Rail transit-oriented channel reorganization security key extraction method and system

Abstract

The application discloses a channel recombination security key extraction method and system for rail transit, and relates to the field of digital information transmission. In the method, a sending end constructs a full connection neural network by taking acquired uplink and downlink channel data as original data. The sending end and a legal receiving end send pilot signals to each other to obtain a first channel estimation sequence of the original channel corresponding to the sending end and a second channel estimation sequence of the original channel corresponding to the legal receiving end. The sending end inputs the first channel estimation sequence into the full connection neural network to obtain a mapping channel estimation sequence. The sending end generates a channel recombination table based on a preset initial key and the mapping channel estimation sequence. After receiving the channel recombination table, the legal receiving end replaces the second channel estimation sequence according to elements in the channel recombination table to obtain a replaced channel estimation sequence. The legal receiving end decodes the replaced channel estimation sequence to obtain a security key. The application is used for improving the security of a rail transit communication system.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of digital information transmission, and particularly relates to a channel reorganization security key extraction method and system for rail transit. BACKGROUND

[0002] In the field of rail transit, in order to improve the real-time performance and reliability of the communication system, a cellular mobile communication architecture is usually adopted, and mainstream schemes mainly include time division duplex (TDD) and frequency division duplex (FDD) transmission modes. Among them, the FDD mechanism configures different frequency resources for the uplink and downlink respectively, realizes the synchronous transmission and reception of data, and thus suppresses the mutual interference problem between the uplink and downlink. However, the rail transit environment itself has high-speed movement, high reliability requirements and variable and complex wireless channel characteristics, which poses more stringent challenges to the security protection of the communication system.

[0003] In the related art, a physical layer key extraction scheme is usually adopted, which realizes the generation of a security key by quantizing a channel characteristic sequence and combining key negotiation and privacy amplification steps. However, due to the influence of noise in the channel detection process, when the observation values of the two parties fall on the two sides of the quantization threshold, the key cannot be used or an error occurs, so that the quantization generated key often has a certain degree of inconsistency, and the system usually needs to perform multiple key negotiations to eliminate the inconsistency, which not only increases the communication overhead, but also may lead to the leakage of key information, thereby reducing the security of the rail transit communication system. SUMMARY

[0004] The application provides a channel reorganization security key extraction method and system for rail transit, which is used to improve the security of the rail transit communication system.

[0005] In a first aspect, the application provides a channel reorganization security key extraction method for rail transit, wherein a sending end constructs a fully connected neural network composed of an input layer, four hidden layers and an output layer by taking the collected uplink and downlink channel data as original data.

[0006] The sending end and a legitimate receiving end send pilot signals to each other to obtain a first channel estimation sequence of the original channel corresponding to the sending end and a second channel estimation sequence of the original channel corresponding to the legitimate receiving end.

[0007] The sending end inputs the first channel estimation sequence into the fully connected neural network to obtain a mapped channel estimation sequence.

[0008] The sending end generates a channel reorganization table based on a preset initial key and the mapped channel estimation sequence.

[0009] After receiving the channel reassembly table, the legitimate receiving end permutes the second channel estimation sequence according to the elements in the channel reassembly table to obtain a permuted channel estimation sequence;

[0010] The legitimate receiving end decodes the permuted channel estimation sequence to obtain the security key.

[0011] By adopting the above technical solution and using a fully connected neural network to process the original channel data, the mapping relationship between the uplink and downlink channels can be fully learned, so that the sending end can more accurately predict the channel state of the legitimate receiving end. The method of generating a channel reorganization table based on a preset initial key and a mapped channel estimation sequence introduces randomness while ensuring channel reciprocity, increasing the uncertainty of the key generation process. The legitimate receiving end performs a permutation operation on the second channel estimation sequence according to the channel reorganization table, so that the final generated security key maintains the channel characteristics and has strong randomness. This key generation method based on neural networks and channel reorganization improves the randomness and security of the key. At the same time, due to the use of channel reciprocity, it ensures that the sending end and the receiving end can generate consistent keys, improves the reliability of key generation, and thus improves the security of the rail transit communication system.

[0012] In combination with some embodiments of the first aspect, in some embodiments, the fully connected neural network is a neural network with a fully connected structure including an input layer, four hidden layers and an output layer, the input layer receives a first channel estimation sequence as an input feature and the number of neurons is the same as the dimension of the first channel estimation sequence, the four hidden layers include a first preset number, a second preset number, a third preset number and a fourth preset number of neurons, respectively, the four hidden layers perform inter-layer nonlinear transformation through the ReLU activation function and a Dropout layer with a preset second ratio is set after each hidden layer, the output layer generates a mapping channel estimation sequence with the same dimension as the second channel estimation sequence, the transmitting end uses the Adam optimization algorithm to train the fully connected neural network based on the uplink and downlink channel data as a training set and a test set, and uses a loss function to minimize the difference between the mapping channel estimation sequence output by the network and the second channel estimation sequence, and the sizes of the first preset number, the second preset number, the third preset number and the fourth preset number decrease in sequence.

[0013] By adopting the above technical solution, a four-layer hidden layer structure is constructed with a decreasing number of neurons, and nonlinear transformations are performed with the ReLU activation function, enabling the network to extract abstract representations of channel features layer by layer. Placing a Dropout layer after each hidden layer prevents network overfitting and improves the model's generalization capabilities. The network is trained using the Adam optimization algorithm, which adaptively adjusts the learning rate and accelerates network convergence. By minimizing the difference between the mapped channel estimate sequence and the second channel estimate sequence as the training objective, the network learns the mapping relationship between the uplink and downlink channels. This network structure design enables the model to have strong feature extraction and generalization capabilities, improves the accuracy of channel mapping, and thus enhances the quality of the generated key.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, the loss function is: ;

[0015] In the above function, is the mapping channel estimation sequence, is the second channel estimation sequence, is the number of training samples, is the independent variable, For the The mapping channel estimation sequence corresponding to the training samples, For the A corresponding second channel estimation sequence.

[0016] By adopting this technical solution, the normalized mean squared error (NMSE) is used as a loss function. The accuracy of the network output is measured by calculating the mean of the sum of squared errors between the mapped channel estimate sequence and the second channel estimate sequence. Compared to directly using the mean squared error (MSE), the NMSE takes into account the relative differences in channel estimates and can better handle channel data of different scales. By minimizing this loss function, the network can learn more accurate channel mapping relationships and reduce prediction errors.

[0017] In conjunction with some embodiments of the first aspect, in some embodiments, the transmitting end generates a channel reassembly table based on a preset initial key and a mapped channel estimation sequence, specifically including:

[0018] The sender encodes the preset initial key to obtain a codeword sequence. The preset initial key is generated by a random source that obeys a 0 / 1 uniform distribution.

[0019] The transmitting end sorts the mapping channel estimation sequence by size to obtain a sorted mapping channel estimation sequence, and determines the position of each element in the sorted mapping channel estimation sequence in the mapping channel estimation sequence;

[0020] The transmitting end stores the elements with the first preset value in the codeword sequence in the order of the sorted mapping channel estimation sequence to the position, and stores the elements with the second preset value in the codeword sequence in the reverse order of the sorted mapping channel estimation sequence to the position, to obtain a channel reorganization table.

[0021] By adopting the above technical solution, a method for generating a channel reassembly table based on a preset initial key and a mapped channel estimation sequence introduces randomness while preserving channel characteristics by sorting the mapped channel estimation sequence and storing it in either forward or reverse order based on the codeword sequence value. Because the preset initial key is generated from a random source with a uniform 0 / 1 distribution, the reassembly process is unpredictable. This channel reassembly method preserves channel reciprocity while adding additional randomness, resulting in a more random and unpredictable key. Furthermore, the reversible reassembly process ensures that the sender and receiver can generate the same key.

[0022] In conjunction with some embodiments of the first aspect, in some embodiments, after decoding the permuted channel estimation sequence at the legitimate receiving end to obtain the security key, the method further includes:

[0023] The transmitting end obtains a first Doppler frequency shift value, and the legal receiving end obtains a second Doppler frequency shift value;

[0024] When the difference between the first Doppler frequency shift value and the second Doppler frequency shift value is greater than a first preset threshold, it is determined that a train passing scene has been entered;

[0025] The transmitting end inputs the first Doppler frequency shift value into a pre-trained multi-scale convolutional neural network to obtain an output result;

[0026] The sending end determines the number of spare neurons that need to be temporarily activated in each hidden layer based on the output results;

[0027] The transmitter initializes the weights of the number of spare neurons according to the channel characteristics in the train passing scenario to obtain a corrected fully connected neural network;

[0028] After the fully connected neural network is replaced with the corrected fully connected neural network, the transmitting end performs the step of inputting the first channel estimation sequence into the fully connected neural network.

[0029] By adopting the above technical solution, the difference in Doppler frequency shift values ​​is detected to determine the train passing scenario, and a multi-scale convolutional neural network is used to analyze the first Doppler frequency shift value to obtain an output result. Based on the output result, the number of spare neurons that need to be temporarily activated is determined. The weights of these spare neurons are then initialized according to the channel characteristics in the train passing scenario, and finally a corrected fully connected neural network is obtained. In the train passing scenario, due to the high-speed movement of the two trains towards each other, the channel characteristics will change significantly, and conventional neural networks may not be able to accurately map the channel characteristics. The corrected fully connected neural network dynamically adjusts the network structure and adds highly targeted spare neurons. The weights of these neurons are specifically optimized based on the channel characteristics of the train passing scenario, allowing the network to better adapt to the drastic changes in channel characteristics. The corrected neural network can maintain a high channel mapping accuracy in the train passing scenario, improve the matching degree between the generated mapped channel estimation sequence and the actual channel characteristics, and thus improve the reliability of the final generated security key.

[0030] In conjunction with some embodiments of the first aspect, in some embodiments, the sending end determines the number of spare neurons that need to be temporarily activated in each hidden layer based on the output result, specifically including:

[0031] The sending end divides the output of each hidden layer of the fully connected neural network into multiple subsequences according to a preset number;

[0032] The sending end performs Fourier transform on each subsequence to obtain the frequency domain feature sequence;

[0033] The transmitter determines the number of spare neurons that need to be temporarily activated in each hidden layer based on the proportion of high-frequency components in the frequency domain feature sequence. The proportion of high-frequency components is proportional to the number of spare neurons by a preset coefficient.

[0034] By adopting the above technical solution, the output results of the multi-scale convolutional neural network are divided into multiple subsequences according to a preset number, and a Fourier transform is performed on each subsequence to obtain a frequency domain feature sequence. The number of spare neurons that need to be temporarily activated in each hidden layer is determined based on the proportion of high-frequency components in the frequency domain feature sequence. By performing segmented frequency domain analysis on the output results, the changing pattern of the channel characteristics can be more carefully portrayed, and the proportion of high-frequency components reflects the severity of the change in channel characteristics. A proportional relationship is established between the proportion of high-frequency components and the number of spare neurons, so that more spare neurons are activated in scenarios where the channel characteristics change more drastically, making the corrected neural network have stronger environmental adaptability and improving the accuracy of the neural network's mapping of channel characteristics.

[0035] In conjunction with some embodiments of the first aspect, in some embodiments, after decoding the permuted channel estimation sequence at the legitimate receiving end to obtain the security key, the method further includes:

[0036] The transmitting end obtains the first angular velocity and the first position information based on the inertial navigation unit, and the legal receiving end obtains the second angular velocity and the second position information based on the inertial navigation unit;

[0037] When the first angular velocity is greater than a second preset threshold and the distance between the first position information and the second position information is less than a third preset threshold, it is determined that the vehicle enters a curve scene;

[0038] The transmitting end fuses the first angular velocity and the first position information to generate a curve feature vector;

[0039] The sending end inputs the curve feature vector into the pre-trained generative adversarial network to generate a new neuron connection method;

[0040] The sending end reconstructs the topology of the fully connected neural network according to the new neuron connection mode to obtain a reconstructed fully connected neural network;

[0041] After the fully connected neural network is replaced with the reconstructed fully connected neural network, the transmitting end performs the step of inputting the first channel estimation sequence into the fully connected neural network.

[0042] By adopting the above technical solution, the angular velocity and position information are detected to determine the curve scene, and the angular velocity and position information are fused to generate a curve feature vector. The pre-trained generative adversarial network is used to generate a new neuron connection method based on the curve feature vector, and the topology of the fully connected neural network is reconstructed accordingly. In the curve scene, due to the multipath effect and scattering effect caused by the train's turning movement, the channel characteristics will show a unique change pattern. The generative adversarial network obtains the mapping relationship between the channel characteristics and the optimal network structure in the curve scene through adversarial training, and can generate a matching neuron connection method based on the specific curve characteristics. This method of dynamically reconstructing the network topology enables the structure of the neural network to be optimized and adjusted according to the characteristics of the curve scene, avoiding the performance degradation problem that may occur in the fixed network structure in the curve scene, and improving the channel feature mapping accuracy of the neural network in the curve scene.

[0043] In the second aspect, an embodiment of the present application provides a channel reorganization security key extraction system for rail transit, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and one or more processors call the computer instructions to enable the system to execute the method described in the first aspect and any possible implementation method of the first aspect.

[0044] In a third aspect, an embodiment of the present application provides a computer-readable storage medium comprising instructions, which, when executed on a system, enables the system to execute the method described in the first aspect and any possible implementation of the first aspect.

[0045] In a fourth aspect, an embodiment of the present application provides a computer program product, which, when executed on a system, enables the system to execute the method described in any possible implementation manner in the first aspect.

[0046] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0047] 1. The present application provides a channel reorganization security key extraction method for rail transit, which uses a fully connected neural network to process the original channel data, and can fully learn the mapping relationship between the uplink and downlink channels, so that the sending end can more accurately predict the channel state of the legitimate receiving end. The method of generating a channel reorganization table based on a preset initial key and a mapped channel estimation sequence introduces randomness while ensuring channel reciprocity, increasing the uncertainty of the key generation process. The legitimate receiving end performs a permutation operation on the second channel estimation sequence according to the channel reorganization table, so that the final generated security key maintains the channel characteristics and has strong randomness. This key generation method based on neural networks and channel reorganization improves the randomness and security of the key. At the same time, due to the use of channel reciprocity, it ensures that the sending end and the receiving end can generate consistent keys, improves the reliability of key generation, and thus improves the security of the rail transit communication system.

[0048] 2. The present application provides a channel reorganization security key extraction method for rail transit, which detects the difference in Doppler frequency shift values ​​to determine the train passing scenario, and uses a multi-scale convolutional neural network to analyze the first Doppler frequency shift value to obtain an output result. Based on the output result, the number of spare neurons that need to be temporarily activated is determined, and then the weights of these spare neurons are initialized according to the channel characteristics in the train passing scenario, and finally a corrected fully connected neural network is obtained. In the train passing scenario, due to the high-speed opposite movement of the two trains, the channel characteristics will change significantly, and conventional neural networks may not be able to accurately map the channel characteristics. The corrected fully connected neural network adds highly targeted spare neurons by dynamically adjusting the network structure. The weights of these neurons are specially optimized based on the channel characteristics of the train passing scenario, so that the network can better adapt to the drastic changes in channel characteristics. The corrected neural network can maintain a high channel mapping accuracy in the train passing scenario, improve the matching degree between the generated mapping channel estimation sequence and the actual channel characteristics, and thus improve the reliability of the security key finally generated.

[0049] 3. The present application provides a channel reconstruction security key extraction method for rail transit, which detects angular velocity and position information to determine the curve scene, and fuses the angular velocity and position information to generate a curve feature vector, and uses a pre-trained generative adversarial network to generate a new neuron connection method based on the curve feature vector, and reconstructs the topology of the fully connected neural network accordingly. In the curve scene, due to the multipath effect and scattering effect caused by the turning movement of the train, the channel characteristics will show a unique change pattern. The generative adversarial network obtains the mapping relationship between the channel characteristics and the optimal network structure in the curve scene through adversarial training, and can generate a matching neuron connection method based on the specific curve characteristics. This method of dynamically reconstructing the network topology enables the structure of the neural network to be optimized and adjusted according to the characteristics of the curve scene, avoiding the performance degradation problem that may occur in the fixed network structure in the curve scene, and improving the channel feature mapping accuracy of the neural network in the curve scene. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of a channel reorganization security key extraction method for rail transit in an embodiment of the present application.

[0051] Figure 2 This is another flow chart of a channel reorganization security key extraction method for rail transit in an embodiment of the present application.

[0052] Figure 3 This is a schematic diagram of the physical device structure of a channel reorganization security key extraction system for rail transit provided in an embodiment of the present application. DETAILED DESCRIPTION

[0053] The terms used in the following examples of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification and appended claims of this application, the singular expressions "a," "an," "said," "above," "the," and "this" are intended to include plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in this application refers to any or all possible combinations comprising one or more of the listed items.

[0054] In the following, the terms "first" and "second" are used for descriptive purposes only and should not be understood to imply or suggest relative importance or implicitly indicate the number of the technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of this application, unless otherwise specified, "plurality" means two or more.

[0055] The following uses an embodiment and combines Figure 1 , a channel reorganization security key extraction method for rail transit in an embodiment of the present application is described:

[0056] See also Figure 1 , which is a flow chart of a channel reorganization security key extraction method for rail transit in an embodiment of the present application.

[0057] S101, the transmitter uses the collected uplink and downlink channel data as raw data to construct a fully connected neural network consisting of an input layer, four hidden layers and an output layer;

[0058] The transmitting end uses the collected uplink and downlink channel data as raw data to construct a fully connected neural network consisting of an input layer, four hidden layers, and an output layer. The system first collects the uplink and downlink channel data as raw data. The system then uses this raw data to construct a fully connected neural network consisting of an input layer, four hidden layers, and an output layer. A fully connected neural network is a common neural network architecture in which neurons in each layer are connected to all neurons in the previous layer. The input layer receives the first channel estimate sequence as input features and has the same number of neurons as the first channel estimate sequence. The four hidden layers include, in descending order, a first preset number, a second preset number, a third preset number, and a fourth preset number of neurons. Nonlinear transformations are performed between hidden layers using the ReLU activation function, and a dropout layer with a preset second ratio is placed after each hidden layer to prevent overfitting. The output layer generates a mapped channel estimate sequence with the same dimensions as the second channel estimate sequence. When constructing the neural network, the number of hidden layers and the number of neurons in each layer can be adjusted according to specific needs and are not limited here.

[0059] The system can randomly divide the raw data into a training set and a test set. The training set data is then used to train the neural network, and the test set data is used to test the trained neural network to evaluate its performance. During the training process, the system can use the Adam optimization algorithm to update the parameters of the neural network and use the mean square error as the loss function to minimize the difference between the mapped channel estimation sequence output by the network and the actual second channel estimation sequence.

[0060] The system can also use cross-validation to train and evaluate neural networks. Specifically, the system divides the original data into k parts (usually k = 5 or 10), then performs k training and testing cycles. Each time, one part of the data is selected as the test set, and the remaining k-1 parts are used as the training set. This results in k trained neural networks and k test results. Finally, the system averages these k test results as an evaluation metric for the neural network's performance. This method makes better use of data and reduces the effects of chance caused by improper data partitioning.

[0061] It should be noted that the fully connected neural network is a neural network with a fully connected structure including an input layer, four hidden layers and an output layer. The input layer receives the first channel estimation sequence as an input feature and the number of neurons is the same as the dimension of the first channel estimation sequence. The four hidden layers include a first preset number, a second preset number, a third preset number and a fourth preset number of neurons, respectively. The four hidden layers perform inter-layer nonlinear transformation through the ReLU activation function and a Dropout layer with a preset second ratio is set after each hidden layer. The output layer generates a mapping channel estimation sequence with the same dimension as the second channel estimation sequence. The transmitting end uses the Adam optimization algorithm to train the fully connected neural network based on the uplink and downlink channel data as a training set and a test set, and uses a loss function to minimize the difference between the mapping channel estimation sequence output by the network and the second channel estimation sequence. The sizes of the first preset number, the second preset number, the third preset number and the fourth preset number decrease in sequence.

[0062] Among them, the loss function is: ;

[0063] In the above function, is the mapping channel estimation sequence, is the second channel estimation sequence, is the number of training samples, is the independent variable, For the The mapping channel estimation sequence corresponding to the training samples, For the A corresponding second channel estimation sequence.

[0064] The physical meaning of this loss function is to measure the difference between the channel state predicted by the neural network and the actual channel state. Mean squared error (MSE) is a commonly used loss function for regression tasks. It is sensitive to the difference between the predicted and true values ​​and can well reflect the degree of deviation between the two. During training, the system continuously adjusts the parameters of the neural network to minimize the loss function value, that is, to ensure that the mapped channel estimate sequence output by the network is as close as possible to the actual second channel estimate sequence.

[0065] S102, the sending end and the legitimate receiving end send pilot signals to each other to obtain a first channel estimation sequence of the original channel corresponding to the sending end and a second channel estimation sequence of the original channel corresponding to the legitimate receiving end;

[0066] In this step, the sending end and the legitimate receiving end in the system estimate the state of the channel by exchanging pilot signals. The pilot signal is a pre-designed known signal, usually a pseudo-random sequence or an orthogonal sequence, which has good correlation and channel estimation performance. The sending end and the legitimate receiving end alternately send pilot signals within the coherence time of the channel, and estimate the channel impulse response or channel gain according to the received pilot signals.

[0067] Specifically, the sending end sends a known pilot signal sequence before sending data. After receiving the pilot signal, the legitimate receiving end compares it with the locally stored pilot signal copy, estimates the second channel estimation sequence of the original channel through correlation operation or least mean square error estimation method, etc. Similarly, the legitimate receiving end also sends a pilot signal sequence to the sending end, and the sending end estimates the first channel estimation sequence of the original channel through the same method. The first channel estimation sequence and the second channel estimation sequence respectively reflect the channel state in the two directions from the sending end to the legitimate receiving end and from the legitimate receiving end to the sending end.

[0068] In actual systems, the design of the pilot signal needs to consider many factors, such as the length, power, transmission time and frequency of the pilot signal. The length of the pilot signal needs to be long enough to provide sufficient channel estimation accuracy; at the same time, it cannot be too long to occupy too many channel resources. The power of the pilot signal needs to be moderate, which needs to ensure sufficient signal-to-noise ratio to perform reliable channel estimation, and also needs to control the interference of the pilot signal to the data signal. The transmission time and frequency of the pilot signal need to be set reasonably according to the coherence time and coherence bandwidth of the channel to track the changes of the channel state.

[0069] S103, the sending end inputs the first channel estimation sequence into the fully connected neural network to obtain a mapped channel estimation sequence;

[0070] In this step, the sending end of the system inputs the first channel estimation sequence into the fully connected neural network, and after forward propagation calculation, obtains the result of the output layer, i.e. the mapped channel estimation sequence.

[0071] Specifically, the sending end first encodes the first channel estimation sequence into the input format of the neural network. Usually, the channel estimation sequence is a complex number vector, and each element represents the channel gain at different subcarriers or different time points. In order to facilitate the processing of the neural network, it is necessary to convert the complex number vector into a real number vector, which can be encoded by splicing the real part and the imaginary part or splicing the amplitude and the phase.

[0072] The transmitter then performs forward propagation calculations layer by layer, following the neural network structure. The input layer receives the encoded first channel estimate sequence and passes it to the first hidden layer. Each neuron in the hidden layer calculates a weighted sum based on the input and weights and generates an output through an activation function. The activation function introduces nonlinear transformations, enhancing the expressive power of the neural network. Common activation functions include sigmoid, tanh, and ReLU. The output of the first hidden layer is then passed to the second hidden layer, and so on, until the last hidden layer.

[0073] The output of the last hidden layer is passed to the output layer. The number of neurons in the output layer is equal to the length of the second channel estimate sequence, with each neuron corresponding to a subcarrier or a time point. The output layer typically uses a linear activation function or an identity activation function to generate an output in the same format as the second channel estimate sequence. The result of the output layer is a mapped channel estimate sequence, which represents the transmitter's prediction of the legitimate receiver's channel state.

[0074] In practical applications, regularization techniques, such as L1 / L2 regularization, dropout, and early stopping, are often used to improve the performance and generalization of neural networks. These techniques prevent neural networks from overfitting the training data and improve their prediction performance on new data. Furthermore, optimization techniques, such as batch normalization, gradient clipping, and learning rate scheduling, can be used to accelerate neural network training and improve training efficiency.

[0075] S104, the transmitting end generates a channel reassembly table based on the preset initial key and the mapped channel estimation sequence;

[0076] The transmitting end generates a channel reorganization table based on a preset initial key and a mapping channel estimation sequence. Specifically, the transmitting end encodes the preset initial key to obtain a codeword sequence, where the preset initial key is generated by a random source that obeys a 0 / 1 uniform distribution; the transmitting end sorts the mapping channel estimation sequence by size to obtain a sorted mapping channel estimation sequence, and determines the position of each element in the sorted mapping channel estimation sequence in the mapping channel estimation sequence; the transmitting end stores the elements with a first preset value in the codeword sequence in the order of the sorted mapping channel estimation sequence in the position, and stores the elements with a second preset value in the codeword sequence in the reverse order of the sorted mapping channel estimation sequence in the position, to obtain the channel reorganization table.

[0077] In this step, the system's transmitter first encodes a preset initial key to obtain a codeword sequence. This preset initial key can be generated from a random source with a uniform 0 / 1 distribution. The transmitter then sorts the mapped channel estimation sequence obtained in step S103 by size to obtain a sorted mapped channel estimation sequence and determines the position of each element in the sorted mapped channel estimation sequence within the original mapped channel estimation sequence. Next, the transmitter generates a channel reassembly table based on the codeword sequence and the sorted mapped channel estimation sequence. Specifically, the transmitter stores elements in the codeword sequence with a first preset value (e.g., 0) in the order of the sorted mapped channel estimation sequence, and stores elements in the codeword sequence with a second preset value (e.g., 1) in the reverse order of the sorted mapped channel estimation sequence. The generated channel reassembly table contains information about the preset initial key and the mapped channel estimation sequence. When generating the channel reassembly table, the selection and encoding method of the preset initial key can be adjusted based on actual needs and are not limited here.

[0078] The system can use some classic encoding methods, such as Manchester encoding and differential Manchester encoding, to encode the preset initial key. These encoding methods have certain error correction capabilities and can resist the influence of channel noise to a certain extent.

[0079] The system can also design special encoding methods to increase the randomness and security of the key. For example, the system can permute the preset initial key before encoding it, or perform an XOR operation on the preset initial key with some random number before encoding it. These special encoding methods can make the final key more difficult to crack.

[0080] During the channel reassembly table generation process, identical elements may appear in the sorted and mapped channel estimate sequence, resulting in non-unique mapping relationships within the reassembly table. To address this issue, the system randomizes the positions of identical elements. Specifically, when storing these identical elements, their positions within the channel reassembly table are randomly selected. This ensures that the channel reassembly tables generated by the sender and receiver are consistent, while also introducing a degree of randomness and improving key security.

[0081] S105. After receiving the channel reassembly table, the legitimate receiving end permutes the second channel estimation sequence according to elements in the channel reassembly table to obtain a permuted channel estimation sequence.

[0082] In this step, after receiving the channel reassembly table from the transmitter, the authorized receiver in the system uses the information in the channel reassembly table to permute the second channel estimate sequence to generate a permuted channel estimate sequence. This process can be considered an encryption operation, shuffling and reordering the second channel estimate sequence to match the sorted mapped channel estimate sequence generated by the transmitter.

[0083] Specifically, the channel reorganization table is actually a position mapping table that records the position of each element in the sorted mapped channel estimate sequence within the original mapped channel estimate sequence. The legitimate receiver uses this position mapping to rearrange the elements in the second channel estimate sequence to obtain a permuted channel estimate sequence. For example, if the second channel estimate sequence is [h1, h2, h3, h4] and the position mapping recorded in the channel reorganization table is [3, 1, 4, 2], then the permuted channel estimate sequence is [h2, h4, h1, h3].

[0084] It's important to note that the permutation operation here is a symmetric encryption operation, and the permutation order is completely determined by the channel reassembly table. This table is generated by the transmitter based on a preset initial key and a sorted and mapped channel estimation sequence. A legitimate receiver can only obtain the same channel reassembly table, and thus the same permuted channel estimation sequence, if it possesses the same preset initial key and sorted and mapped channel estimation sequence. This ensures the security of the permutation operation and prevents unauthorized receivers from obtaining communication content.

[0085] Furthermore, since the channel estimation sequence is obtained within the channel coherence time, the permutation operation must be completed within the coherence time. Once the channel state changes, the channel estimation sequence also changes, and the previous channel reassembly table is no longer applicable, requiring the channel estimation and key generation process to be repeated. This also improves communication security from another perspective.

[0086] S106: The legal receiving end decodes the permuted channel estimation sequence to obtain a security key.

[0087] In this step, the legitimate receiver of the system processes the permuted channel estimate sequence, extracts the key information contained therein, and generates the final security key. This process can be considered a decryption operation, which converts the encrypted permuted channel estimate sequence back into the key.

[0088] The legitimate receiver uses a key extraction algorithm agreed upon with the transmitter to analyze and calculate the permuted channel estimate sequence. This algorithm must fully exploit the randomness and uniqueness of the permuted channel estimate sequence while ensuring the security and consistency of the key. Common key extraction algorithms include quantization, error correction, masking, and hashing.

[0089] Taking the quantization algorithm as an example, a legitimate receiver can quantize the amplitude or phase of the permuted channel estimate sequence to obtain a bit sequence. This bit sequence can then be used as a security key. The quantization threshold and interval must be appropriately set based on the distribution characteristics of the permuted channel estimate sequence to balance the randomness and consistency of the key.

[0090] If an error correction algorithm is used, the legitimate receiver can treat the permuted channel estimate sequence as a key sequence transmitted through a random error channel and decode it using an error-correcting code (such as BCH or LDPC) to obtain the original key sequence. This effectively improves key consistency and reduces the impact of channel errors.

[0091] The masking algorithm uses a pre-shared random sequence (i.e., mask) to perform an XOR operation on the permuted channel estimate sequence to obtain the masked key sequence. This method can further enhance the security of the key and prevent it from being intercepted and cracked.

[0092] A hash algorithm uses a hash function to map a permuted channel estimate sequence to a fixed-length hash value, which serves as a secure key. Hash functions are one-way and collision-resistant, meaning it's difficult to decipher the original input from the hash value, and different inputs rarely produce the same hash value. Therefore, the keys generated by hash algorithms are highly random and unique.

[0093] In the above embodiment, a fully connected neural network is used to process the original channel data, which can fully learn the mapping relationship between the uplink and downlink channels, so that the sending end can more accurately predict the channel state of the legitimate receiving end. The method of generating a channel reorganization table based on a preset initial key and a mapped channel estimation sequence introduces randomness while ensuring channel reciprocity, increasing the uncertainty of the key generation process. The legitimate receiving end performs a permutation operation on the second channel estimation sequence according to the channel reorganization table, so that the final generated security key maintains the channel characteristics and has strong randomness. This key generation method based on neural networks and channel reorganization improves the randomness and security of the key. At the same time, due to the use of channel reciprocity, it ensures that the sending end and the receiving end can generate consistent keys, improves the reliability of key generation, and thus improves the security of the rail transit communication system.

[0094] The above embodiment describes a basic channel reorganization key extraction method, which uses a fully connected neural network to achieve channel feature mapping and completes key generation in combination with a channel reorganization table. However, in actual rail transit scenarios, the train operating environment is very complex, especially in special scenarios such as train intersections and curves, where channel features can change dramatically. In order to make the system better adapt to these special scenarios, the following is combined with Figure 2, another channel reorganization security key extraction method for rail transit in an embodiment of the present application is described:

[0095] See also Figure 2 , which is another flow chart of a channel reorganization security key extraction method for rail transit in an embodiment of the present application.

[0096] S201: A transmitting end obtains a first Doppler frequency shift value, and a legal receiving end obtains a second Doppler frequency shift value;

[0097] In this step, the transmitter and legitimate receiver in the system each obtain their respective measured Doppler shift values. Doppler shift is a signal frequency deviation caused by relative motion between the transmitter and receiver, reflecting the train's motion and speed. In rail transit scenarios, the impact of Doppler shift is significant due to the high speed of trains and needs to be considered.

[0098] The transmitter and authorized receiver can use pilot signals to estimate Doppler shift. The transmitter sends a known pilot signal to the authorized receiver, which analyzes the frequency offset of the received pilot signal to estimate the second Doppler shift value. Similarly, the authorized receiver sends a pilot signal to the transmitter, which estimates the first Doppler shift value. The pilot signal frequency can be selected to balance estimation accuracy and frequency resource utilization.

[0099] S202: When the difference between the first Doppler frequency shift value and the second Doppler frequency shift value is greater than a first preset threshold, it is determined that a train passing scene has been entered;

[0100] In this step, the system compares the difference between the Doppler shift values ​​obtained by the transmitter and the authorized receiver with a preset threshold to determine whether a train has entered a passing scenario. In a passing scenario, there is significant relative motion between the two trains traveling in opposite directions, and the Doppler shift values ​​measured by each train typically differ significantly. Therefore, by setting an appropriate threshold, the difference in Doppler shift values ​​can be used to detect a passing train scenario.

[0101] Specifically, the system first calculates the absolute value of the difference between the first and second Doppler shift values, then compares this value to a first preset threshold. If the absolute value of the difference exceeds the first threshold, it is determined that a train passing scenario has occurred; otherwise, the train is considered to be in normal operation. The selection of the first threshold requires a comprehensive consideration of factors such as the train's operating speed, track layout, and communication system parameters to balance detection sensitivity and false alarm rate.

[0102] In practical applications, the first preset threshold can be set in the following two ways:

[0103] Method 1: Based on prior knowledge of train speeds and track layout, the theoretical Doppler shift difference range for a passing train scenario is pre-calculated and used as the first preset threshold. This method is simple and intuitive, eliminating the need for extensive field testing and data analysis. However, its disadvantage is that the threshold is fixed and cannot adapt to changes in train operating conditions and the environment.

[0104] Method 2: Statistically analyze a large amount of field test data to obtain an empirical distribution of Doppler shift differences in train passing scenarios. Based on the requirements for false alarm rate and missed detection rate, an appropriate quantile from this empirical distribution is selected as the first preset threshold. The advantage of this method is that the threshold can be adjusted according to actual needs. However, the disadvantage is that it requires extensive field testing and data analysis, which is a heavy workload.

[0105] During the train passing detection process, false alarms or missed detections may occur. A false alarm refers to the misjudgment of a normal operating state as a train passing scenario, and a missed detection refers to the misjudgment of a train passing scenario as a normal operating state. To reduce the false alarm and missed detection rates, the system can introduce a multi-level decision mechanism. Specifically, the system not only examines the Doppler frequency shift difference at the current moment, but also examines the historical difference within a certain time window. Only when the difference at multiple consecutive moments exceeds the first preset threshold is it determined to have entered the train passing scenario. This can effectively avoid false alarms caused by transient interference or measurement errors. At the same time, the system can also set a second preset threshold. When the difference at multiple consecutive moments is lower than the second preset threshold, it is determined to have exited the train passing scenario. The second preset threshold should be lower than the first preset threshold. The two thresholds form a hysteresis interval, which can avoid frequent entry and exit of the train passing scenario under critical conditions.

[0106] S203: The transmitting end inputs the first Doppler frequency shift value into a pre-trained multi-scale convolutional neural network to obtain an output result;

[0107] In this step, the system's transmitter processes the first Doppler shift value using a pre-trained multi-scale convolutional neural network (CNN) to generate the corresponding output. A CNN is a specialized CNN structure that uses convolution kernels of varying scales to extract multi-scale features from the input data, enhancing the network's expressiveness and robustness. In train-passing scenarios, the channel environment varies dramatically, with multi-scale scattering and fading effects present. Using a CNN can better characterize these complex channel characteristics.

[0108] Specifically, the transmitter inputs a sequence of first Doppler shift values ​​into a multi-scale convolutional neural network. The network's input layer receives the Doppler shift sequence, which is then passed through multiple convolutional and pooling layers to extract multi-scale features of the Doppler shift. Each convolutional layer consists of multiple convolution kernels, whose sizes can be set to different values ​​to capture frequency shift features at different scales. The pooling layer is used to reduce the size of the feature map and minimize computational effort. Convolutional and pooling layers are stacked alternately to gradually form a more abstract and high-level feature representation. Finally, the outputs of the convolutional and pooling layers are combined through a fully connected layer to generate the final output.

[0109] In practical systems, the structure and parameters of multi-scale convolutional neural networks can be designed and optimized according to specific needs. The following are two possible design schemes:

[0110] Solution 1: Use the classic VGG network architecture, consisting of multiple small 3x3 convolution kernels and max pooling layers. For example, you can set up five convolution layers with 32, 64, 128, 256, and 512 kernels, respectively. Max pooling layers are inserted between the convolution layers, with a pooling window size of 2x2 and a stride of 2. This architecture has the advantages of simplicity and efficiency, and has been widely verified in fields such as image recognition. However, its disadvantage is a smaller receptive field, which may not effectively capture large-scale frequency shift features.

[0111] Solution 2: Use techniques such as dilated convolution or pyramid pooling to expand the receptive field of the convolution kernel. For example, a dilated convolution layer can be inserted between standard convolution layers, with the dilation rate increasing exponentially, such as 1, 2, 4, 8, and so on. This allows for a larger receptive field without increasing the number of parameters. Alternatively, a pyramid pooling module can be added at the end of the network to adaptively global pool feature maps of different scales, thereby capturing global contextual information. The advantage of this structure is a large receptive field and the ability to extract multi-scale features; the disadvantage is a high computational load and a relatively complex network structure.

[0112] S204: The sending end determines the number of spare neurons that need to be temporarily activated in each hidden layer according to the output result;

[0113] The sending end determines the number of spare neurons that need to be temporarily activated in each hidden layer based on the output results, specifically including: the sending end divides the output results of each hidden layer of the fully connected neural network into multiple subsequences according to a preset number; the sending end performs Fourier transform on each subsequence to obtain a frequency domain feature sequence; the sending end determines the number of spare neurons that need to be temporarily activated in each hidden layer based on the proportion of high-frequency components in the frequency domain feature sequence, and the proportion of high-frequency components is proportional to the number of spare neurons by a preset coefficient.

[0114] In this step, the system's transmitter determines the number of backup neurons that need to be temporarily activated in each hidden layer of the subsequent fully connected neural network based on the output of the multi-scale convolutional neural network. The fully connected neural network can flexibly adjust the number of hidden layer neurons based on changes in the channel environment to adapt to feature mapping relationships in different scenarios. Backup neurons are neurons that are dormant under normal circumstances and do not participate in calculations. They are activated only in special scenarios. By analyzing the output of the multi-scale convolutional neural network, the transmitter can estimate the complexity of the current channel environment and determine the number of backup neurons to activate accordingly.

[0115] Specifically, the transmitter uses the segmented Fourier transform method to perform time-frequency domain analysis on the feature sequence output by the multi-scale convolutional neural network. First, the transmitter divides the output feature sequence into multiple subsequences, each of which contains a preset number of feature values. Then, the transmitter performs a fast Fourier transform (FFT) on each subsequence to obtain the corresponding frequency domain feature sequence. The frequency domain feature sequence reflects the distribution of the original feature sequence on different frequency components. Next, the transmitter counts the proportion of high-frequency components in each frequency domain feature sequence as an indicator of the complexity of the channel environment. Finally, the transmitter determines the number of spare neurons that need to be temporarily activated in each hidden layer based on the proportional relationship between the proportion of high-frequency components and the number of spare neurons. The higher the proportion of high-frequency components, the more complex the channel environment, and the more spare neurons that need to be activated.

[0116] In actual systems, the following two methods can be used to establish the mapping relationship between the proportion of high-frequency components and the number of spare neurons:

[0117] Method 1: Based on an empirical formula. By statistically analyzing a large amount of field test data, an empirical formula is developed to correlate the proportion of high-frequency components with the number of backup neurons. For example, a piecewise linear function can be set: when the proportion of high-frequency components is below threshold A, the number of backup neurons is 0; when the proportion of high-frequency components is between thresholds A and B, the number of backup neurons and the proportion of high-frequency components are linearly related; when the proportion of high-frequency components is above threshold B, the number of backup neurons reaches the upper limit. Thresholds A and B and the linear scaling factor can be set according to actual needs.

[0118] Method 2: Lookup table-based. A lookup table is pre-designed to store the number of backup neurons corresponding to different high-frequency component ratio intervals. For example, the high-frequency component ratio can be divided into 10 equal intervals from 0 to 1, with each interval corresponding to a specific number of backup neurons. The transmitter first determines which interval the high-frequency component ratio falls within and then directly reads the corresponding number of backup neurons from the lookup table. The granularity and value of the lookup table can be set according to actual needs.

[0119] When determining the number of backup neurons, the proportion of high-frequency components may be too high, causing the number of backup neurons to exceed the total number of neurons in the hidden layer. To avoid this, the system can set an upper limit for the number of backup neurons. When the number of backup neurons calculated based on the proportion of high-frequency components exceeds the upper limit, it is truncated to the upper limit. The upper limit should take into account factors such as neural network performance, computational complexity, and storage overhead. Generally speaking, the number of backup neurons should not exceed 20% of the total number of hidden layer neurons to avoid adversely affecting the stability and generalization ability of the neural network. In addition, the system can introduce a smoothing factor to perform a weighted average of the number of backup neurons at the current time and historical times to reduce sharp fluctuations in the number of backup neurons. This can avoid outliers caused by transient interference or measurement errors and improve system stability.

[0120] S205: The sending end initializes the weights of the number of spare neurons according to the channel characteristics in the train passing scenario to obtain a corrected fully connected neural network;

[0121] In this step, the system's transmitter uses spare neurons to modify and expand the original fully connected neural network, resulting in a corrected fully connected neural network adapted for train passing scenarios. Based on the original fully connected neural network, the transmitter adds a corresponding number of neurons to each hidden layer based on the number of spare neurons determined in step S204, and initializes their weights to values ​​appropriate for the channel characteristics of train passing scenarios. This corrected fully connected neural network is more structurally complex and parameterized, enabling it to better learn and characterize the channel feature mapping relationships in train passing scenarios.

[0122] Specifically, for each hidden layer, the transmitter adds an equal number of neurons to the existing neurons in the network. These newly added neurons use the same activation function and connection method as the original neurons, forming an expanded hidden layer. Next, the transmitter uses prior information about the train passing scenario to design a weight initialization strategy to initialize the weight matrix of the newly added neurons. This strategy needs to consider the following factors:

[0123] Statistical distribution of channel characteristics. In train-passing scenarios, the channel typically exhibits non-stationary, non-Gaussian characteristics, such as impulse noise, Doppler shift, and delay spread. These characteristics can be statistically analyzed using a large amount of field test data to derive their probability density functions or empirical distribution functions. The weight initialization strategy should closely match these statistical distributions, enabling the calibrated neural network to better adapt to the train-passing channel.

[0124] The physical mechanisms of channel characteristics. Channel characteristics in train passing scenarios are primarily caused by physical mechanisms such as relative motion between trains, vehicle occlusion, and electromagnetic scattering. By modeling and analyzing these physical mechanisms, analytical expressions or numerical relationships between channel characteristics and environmental parameters can be derived. Weight initialization strategies should fully leverage prior knowledge of these physical mechanisms to rationally set initial weight values ​​that conform to the physical laws of channel characteristics.

[0125] Convergence and generalization of neural networks. Weight initialization strategies must not only consider matching channel characteristics but also balance the convergence speed and generalization capabilities of neural network training. Generally speaking, initial weight values ​​should not be set too large or too small, as this can cause gradient explosion or vanishing, affecting network convergence. Furthermore, initial weight values ​​should exhibit a certain degree of randomness and diversity to prevent the network from falling into local optima and improve generalization. Common weight initialization methods include Gaussian distribution initialization, Xavier initialization, and Kaiming initialization.

[0126] Based on the above considerations, the following two weight initialization strategies can be adopted:

[0127] Strategy 1: Initialization based on channel statistical distribution. First, the transmitter generates a set of random samples equal to the number of spare neurons based on the empirical distribution function of channel characteristics in train passing scenarios. The transmitter then uses this set of random samples as the initial weights for the newly added neurons. This initialization strategy has the advantages of being direct and simple, and it closely matches the statistical laws of channel characteristics. However, it does not consider the physical mechanisms of channel characteristics and may not fully reflect the influence of environmental parameters.

[0128] Strategy 2: Initialization based on channel physics. First, the transmitter establishes a channel feature generation model based on the physical relationship between channel characteristics and environmental parameters in the train passing scenario. This model takes environmental parameters as input and channel characteristics as output, and can be implemented in the form of analytical expressions, lookup tables, interpolation algorithms, etc. The transmitter then uses this model to generate a set of channel feature samples equal to the number of spare neurons and uses them as the initial weight values ​​of the newly added neurons. The advantage of this initialization strategy is that it incorporates prior knowledge of the physical characteristics of the channel and can adaptively adjust the weights to adapt to changes in environmental parameters; the disadvantage is that it requires modeling the physical mechanism of the channel, which is relatively complex to implement.

[0129] In practical applications, the two weight initialization strategies described above can be combined to leverage their strengths and overcome their weaknesses. For example, an initialization strategy based on channel statistics can be used to generate initial weight values. Subsequently, an initialization strategy based on channel physics can be used to fine-tune and optimize these initial weight values. This ensures that the initial weight distribution matches the channel characteristics while also accounting for the influence of environmental parameters, ultimately improving the performance of the correction neural network.

[0130] S206: After replacing the fully connected neural network with the corrected fully connected neural network, the transmitting end performs the step of inputting the first channel estimation sequence into the fully connected neural network.

[0131] In this step, the system's transmitter replaces the original fully connected neural network with a corrected fully connected neural network and then re-executes the channel mapping process. Specifically, the transmitter inputs the first channel estimation sequence collected in the train passing scenario into the corrected fully connected neural network, and after forward propagation calculation, obtains the corresponding mapped channel estimation sequence. Because the corrected fully connected neural network is more complex in structure and more specialized in parameters, the mapped channel estimation sequence it outputs can more accurately reflect the characteristics of the train passing channel, thereby improving the reliability and security of key generation.

[0132] It's important to note that while backup neurons are introduced and the fully connected neural network is corrected for the train passing scenario, this doesn't render the original fully connected neural network completely useless. In fact, the original fully connected neural network was trained under normal conditions and is well adapted to typical channel environments. When the train exits the passing scenario and returns to normal operation, the transmitter can switch back to the original fully connected neural network, saving computational resources and storage overhead. Therefore, in a practical system, the transmitter needs to maintain two versions of the fully connected neural network: the original version, used for channel mapping in normal scenarios, and the corrected version, used for channel mapping in train passing scenarios. The transmitter can dynamically select which version of the fully connected neural network to use based on changes in the Doppler shift value, enabling adaptive allocation and optimization of computational resources.

[0133] In the above embodiment, the difference in Doppler shift values ​​is detected to determine a train passing scenario, and a multi-scale convolutional neural network is used to analyze the first Doppler shift value to obtain an output result. Based on the output result, the number of spare neurons that need to be temporarily activated is determined. The weights of these spare neurons are then initialized according to the channel characteristics in the train passing scenario, ultimately resulting in a corrected fully connected neural network. In a train passing scenario, due to the high-speed, opposite-direction movement of the two trains, the channel characteristics will change significantly, and conventional neural networks may not be able to accurately map the channel characteristics. The corrected fully connected neural network, however, dynamically adjusts the network structure to add highly targeted spare neurons. The weights of these neurons are specifically optimized based on the channel characteristics of the train passing scenario, enabling the network to better adapt to drastic changes in channel characteristics. The corrected neural network can maintain high channel mapping accuracy in train passing scenarios, improve the matching degree between the generated mapped channel estimation sequence and the actual channel characteristics, and thereby improve the reliability of the ultimately generated security key.

[0134] Furthermore, in another embodiment, after the legitimate receiving end decodes the permuted channel estimation sequence to obtain the security key, the transmitting end obtains the first angular velocity and the first position information based on the inertial navigation unit, and the legitimate receiving end obtains the second angular velocity and the second position information based on the inertial navigation unit;

[0135] When the first angular velocity is greater than a second preset threshold and the distance between the first position information and the second position information is less than a third preset threshold, it is determined that the vehicle enters a curve scene;

[0136] The transmitting end fuses the first angular velocity and the first position information to generate a curve feature vector;

[0137] The sending end inputs the curve feature vector into the pre-trained generative adversarial network to generate a new neuron connection method;

[0138] The sending end reconstructs the topology of the full connection neural network according to the new neuron connection mode, and obtains a reconstructed full connection neural network;

[0139] After the full connection neural network is replaced by the reconstructed full connection neural network, the sending end performs the step of inputting the first channel estimation sequence into the full connection neural network.

[0140] In this embodiment, the system introduces a channel feature mapping method in a curve scenario. When the train enters a curve, due to factors such as train body turning and environmental scattering, the channel feature will show a different change pattern from that in a straight line scenario. In order to adapt to this change, the system uses a dynamic network reconstruction technology based on a generative adversarial network to adjust the structure of the full connection neural network in real time according to the curve feature, so as to improve the accuracy of channel feature mapping.

[0141] Specifically, the sending end and the legitimate receiving end first obtain their own angular velocity and position information by using an inertial navigation unit. The angular velocity reflects the turning rate of the train, and the position information reflects the spatial position of the train on the track. The system judges whether the train has entered a curve scenario by judging whether the angular velocity exceeds a preset threshold and whether the distance between the sending end and the legitimate receiving end is less than a preset threshold. When both conditions are met, the system considers that the train is in a curve running state, and needs to start the channel feature mapping mechanism in the curve scenario.

[0142] After entering the curve scenario, the sending end fuses the angular velocity and position information to generate a vector reflecting the curve feature. The vector contains key information such as curve curvature, curve length, and turning direction, which is used to characterize the channel change characteristics in the curve scenario. After obtaining the curve feature vector, the sending end inputs it into a pre-trained generative adversarial network.

[0143] The generative adversarial network is a neural network model based on the idea of adversarial learning, which consists of a generator and a discriminator. The generator is used to generate false samples as similar as possible to real samples, and the discriminator is used to judge whether the sample is real or false. Through the game learning between the generator and the discriminator, the generative adversarial network can learn the internal distribution and generation rule of the data. In this embodiment, the system uses the generative adversarial network to learn the mapping relationship between the channel feature in the curve scenario and the network structure.

[0144] Specifically, the generator of the generative adversarial network takes the bend feature vector as input and outputs the corresponding neuron connection mode. The connection mode includes the connection topology between neurons, connection weights and other information, which determines the structure and parameters of the neural network. Through continuous adversarial learning with the discriminator, the generator gradually masters the mapping rule between the bend feature and the optimal network structure. On the one hand, the generator generates a connection mode that matches the bend feature as much as possible, so that the reconstructed neural network can well adapt to the channel changes in the bend scene; on the other hand, the generator generates samples close to the real connection mode distribution as much as possible to deceive the identification of the discriminator. After multiple rounds of adversarial training, the generator finally learns to generate the optimal neuron connection mode according to the bend feature.

[0145] After obtaining the new connection mode, the sending end reconstructs the original fully connected neural network accordingly to obtain a new network that adapts to the bend scene. The reconstruction process includes adjusting the connection relationship between neurons, modifying the connection weights, adding or deleting neurons, and other operations, so that the structure of the network can dynamically adapt to the changes in the bend feature. It should be noted that the network reconstruction is performed on the basis of the original network, rather than completely starting over. This is because the original network is trained on a large amount of normal scene data and already has certain feature extraction and mapping capabilities, and completely retraining is too costly. By making local adjustments based on the original network, the existing knowledge can be preserved while the network structure is optimized to adapt to the particularity of the bend scene.

[0146] After reconstructing the fully connected neural network, the sending end inputs the first channel estimation sequence into the new network to generate the corresponding mapping channel estimation sequence. Since the structure of the reconstructed network is optimized for the bend scene, the mapping sequence generated by the network can more accurately reflect the feature changes of the bend channel, thereby improving the reliability of key generation. When the train leaves the bend and returns to the straight track, the sending end can switch back to the original fully connected neural network to save computing resources.

[0147] In the above embodiment, the angular velocity and position information are detected to determine the curve scene, and the angular velocity and position information are fused to generate a curve feature vector, and a pre-trained generative adversarial network is used to generate a new neuron connection method based on the curve feature vector, and the topological structure of the fully connected neural network is reconstructed accordingly. In the curve scene, due to the multipath effect and scattering effect caused by the turning movement of the train, the channel characteristics will show a unique change pattern. The generative adversarial network obtains the mapping relationship between the channel characteristics and the optimal network structure in the curve scene through adversarial training, and can generate a matching neuron connection method based on the specific curve characteristics. This method of dynamically reconstructing the network topology enables the structure of the neural network to be optimized and adjusted according to the characteristics of the curve scene, avoiding the performance degradation problem that may occur in the fixed network structure in the curve scene, and improving the channel feature mapping accuracy of the neural network in the curve scene.

[0148] The following describes the system in the embodiment of the present invention from the perspective of hardware processing. Figure 3 , which is a schematic diagram of the physical device structure of a channel reconstruction security key extraction system for rail transit provided in an embodiment of the present application.

[0149] It should be noted that Figure 3 The structure of the system shown is only an example and should not limit the functions and scope of use of the embodiments of the present invention.

[0150] like Figure 3 As shown, the system includes a central processing unit (CPU) 301, which can perform various appropriate actions and processes, such as the methods described in the above embodiments, based on programs stored in a read-only memory (ROM) 302 or programs loaded from a storage unit 308 into a random access memory (RAM) 303. RAM 303 also stores various programs and data required for system operation. CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to bus 304.

[0151] The following components are connected to the I / O interface 305: an input section 306 including a camera, infrared sensor, and the like; an output section 307 including a liquid crystal display (LCD) and speakers; a storage section 308 including a hard disk and the like; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card or a modem. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the I / O interface 305 as needed. Removable media 311, such as a magnetic disk, optical disk, magneto-optical disk, or semiconductor memory, is installed in the drive 310 as needed, so that computer programs read from the media can be installed in the storage section 308 as needed.

[0152] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for executing the methods illustrated in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via the communication section 309 and / or installed from removable media 311. When executed by the central processing unit (CPU) 301, the computer program performs the various functions defined in the present invention.

[0153] It should be noted that the computer-readable medium described in the embodiments of the present invention may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, an electrical connection having one or more conductors, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium may include a data signal transmitted in baseband or as part of a carrier wave, which carries a computer-readable computer program. Such a propagated data signal may take any of a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof.

[0154] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. Each box in the flowchart or block diagram can represent a module, program segment, or part of the code, and the above-mentioned module, program segment, or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the boxes can also occur in an order different from that marked in the accompanying drawings. For example, two boxes shown in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or can be implemented using a combination of dedicated hardware and computer instructions.

[0155] As another aspect, the present invention further provides a computer-readable storage medium, which may be included in the system described in the above embodiments, or may exist independently and not incorporated into the system. The storage medium carries one or more computer programs, and when executed by a processor of a system, the system implements the methods provided in the above embodiments.

[0156] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

[0157] As used in the above embodiments, the term “when…” may be interpreted as “if…” or “after…” or “in response to determining…” or “in response to detecting…”, depending on the context. Similarly, the phrases “upon determining…” or “if (stated condition or event) is detected” may be interpreted as “if determining…” or “in response to determining…” or “upon detecting (stated condition or event)” or “in response to detecting (stated condition or event)”, depending on the context.

[0158] In the above embodiments, all or part of the embodiments can be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, optical fiber, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, hard disk, tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive).

[0159] Those skilled in the art will appreciate that all or part of the process steps in the above-described method embodiments can be implemented by a computer program instructing the relevant hardware. The program can be stored in a computer-readable storage medium, and when executed, the program can include the process steps in the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A channel reorganization security key extraction method for rail transit, applied to the system, characterized in that: The system includes a sending end and a legal receiving end, and the method includes: The transmitting end uses the collected uplink and downlink channel data as raw data to construct a fully connected neural network consisting of an input layer, four hidden layers and an output layer; The transmitting end and the legitimate receiving end send pilot signals to each other to obtain a first channel estimation sequence of the original channel corresponding to the transmitting end and a second channel estimation sequence of the original channel corresponding to the legitimate receiving end; The transmitting end inputs the first channel estimation sequence into the fully connected neural network to obtain a mapped channel estimation sequence; The transmitting end generates a channel reassembly table based on a preset initial key and the mapped channel estimation sequence; After receiving the channel reassembly table, the legal receiving end permutes the second channel estimation sequence according to elements in the channel reassembly table to obtain a permuted channel estimation sequence; The legitimate receiving end decodes the permuted channel estimation sequence to obtain a security key; The fully connected neural network is a neural network with a fully connected structure including one input layer, four hidden layers, and one output layer. The input layer receives the first channel estimation sequence as an input feature, and the number of neurons is the same as the dimension of the first channel estimation sequence. The four hidden layers include a first preset number, a second preset number, a third preset number, and a fourth preset number of neurons, respectively. The four hidden layers perform inter-layer nonlinear transformation through a ReLU activation function, and a Dropout layer with a preset second ratio is set after each hidden layer. The output layer generates a mapping channel estimation sequence with the same dimension as the second channel estimation sequence. The transmitting end trains the fully connected neural network using the Adam optimization algorithm based on the uplink and downlink channel data as a training set and a test set, and uses a loss function to minimize the difference between the mapping channel estimation sequence output by the network and the second channel estimation sequence. The first preset number, the second preset number, the third preset number, and the fourth preset number are in decreasing order. The transmitting end generates a channel reassembly table based on a preset initial key and the mapped channel estimation sequence, specifically including: The transmitting end encodes the preset initial key to obtain a codeword sequence, where the preset initial key is generated by a random source that obeys a 0 / 1 uniform distribution; The transmitting end sorts the mapped channel estimation sequence by size to obtain a sorted mapped channel estimation sequence, and determines a position of each element in the sorted mapped channel estimation sequence in the mapped channel estimation sequence; The transmitting end stores the elements of the codeword sequence with the first preset value in the order of the sorted mapping channel estimation sequence to the position in sequence, and stores the elements of the codeword sequence with the second preset value in the reverse order of the sorted mapping channel estimation sequence to the position in sequence, to obtain a channel reorganization table.

2. The method according to claim 1, characterized in that The loss function is: ； In the above function, is the mapping channel estimation sequence, the is the second channel estimation sequence, is the number of training samples, is the independent variable, For the The mapping channel estimation sequence corresponding to the training samples, For the said A corresponding second channel estimation sequence.

3. The method according to claim 1, characterized in that After the legal receiving end decodes the permuted channel estimation sequence to obtain a security key, the method further includes: The transmitting end obtains a first Doppler frequency shift value, and the legal receiving end obtains a second Doppler frequency shift value; When the difference between the first Doppler frequency shift value and the second Doppler frequency shift value is greater than a first preset threshold, determining that a train passing scene has been entered; The transmitting end inputs the first Doppler frequency shift value into a pre-trained multi-scale convolutional neural network to obtain an output result; The transmitting end determines the number of spare neurons that need to be temporarily activated in each hidden layer according to the output result; The transmitting end initializes the weight of the number of spare neurons according to the channel characteristics in the train passing scenario to obtain a corrected fully connected neural network; After replacing the fully connected neural network with the corrected fully connected neural network, the transmitting end performs the step of inputting the first channel estimation sequence into the fully connected neural network.

4. The method according to claim 3, characterized in that The sending end determines the number of spare neurons that need to be temporarily activated in each hidden layer according to the output result, specifically including: The transmitting end divides the output result into a plurality of subsequences according to a preset number for each hidden layer of the fully connected neural network; The transmitting end performs Fourier transform on each of the subsequences to obtain a frequency domain feature sequence; The transmitting end determines the number of spare neurons that need to be temporarily activated in each hidden layer according to the proportion of high-frequency components in the frequency domain feature sequence, and the proportion of the high-frequency components is proportional to the number of spare neurons by a preset coefficient.

5. The method according to claim 1, wherein After the legal receiving end decodes the permuted channel estimation sequence to obtain a security key, the method further includes: The transmitting end obtains a first angular velocity and a first position information based on an inertial navigation unit, and the legal receiving end obtains a second angular velocity and a second position information based on the inertial navigation unit; When the first angular velocity is greater than a second preset threshold and the distance between the first position information and the second position information is less than a third preset threshold, determining that the vehicle enters a curve scene; The transmitting end performs a fusion process on the first angular velocity and the first position information to generate a curve feature vector; The transmitting end inputs the curve feature vector into a pre-trained generative adversarial network to generate a new neuron connection mode; The transmitting end reconstructs the topological structure of the fully connected neural network according to the new neuron connection mode to obtain a reconstructed fully connected neural network; After replacing the fully connected neural network with the reconstructed fully connected neural network, the transmitting end performs the step of inputting the first channel estimation sequence into the fully connected neural network.

6. A channel reorganization security key extraction system for rail transit, characterized in that: The system comprises: One or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to cause the system to execute the method according to any one of claims 1 to 5.

7. A computer-readable storage medium comprising instructions, characterized in that: When the instructions are executed on a system, the system is caused to perform the method according to any one of claims 1 to 5.

8. A computer program product, characterized in that When the computer program product is run on a system, the system is caused to perform the method according to any one of claims 1 to 5.

Images