vector to obtain a super-dimensional vector feature; and taking the super-dimensional vector features as input, training a random connection neural network, and obtaining a radiation source identification model. The method is based on the density coding technology and the super-dimensional calculation technology, is combined with the machine connection neural network, reduces the network calculation amount, accelerates the model reasoning speed, has the advantages of high recognition precision, high safety, high reliability and the like, and can be used on edge equipment under the condition that resources are limited.

Description

Technical Field

[0001] The present invention relates to the technical field of target recognition, and in particular to a radiation source recognition model training method and a recognition method based on a random connection network. Background Art

[0002] In the traditional convolutional neural network-based radiation source individual identification scheme, due to problems such as many parameters, large models, complex back propagation calculations and high power consumption, it cannot be well used in resource-constrained edge networks.

[0003] The random-link neural network (FLNN) can include convolution operations but does not require back propagation. It does not have the long learning time and slow convergence speed of convolutional neural networks, and can be used in scenarios with limited computing resources. However, since the bias and weight parameters of the random-link neural network model are randomly generated, the model's individual radiation source identification accuracy and generalization ability are poor.

[0004] Therefore, in order to quickly and accurately identify individual radiation sources in resource-constrained scenarios, it is necessary to improve the existing neural network-based individual radiation source identification method to improve the recognition accuracy and model generalization ability and reduce the amount of calculation and power consumption. Summary of the invention

[0005] The present invention provides a radiation source identification model training method and an identification method based on a randomly connected network, so as to solve the problem of how to quickly and accurately identify radiation sources under the condition of limited resources.

[0006] The present invention is achieved through the following technical solutions:

[0007] A first aspect of the present invention provides a radiation source identification model training method based on a randomly connected network, comprising:

[0008] Extracting features and reducing the dimension of features on the electromagnetic sample signal to obtain a sample feature vector, and normalizing and quantizing the sample feature vector;

[0009] Density encoding is performed on the sample feature vector after normalization and quantization to obtain a bit vector;

[0010] Performing hyperdimensional encoding on the bit vector to obtain a hyperdimensional vector feature;

[0011] The hyper-dimensional vector feature is used as input to train a randomly connected neural network to obtain a radiation source recognition model.

[0012] The present invention utilizes the characteristic that a randomly connected neural network does not require complex back-propagation calculations, and trains it as a radiation source identification model. Since the randomly connected neural network uses randomly generated model biases and weights, the recognition accuracy and generalization ability are poor. In order to improve the recognition accuracy of the randomly connected neural network and further reduce the amount of calculation and power consumption, the present invention performs feature dimension reduction, density coding and hyper-dimensional coding on the features of the input training model. The key feature parameters are obtained by feature dimension reduction, the amount of data is reduced without affecting the feature expression, and the floating-point input data is converted into integer data by normalization and quantization, thereby improving the model's data processing efficiency. And by performing density coding and hyper-dimensional coding on scalar features, the model's ability to learn difficult samples is improved, while data privacy is also protected. Based on density coding, there is no need to binarize the input features. Among them, density coding is a coding method based on information density. It does not simply convert information into fixed symbols, but considers the density distribution of information for encoding. Compared with traditional coding, density coding pays more attention to the statistical characteristics and storage efficiency of information; hyperdimensional coding uses high-dimensional space to represent information. It maps information to a high-dimensional vector space, and each information unit (such as a word, an image feature, etc.) corresponds to a high-dimensional vector. Traditional coding is mainly a symbolic representation, while hyperdimensional coding is based on the representation of vector space. The operation of hyperdimensional coding is based on simple arithmetic operations of high-dimensional vectors and has a high degree of parallelism. During data processing, density coding can be combined with the parallel computing characteristics of hyperdimensional coding to further improve computing efficiency. It can reduce the storage space of the model without losing model performance, making it more suitable for resource-constrained devices and scenarios with high real-time requirements. Therefore, compared with the traditional target recognition method based on randomly connected neural networks, the method of the present invention has a faster network reasoning speed and more reliable recognition results.

[0013] In one implementation, density encoding the normalized and quantized sample feature vector to obtain a bit vector includes:

[0014] Bipolar encoding is used to encode the scalar number n as n -1s, with -1 occupying the first n bits and the other bits filled with 1;

[0015] Bipolar encoding is used for each scalar number in the quantized sample feature vector to obtain a D×N-dimensional bit vector, wherein D is the dimension of the sample feature vector and N is the quantization representation dimension.

[0016] In one implementation, performing super-dimensional encoding on the bit vector to obtain a super-dimensional vector feature includes:

[0017] Randomly generate an input vector of the same dimension as the bit vector, and perform Hadamard multiplication on the bit vector and the input vector to obtain a product vector;

[0018] The product vector is summed up by column to obtain a super-dimensional vector feature.

[0019] In one embodiment, before taking the hyperdimensional vector feature as input and training the randomly connected neural network, the method further includes: performing wave elimination processing on the hyperdimensional vector feature.

[0020] In one embodiment, the randomly connected neural network includes a convolutional network, a single-layer feedforward neural network and an output layer connected in sequence; the single-layer feedforward neural network includes a first linear layer, a second linear layer and an activation layer, and a direct link is added between the linear layer and the activation layer to form a nonlinear transformation from the input layer to the second linear layer, a linear transformation from the second linear layer to the output layer, and a linear transformation from the input layer to the output layer.

[0021] In one embodiment, the training method of the randomly connected neural network is:

[0022] Step 1, randomly initializing the weights and biases of the randomly connected neural network, fixing the second parameter, and obtaining a first output through the activation layer of the single-layer feedforward neural network, wherein the second parameter is the weight and bias between the input layer and the second linear layer;

[0023] Step 2, based on a linear regression formula, solving a first parameter and a third parameter from the first output, wherein the first parameter is a weight between the input layer and the output layer, and the third parameter is a weight between the second linear layer and the output layer;

[0024] Step 3: Use the first parameter, the second parameter and the third parameter as network parameters of the randomly connected neural network to obtain a trained radiation source recognition model.

[0025] In one embodiment, the training method further comprises:

[0026] Generate different combinations of hyperparameters, and obtain different radiation source recognition models based on the different combinations of hyperparameters training;

[0027] The classification accuracy of the radiation source identification model is evaluated through a test set, and the optimal model is selected for radiation source identification.

[0028] A second aspect of the present invention provides a radiation source identification method, comprising:

[0029] Acquiring detected electromagnetic signals;

[0030] Extracting a characteristic vector from the electromagnetic signal, and performing normalization and quantization processing on the characteristic vector;

[0031] Density encoding is performed on the normalized and quantized feature vector to obtain a bit vector;

[0032] Performing hyperdimensional encoding on the bit vector to obtain a hyperdimensional vector feature;

[0033] The hyperdimensional vector feature is input into a radiation source identification model to obtain the type of the electromagnetic signal; the radiation source identification model is trained by the radiation source identification model training method based on a random connection network described in any one of the first aspects of the present invention.

[0034] The third aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the radiation source identification model training method based on a randomly connected network as described in any one of the first aspect of the present invention, or the radiation source identification method as described in the second aspect of the present invention.

[0035] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that when the computer program is executed by a processor, the radiation source identification model training method based on a randomly connected network as described in any one of the first aspect of the present invention, or the radiation source identification method as described in the second aspect of the present invention is implemented.

[0036] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0037] The feature characterization technology based on density coding and hyperdimensional coding is combined with a randomly connected neural network to identify individual radiation sources, forming a resource-efficient full-integer randomly connected neural network, which can improve the accuracy of individual radiation source identification, accelerate the reasoning speed, and meet the use of FPGA-based edge terminals under resource-constrained conditions.

[0038] Based on the design idea of ​​regularized density coding, it can improve the model recognition accuracy and enhance the model generalization ability.

[0039] Based on the design idea of ​​binding and bundling operations of hyperdimensional computing, the network input is made into integer type, which effectively improves the algorithm reasoning speed.

[0040] A convolutional randomly connected neural network is used, which contains convolution operations but does not require back-propagation calculations, shortening the model learning time, making the model converge quickly, and reducing training costs.

[0041] By performing wave elimination on the input features, the storage memory occupied by the features is reduced and resource consumption is reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0043] Figure 1 is a flow chart of a radiation source identification model training method based on a randomly connected network according to an embodiment of the present invention;

[0044] Figure 2 It is an overall framework diagram of an embodiment of the present invention based on density coding, hyperdimensional coding and randomly connected neural network;

[0045] Figure 3 is a schematic diagram of a density coding method according to an embodiment of the present invention;

[0046] Figure 4 Schematic diagram of the structure of a convolutional randomly connected neural network according to an embodiment of the present invention;

[0047] Figure 5 This is an algorithm flow chart of a convolutional random connection neural network in an embodiment of the present invention.

[0048] Figure 6 It is a schematic diagram of model training and recognition according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0050] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to or inherent to other steps or units of the device.

[0051] The terms used in the various embodiments of the application are only used to describe the purpose of specific embodiments and are not intended to limit the various embodiments of the application. As used herein, the singular form is intended to also include the plural form, unless the context clearly indicates otherwise. Unless otherwise limited, all terms used here (including technical terms and scientific terms) have the same meaning as the meanings commonly understood by ordinary technicians in the field of the various embodiments of the application. The terms (such as the terms defined in the dictionary generally used) will be interpreted as having the same meaning as the contextual meaning in the relevant technical field and will not be interpreted as having an idealized meaning or an overly formal meaning, unless clearly defined in the various embodiments of the application.

[0052] The present invention trains a randomly connected neural network as a radiation source identification model. In order to improve the accuracy of the randomly connected neural network model, reduce the amount of calculation and power consumption, and meet the use of the algorithm on resource-constrained edge devices, a new radiation source individual identification method of a randomly connected neural network is proposed by combining the density coding technology of random computing, the binding and bundling technology of hyperdimensional coding, and the advantages of fast reasoning of randomly connected neural networks. The method solves the problems of low accuracy, slow speed, and weak generalization ability of the randomly connected neural network model, thereby meeting the use conditions of edge devices and having the advantages of low latency, good interpretability, high security, data privacy protection, and high reliability.

[0053] like Figure 1 The figure is a flow chart of a radiation source identification model training method based on a randomly connected network of the present invention, which includes the following steps.

[0054] S1, extracting features and reducing the dimension of electromagnetic sample signals to obtain sample feature vectors, and normalizing and quantizing the sample feature vectors.

[0055] S2, density encoding is performed on the normalized and quantized sample feature vector to obtain a bit vector.

[0056] S3, perform hyper-dimensional encoding on the bit vector to obtain a hyper-dimensional vector feature.

[0057] S4, taking the hyperdimensional vector features as input, training the randomly connected neural network to obtain a radiation source recognition model.

[0058] First, an electromagnetic sample signal is obtained, including a radiation source of known type and the electromagnetic signal radiated by it, and the radiation source type is used as a sample label of the electromagnetic sample signal. In step S1, principal component analysis (PCA), linear discriminant analysis (LDA), and singular value decomposition (SVD) can be used for feature extraction and dimensionality reduction.

[0059] Taking principal component analysis as an example, we first use principal component analysis to extract useful features of electromagnetic signals, reduce the dimensionality of the signal, and extract the key features of the signal to form a feature vector. The dimension of the feature vector can be set, and each feature vector is D-dimensional, with a value range of [20,100].

[0060] During the training process, the sample data can be divided into a training set and a test set in proportion. The recognition model is trained by the training set, and the performance of the trained model is tested by the test set. The feature vector after dimension reduction by principal component analysis forms a data set of M categories (M≥1), and the size of each category L is determined according to the actual situation. If the total number of samples in each category of the data set is 600, the test set and training set are divided according to the ratio of 8:2, with 480 samples in the training set and 160 samples in the test set.

[0061] like Figure 2 The figure shows the overall framework diagram based on density coding, hyperdimensional coding and randomly connected neural network. First, the sample feature vector is normalized and quantized, the floating point data is converted into integer data, and then feature encoding is performed, which can effectively improve the algorithm reasoning speed and accuracy.

[0062] In one embodiment, the feature vector is normalized to the range of [0, 1] using a minimum-maximum normalization operation, and the normalization formula is as follows:

[0063]

[0064] Among them, x i is the i-th eigenvalue of the current eigenvector, x min and x max represents the minimum and maximum values ​​of the current eigenvector, Represents the i-th eigenvalue of the current eigenvector after normalization.

[0065] In one embodiment, the quantization process determines the quantized value according to the dimension of the final vector representation.

[0066] For example, the feature vector dimension D is 4, and the dimension N of the quantitative representation is set to 10. Then, according to the formula quantization, Round(·) represents the rounding function. The actual value of N will be determined according to the classification accuracy, and the value of N is relatively large. The feature vector reduced by principal component analysis is quantized to an integer in the range of [0,1] through the quantization operation, such as Figure 2 As shown, the 4-dimensional feature vector [0.423, 0.638, 0.967, 0.128] becomes [4, 6, 10, 1] after quantization, which satisfies the integer in [0, 10].

[0067] The core idea of ​​stochastic computing is that Boolean operations on bit vectors can be used to implement efficient algorithmic operations on scalars. In the field of stochastic computing, scalars are often represented as random bit vectors, where the encoded scalar determines the probability of generating 1. The method of encoding a scalar using the density of 1s in the bit vector (the proportion of 1s in the total number of bits) is called density-based encoding. This method does not require the representation of the scalar encoding to be random, so a structured density-based encoding method is used.

[0068] In one implementation, just like the binary coding method, that is, 1 represents 1 and 0 represents 0, this embodiment uses bipolar coding to perform density encoding on the scalar feature, and encodes the scalar number n as n -1s, -1 occupies the first n bits, and the other bits are filled with 1, that is, for the quantized integer value, there are as many -1s as the integer, and the other coding bits are 1, and the other coding bits are 1.

[0069] like Figure 3 The figure shows a schematic diagram of the density coding method. The third feature in the quantized features [0,1,2] is quantized to an integer of 2. There are 2 red circles, which are used to represent the number of -1 (i.e. 2 -1), and the light blue represents the number of 1.

[0070] Bipolar encoding is used for each scalar number in the quantized sample feature vector to obtain a D×N dimensional bit vector, where D is the dimension of the sample feature vector and N is the quantization representation dimension. For example, for the quantized feature [4, 6, 10, 1], the quantization representation dimension is N = 10, and the encoded bit vector is a 4×10 vector, which is expressed as:

[0071]

[0072] Hyperdimensional coding mainly uses the binding operation and bundling operation in hyperdimensional computing. The binding operation multiplies the density-encoded feature vector and the randomly generated input vector, and the bundling operation sums the temporary feature matrix calculated by the binding operation by column.

[0073] Specifically, an input vector with the same dimension as the bit vector is first randomly generated, and then the bit vector is Hadamard multiplied with the randomly generated input vector to obtain a product vector, and the dimension of the product vector remains unchanged. Figure 2 The random weight matrix Win is generated as shown. The element values ​​of the matrix Win are randomly generated and drawn from {-1,1} with equal probability. Therefore, Win is also a bipolar coding matrix, that is, D N-dimensional bipolar super-dimensional vectors. The temporary feature matrix is ​​composed of a binding operation between the random input super-dimensional vector and its current density-based coding feature vector, that is, multiplying the corresponding elements of the vector to obtain a temporary product vector.

[0074] Finally, the product vector cases are summed to obtain the hyperdimensional vector feature after hyperdimensional encoding, which is an N-dimensional vector, such as Figure 2 As shown in Figure 1, this step is equivalent to encrypting the input vector and reducing the dimension of the bit vector, and inputting the feature into the randomly connected neural network for model training. Due to the pre-processing of principal component extraction and feature encoding, the recognition accuracy and model generalization ability of the randomly connected neural network are improved.

[0075] Furthermore, before training, the hyper-dimensional vector features are also subjected to wave elimination processing.

[0076] When matrices are added column by column, the eigenvector space (the value of the vector) becomes unrestricted. Therefore, in order to reduce the computational complexity and control the range of the eigenvalues, a clipping function is used to limit the value of the result so that the eigenvector input is always an integer in the range [-d, d]. The clipping function in this article is as follows:

[0077]

[0078] Here, d is a configurable threshold parameter. In addition, even after clipping, the eigenvector eigenvalues ​​are integers in the range [-d, d] (d <D), Figure 2 The clipping operation shown in is clipped according to d=2. Therefore, each eigenvector eigenvalue can be represented using only log2(2d+1) bits of memory. For example, when d=2, there are 5 representations of the eigenvalue, which only require three bits to store, saving a lot of resource usage.

[0079] In one embodiment, the model training uses a convolutional randomly connected neural network as a classifier. The structural diagram of the convolutional randomly connected neural network is as follows: Figure 4As shown, it includes a convolutional network, a single-layer feedforward neural network and an output layer connected in sequence. The convolutional network is composed of multiple convolutional blocks, each of which is composed of one-dimensional convolution, batch normalization, an activation function (ReLU) and maximum pooling. The single-layer feedforward neural network includes a first linear layer (linear layer 1), a second linear layer (linear layer 2) and an activation layer. A direct link is added to one side of the single-layer feedforward neural network, that is, a direct link from linear layer 1 to the activation layer. The addition of this linear model enhances the feature representation capability of the network.

[0080] For model training, the random link network first randomly generates the input weight W and bias b of the enhancement node (neuron node of linear layer 2), and then uses pseudo-inverse analysis to determine the weight connected to the output layer. Compared with the random connection neural network, the convolutional neural network has an additional convolution operation. Specifically, for a given input X (feature vector after clipping operation), such as Figure 4 As shown in the figure, all weights and biases between it and the linear layer 2 are randomly generated, and the formula F=WX+b is used to pass down layer by layer until it is passed to the linear layer 2 (also called the enhancement layer). This set of weights remains unchanged after the linear layer 2. The feature vector passed to the linear layer 2 is obtained by the activation function to obtain U. After that, the original input X and the calculation result U of the linear layer 2 are combined into a matrix V, V=[X|U]. V is multiplied by the weight and added to the bias and then passed to the output layer to obtain the classification result Y'.

[0081] There are three types of link relationships in the convolutional random connection neural network, as shown below:

[0082] (1) Input layer (convolutional network input X) → Linear layer 2: nonlinear transformation;

[0083] (2) Linear layer 2 → output layer: linear transformation;

[0084] (3) Input layer → output layer: linear transformation.

[0085] Among these three connections, only the linear layer 2 has an activation function and has the characteristics of a neural network unit. Arranging the linear layer and the input node together, the network becomes a linear transformation from V to Y'. If all the random weights and random biases in the convolutional random connection neural network are fixed, the training of the entire network is to find the linear transformation W from V to label Y:

[0086] W=V -1 Y

[0087] Among them, the input X is known (the feature vector after the clipping operation), and the predicted output V, V -1 represents the inverse of V, and Y is the known quantity (the label vector of input X). The weight matrix W is solved using the linear regression formula. Considering the computational efficiency, the ridge regression method is used in the actual calculation, and the solution is:

[0088] W=(λI+VV T ) -1 V T Y

[0089] Where I is the identity matrix, the diagonal is all 1, λ is the ridge coefficient, changing the value of λ can change the value of the diagonal of the identity matrix, V T is the transpose of V.

[0090] The algorithm flow of the convolutional randomly connected neural network is as follows Figure 5 As shown, it includes the following operation steps:

[0091] Step 1: Randomly initialize the weights and biases of the randomly connected neural network, fix the second parameter (parameter 2, representing all weights and biases between the input layer and the linear layer 2), and pass the activation layer of the linear layer 2 to obtain the first output V.

[0092] Step 2: Solve the weight matrix according to the linear regression formula (the ridge regression formula W = (λI + VV T ) -1 V T Y), calculate the first parameter (parameter 1) and the third parameter (parameter 3) from the first output V. Parameter 1 represents the mapping weight between the input layer and the output layer, and parameter 3 represents the mapping weight between the input layer and the enhancement layer. These two weights are calculated by the ridge regression formula and a set of input and output data.

[0093] Step 3: Use the first parameter, the second parameter and the third parameter as network parameters of the randomly connected neural network to obtain a trained radiation source recognition model.

[0094] Use the fixed second parameter, the first parameter in step 2, and the third parameter to predict the input data X and obtain the second output Y, i.e. the prediction result. Since the first parameter, the third parameter and the three parameters are combined into the weight W and have been calculated by steps 1-2, there is no need to repeat the training calculation, and prediction and reasoning can be performed directly, saving a lot of training time.

[0095] In one embodiment, during the model training process, the randomly connected network is trained using different hyperparameters to generate different recognition models, and then the best model is selected according to the classification accuracy index of the test set as the final radiation source recognition model.

[0096] In the parameter optimization training scheme based on ridge regression, the hyperparameters include the quantization representation dimension N, the wavelet threshold d, and the ridge coefficient λ. The values ​​of N, d, and λ are selected by grid search to find the optimal parameters. Set the search range and search step of N, d, and λ. As an example, N varies in the range of [500, 11000] with a step size of 50, λ varies in the range of 2[-3, 5] with a step size of 1, and the value range of d is {1, 2, 3, 5, 7, 10, 15}. Evaluate different hyperparameter models through the test set to find the optimal model.

[0097] In summary, the present invention proposes a new type of radiation source individual identification method of randomly connected neural network. By introducing density coding technology in the field of random computing, binding and bundling technology of hyperdimensional computing, and convolutional randomly connected neural network technology in randomly connected neural network, it not only reduces the amount of network calculation, speeds up the reasoning speed, and improves the recognition accuracy, but also can meet the use of FPGA-based edge devices under resource-constrained conditions. It has the advantages of higher security, data privacy protection and high reliability. It is a resource-efficient full-integer randomly connected neural network. The proposed density-based coding method does not require binarization of input features and will not reduce classification accuracy. Compared with the traditional random connection neural network full floating point type on FPGA, the resource-efficient full-integer random connection network has a faster reasoning speed; compared with the classification accuracy of the traditional random connection neural network, the classification accuracy of the method proposed in this scheme can be improved by more than 10%.

[0098] A second aspect of the present invention provides a radiation source identification method, comprising the following steps.

[0099] S11, acquiring the detected electromagnetic signal;

[0100] S12, extracting a feature vector from the electromagnetic signal, and normalizing and quantizing the feature vector;

[0101] S13, performing density encoding on the normalized and quantized feature vector to obtain a bit vector;

[0102] S14, performing hyper-dimensional encoding on the bit vector to obtain a hyper-dimensional vector feature;

[0103] S15, inputting the hyperdimensional vector feature into a radiation source identification model to obtain the type of electromagnetic signal.

[0104] See also Figure 6 The model training and recognition schematic diagram shown in the figure shows that in the recognition method, the processing of the input features is the same as the processing of the sample data. The radiation source recognition model is trained using the radiation source recognition model training method based on the random connection network of the present invention, which will not be described in detail here.

[0105] In a third aspect of the present invention, an electronic device is provided, the electronic device comprising a processor and a memory, and the number of processors may be one or more. The memory, as a computer-readable storage medium, can be used to store software programs, computer executable programs, and modules. The processor executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory, so as to implement the radiation source identification model training method or radiation source identification method based on a random connection network of any of the above embodiments of the present invention.

[0106] The memory may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required for a function; the data storage area may store data created according to the use of the terminal, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage devices.

[0107] In some examples, the memory may further include a memory remotely located relative to the processor, and these remote memories may be connected to the electronic device via a network. Examples of the above network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0108] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a radiation source identification model training method or a radiation source identification method based on a randomly connected network according to any embodiment of the present invention.

[0109] According to a fifth aspect of the present invention, a computer program product is provided. When the computer program product is run on a computer, the computer executes the radiation source identification model training method or radiation source identification method based on a randomly connected network according to any of the above embodiments of the present invention.

[0110] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A radiation source identification model training method based on a randomly connected network, characterized in that: include: Extracting features and reducing the dimension of features on the electromagnetic sample signal to obtain a sample feature vector, and normalizing and quantizing the sample feature vector; Density encoding is performed on the sample feature vector after normalization and quantization to obtain a bit vector; Performing hyperdimensional encoding on the bit vector to obtain a hyperdimensional vector feature; The hyper-dimensional vector feature is used as input to train a randomly connected neural network to obtain a radiation source recognition model.

2. The radiation source identification model training method based on a randomly connected network according to claim 1 is characterized in that: The step of performing density encoding on the normalized and quantized sample feature vector to obtain a bit vector comprises: Bipolar encoding is used to encode the scalar number n as n -1s, with -1 occupying the first n bits and the other bits filled with 1; Bipolar encoding is used for each scalar number in the quantized sample feature vector to obtain a D×N-dimensional bit vector, wherein D is the dimension of the sample feature vector and N is the quantization representation dimension.

3. The radiation source identification model training method based on a randomly connected network according to claim 1 is characterized in that: The step of performing super-dimensional encoding on the bit vector to obtain a super-dimensional vector feature includes: Randomly generate an input vector of the same dimension as the bit vector, and perform Hadamard multiplication on the bit vector and the input vector to obtain a product vector; The product vector is summed up by column to obtain a super-dimensional vector feature.

4. The radiation source identification model training method based on a randomly connected network according to claim 3 is characterized in that: Before taking the hyper-dimensional vector feature as input and training the randomly connected neural network, the method further comprises: performing wave elimination processing on the hyper-dimensional vector feature.

5. The radiation source identification model training method based on a randomly connected network according to claim 1 is characterized in that: The randomly connected neural network includes a convolutional network, a single-layer feedforward neural network and an output layer connected in sequence; the single-layer feedforward neural network includes a first linear layer, a second linear layer and an activation layer, and a direct link is added between the linear layer and the activation layer to form a nonlinear transformation from the input layer to the second linear layer, a linear transformation from the second linear layer to the output layer, and a linear transformation from the input layer to the output layer.

6. The radiation source identification model training method based on a randomly connected network according to claim 5 is characterized in that: The training method of the randomly connected neural network is: Step 1, randomly initializing the weights and biases of the randomly connected neural network, fixing the second parameter, and obtaining a first output through the activation layer of the single-layer feedforward neural network, wherein the second parameter is the weight and bias between the input layer and the second linear layer; Step 2, based on a linear regression formula, solving a first parameter and a third parameter from the first output, wherein the first parameter is a weight between the input layer and the output layer, and the third parameter is a weight between the second linear layer and the output layer; Step 3: Use the first parameter, the second parameter and the third parameter as network parameters of the randomly connected neural network to obtain a trained radiation source recognition model.

7. The radiation source identification model training method based on a randomly connected network according to any one of claims 1 to 6, characterized in that: The training method further comprises: Generate different combinations of hyperparameters, and obtain different radiation source recognition models based on the different combinations of hyperparameters training; The classification accuracy of the radiation source identification model is evaluated through a test set, and the optimal model is selected for radiation source identification.

8. A radiation source identification method, characterized in that: include: Acquiring detected electromagnetic signals; Extracting a characteristic vector from the electromagnetic signal, and performing normalization and quantization processing on the characteristic vector; Density encoding is performed on the normalized and quantized feature vector to obtain a bit vector; Performing hyperdimensional encoding on the bit vector to obtain a hyperdimensional vector feature; The hyperdimensional vector feature is input into a radiation source identification model to obtain the type of the electromagnetic signal; the radiation source identification model is trained by the radiation source identification model training method based on a random connection network described in any one of claims 1-7.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the radiation source identification model training method based on a randomly connected network described in any one of claims 1 to 7, or the radiation source identification method described in claim 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the radiation source identification model training method based on a randomly connected network described in any one of claims 1 to 7 or the radiation source identification method described in claim 8 is implemented.

Images