Cross-frequency channel characteristic parameter migration method and system

Abstract

The invention discloses a cross-frequency channel characteristic parameter migration method and system, and the method comprises the following steps: collecting and preprocessing an original signal of a channel, and obtaining three channel characteristics; channel time-frequency response is regarded as an image, the time-frequency image is input into a scene recognition module to extract image features, the image features are combined with the three channel features, and a scene recognition vector SRV is generated; determining the weight of a lightweight path or a complex path according to the scene recognition vector SRV and the gradient information returned by the path, and optimizing the feature representation to obtain a source frequency feature vector; combining the source frequency feature vector with a frequency control vector from a target domain, and generating a target frequency feature through nonlinear transformation; and carrying out deconvolution operation on the target frequency features, and recovering the low-dimensional feature vectors into a high-dimensional time-frequency graph. By adopting the technical scheme, the channel characteristic parameters are migrated, the consumption of external field actual measurement resources is reduced, and the efficiency of channel measurement and simulation reproduction is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of propagation channels and relates to a method and system for migrating characteristic parameters of cross-frequency channels. Background Art

[0002] The study of propagation channels is the basis for wireless communication system design, network optimization and performance evaluation. The statistical characteristics and dynamic changes of channels determine the overall performance of the communication system and directly affect the reliability, efficiency and security of communication. Channel modeling is to characterize channel characteristics in a mathematical way. An accurate channel model can optimize data transmission quality and reduce communication delay.

[0003] The research on channel modeling cannot be separated from the support of channel data. However, field channel measurement is not only time-consuming and labor-intensive, but is also often affected by problems such as frequency interference. It is difficult to cover all scenarios and frequency bands, resulting in a lack of sufficient data support for scenario channel research at specific frequencies.

[0004] The goal of channel characteristic parameter migration is to use the channel data collected at other frequencies to generate the channel characteristic parameters of the target frequency. With the continuous development and maturity of deep learning technology, the channel generation and modeling methods based on neural networks have received more and more attention.

[0005] Neural networks and deep learning methods are considered to be potential general solutions to the limitations of traditional modeling methods in different scenarios and environments due to their powerful learning and prediction capabilities, excellent nonlinear fitting and adaptive capabilities, and advantages in mining complex features in high-dimensional and highly redundant data. For example, GAN is applied to the field of wireless channel modeling to model additive white Gaussian noise (AWGN), and the target channel model is learned from a large amount of raw channel data by realizing the Nash equilibrium of the minimum-maximum game between the channel data generator and the channel data discriminator. Or a cross-frequency channel data generation and modeling framework based on CycleGAN, which can obtain channel data through generative adversarial networks.

[0006] At present, the cross-frequency channel feature parameter migration framework based on MMD-AAE and CycleGAN has overcome the limitations of traditional modeling technology and provided a new idea for modeling in complex channel environments. However, the existing deep learning architectures such as convolution and GAN networks are too fixed and cannot dynamically adjust the feature flow and the interaction between network layers. They cannot adapt to changes in different environments and fail to effectively utilize the interaction between local and global features, resulting in insufficient feature extraction. Especially in complex channel environments, it is impossible to capture the multi-scale and multi-dimensional features of the input data, which reduces the generalization ability and adaptability of the model. In addition, the existing methods lack an adaptive mechanism when migrating between different frequencies, resulting in poor feature generation for unknown frequencies, affecting the accuracy and consistency of the conversion. Summary of the invention

[0007] The purpose of the present invention is to provide a method and system for migrating channel characteristic parameters across frequencies, so as to migrate channel characteristic parameters, save the consumption of field measurement resources, and improve the efficiency of channel measurement and simulation reproduction.

[0008] In order to achieve the above object, the basic scheme of the present invention is: a method for migrating cross-frequency channel characteristic parameters, comprising the following steps:

[0009] The original channel signal is collected and preprocessed to obtain three channel characteristics, namely power delay spectrum PDP, root mean square delay spread RMS and Rice K factor K. factor ;

[0010] The channel time-frequency response is regarded as an image, and the time-frequency image is input into the scene recognition module to extract image features, and combined with the three channel features extracted above to generate a scene recognition vector SRV;

[0011] According to the scene recognition vector SRV and the gradient information returned by the path, the weights of the lightweight path and the complex path are determined, the feature representation is optimized, and the source frequency feature vector is obtained;

[0012] The source frequency feature vector and the frequency control vector from the target domain are combined, and then subjected to nonlinear transformation to generate a target frequency feature;

[0013] Deconvolution is performed on the target frequency features to restore the low-dimensional feature vector to a high-dimensional time-frequency graph.

[0014] The working principle and beneficial effects of this basic solution are: this technical solution breaks the traditional fixed network structure through a dynamic feature fusion mechanism, and the network can be dynamically adjusted according to the input channel characteristics to adapt to feature generation at different frequencies.

[0015] Through local feature interaction, inter-block feature aggregation, and global feature compression, the present invention efficiently extracts multi-scale and multi-dimensional channel features in the time-frequency diagram and generates a scene recognition vector (SRV) representing the overall scene features. This design enhances the generalization ability and robustness of the model in complex channel environments, can adapt to changing channel scenarios, and improves the adaptability and robustness of the model.

[0016] Based on frequency mapping combined with frequency control vector (FCV), the present invention can achieve accurate conversion from source frequency to target frequency. FCV is generated based on target domain data, providing guidance of target frequency characteristics for mapping, thereby ensuring that the mapping result can better reflect the characteristics of the target frequency and improve the accuracy and consistency of frequency conversion.

[0017] Furthermore, the original channel signal is collected and preprocessed to obtain the three channel characteristics:

[0018] Clean and slice the original channel signal;

[0019] When extracting channel characteristics, the power delay profile PDP is obtained by analyzing the channel impulse response, and the root mean square delay spread τ of each sample is calculated. rms and Rice factor K factor :

[0020]

[0021]

[0022] Among them, τ k ,a k ,p(τ k ) represent the channel delay, amplitude and power of the kth path respectively, P Los In the case of a direct line of sight, it represents the power of the direct line of sight path; in the case of a scene without a direct line of sight path, it represents the power of the strongest multipath component. mpc Represents the power of the remaining multipath components; is the average relative delay, defined as the first-order moment of PDP, which represents the average propagation delay of the multipath component relative to the first path:

[0023]

[0024] Furthermore, the time-frequency image is input into the scene recognition module, image features are extracted, and the steps of generating the scene recognition vector SRV are as follows:

[0025] The format of the source channel time-frequency response image is [B, C, H, W], where B represents the batch, C represents the number of channels, H represents the length, and W represents the width;

[0026] The input time-frequency image is divided into blocks of size nxn to generate a local block set B S ;

[0027] Perform convolution operation on each local block, with the convolution kernel size of nxn and the step size consistent with the convolution kernel size to ensure no overlapping convolution;

[0028] Based on the feature map within the local block, the interaction matrix between local blocks is constructed, and the interaction feature map is obtained by weighted aggregation by calculating the similarity between the local block and its adjacent blocks;

[0029] The interaction feature map [B, C', H', W'] is compressed to [B, C'] by global average pooling, where B represents the batch, C' represents the number of channels of the interaction feature map, H' represents the length of the interaction feature map, and W' represents the width of the interaction feature map. The fully connected layer is used to map the features to a low-dimensional space to obtain Img, and the extracted three channel features are compared with the power delay spectrum PDP, root mean square delay spread RMS and Rice K factor K. factor The scene recognition vector (SRV) is generated by combining the following formula, and the overall characteristics of the channel scene are obtained through layer-by-layer feature extraction and compression:

[0030] SRV=[Img,PDP,RMS,K factor ].

[0031] Furthermore, according to the scene recognition vector SRV and the gradient information returned by the path, the weights of the lightweight path and the complex path are determined, the feature representation is optimized, and the source frequency feature vector is obtained. The specific method is as follows:

[0032] According to SRV, the outputs of the lightweight path and the complex path are weighted summed using the gating unit to generate the final feature representation:

[0033] F=αK L +(1-α)K D ,

[0034] Among them, K L The feature vector representing the output of the lightweight path, K D A feature vector representing the output of a complex path; SRV is the scene recognition vector; gradient information and It is calculated by the network during the forward propagation and back propagation processes, corresponding to the gradients of simple paths and complex paths respectively; the gradient information and SRV are not obtained at the same time. The gradient information that passes through the gating unit each time is obtained in the previous round of propagation. Finally, the Gating function dynamically generates the weight α based on this information to optimize the feature representation.

[0035] According to SRV, appropriate sub-network paths are selected to process channel features of different complexities. The sub-network paths are divided into simple paths (lightweight paths) and complex paths. The simple paths use a shallower network structure to process simple scenarios, such as input data with less changes; the complex paths are used to process complex scenarios, with a deeper network structure and contain more advanced feature extraction and attention mechanisms.

[0036] Furthermore, the source frequency feature vector and the frequency control vector from the target domain are combined, and then the target frequency feature is generated through nonlinear transformation. The specific steps are as follows:

[0037] Input the time-frequency diagram of the target domain [B, C', H', W'], which represents the channel characteristics at the target frequency;

[0038] Use the convolution layer to extract the characteristics of the target frequency to ensure that multi-scale and multi-dimensional frequency domain features are extracted from the time-frequency diagram; the convolution layer parameters are the number of convolution kernels k, the size of the convolution kernel n*n, and the input target frequency time-frequency diagram [B, C', H', W'] to obtain the output convolution feature diagram [B, k, H', W'];

[0039] Perform global pooling on the convolutional features to generate global features of the target domain, that is, input the convolutional features [B, k, H', W'], and output the target domain global feature vector [B, k], where B represents the batch, C' represents the number of channels of the interaction feature map, H' represents the length of the interaction feature map, and W' represents the width of the interaction feature map;

[0040] The global feature vector of the target domain is mapped to FCV through a fully connected layer, that is, the input is the global feature vector [B, k], and the output is FCV [B, D f ], where D f It is the dimension of FCV, which serves as a guide to the target frequency;

[0041] Concatenate the source frequency feature vector [B, F] and FCV to obtain the concatenated feature vector [B, F+D f ];

[0042] Perform linear transformation on the concatenated feature vectors to generate a higher-dimensional feature vector [B, 8k];

[0043] The feature vector [B,8k] is nonlinearly transformed using the ReLU activation function and then mapped to the target frequency feature space to finally generate the target frequency feature [B,F'], where F' is the feature dimension at the target frequency.

[0044] Receives the feature vector from the dynamic path selection module and then maps the source frequency channel features to the target frequency features.

[0045] Furthermore, the target frequency feature is deconvolved to restore the low-dimensional feature vector to a high-dimensional time-frequency graph. The specific steps are as follows:

[0046] The one-dimensional target frequency feature [B, F'] is converted into a four-dimensional tensor and upsampled multiple times to restore the target frequency feature to the same size as the original time-frequency graph;

[0047] Sigmoid activation is used for the target frequency feature to limit the output value range to [0,1] to obtain a high-dimensional time-frequency diagram.

[0048] Deconvolving the target frequency feature vector into a time-frequency graph representing the target frequency is simple.

[0049] Further, the loss function is introduced:

[0050] Frequency consistency loss Measure the difference between the generated colored time-frequency map and the true target frequency map to ensure that the generated time-frequency map meets the target frequency characteristics:

[0051]

[0052] Among them, N represents the batch size, that is, the number of samples input into the network during one training process, and x i is the source frequency image; FCV is the feature representation of the target extraction; G(x i ,FCV): Color time-frequency diagram generated by the frequency mapping layer; i : Colored time-frequency diagram of the true target frequency; |·|2 is the L2 norm;

[0053] Introducing L2 regularization reg , limiting the size of the model parameters and encouraging the model to find smoother and simpler solutions:

[0054]

[0055] Among them, λ: regularization coefficient, which controls the strength of the regularization term and usually needs to be tuned through experiments; θ: the set of trainable parameters in the model; w: each trainable parameter;

[0056] The total loss function L total Combining frequency map consistency loss and regularization term:

[0057] L total =L freq +L reg .

[0058] Set the loss function and optimize the calculation of each part of the model.

[0059] The present invention also provides a system for migrating characteristic parameters of a flying frequency channel based on the method of the present invention, comprising an input layer, a scene recognition module SRM, a dynamic path selection module DPS, a frequency mapping layer FML and an output layer;

[0060] The input layer is used to collect the original channel signal and perform preprocessing to obtain a channel characteristic data set;

[0061] The scene recognition module SRM and the dynamic path selection module DPS are both connected to the input layer, and the scene recognition module SRM includes a local feature interaction layer, an intra-block feature extraction layer, an interactive feature aggregation layer, and a feature compression and scene representation layer connected in sequence;

[0062] The input end of the frequency mapping layer FML is connected to the output end of the dynamic path selection module DPS, the input end of the output layer is connected to the output end of the frequency mapping layer FML, and the output layer includes a fully connected layer, a first deconvolution layer, a second deconvolution layer, a third deconvolution layer and a nonlinear activation layer.

[0063] This system uses various modules and levels to obtain accurate characteristic parameter models through network-based training of measured source data, and then realizes the migration of channel characteristic parameters, which greatly saves the consumption of field measurement resources and improves the efficiency of channel measurement and simulation reproduction.

[0064] Further, the dynamic path selection module DPS includes a lightweight path and a complex path;

[0065] The lightweight path includes 2 convolutional layers, 2 pooling layers, 1 fully connected layer, and 1 regularization layer:

[0066] The first convolutional layer is used to initially extract features, and the spatial dimension is consistent with the input;

[0067] The first pooling layer is used to reduce the spatial dimension and reduce the amount of calculation;

[0068] The second convolutional layer is used to extract higher-level features;

[0069] The second pooling layer is used to reduce the spatial dimension;

[0070] Fully connected layer, used to output the final feature vector K L ;

[0071] Regularization layer Dropout is used to prevent overfitting;

[0072] The complex path includes 4 convolutional layers, 2 residual blocks, 2 pooling layers, and 1 fully connected layer:

[0073] Convolutional layer 1: preserves spatial dimensions;

[0074] Residual block 1: contains two convolutional layers: convolutional layer 2 and convolutional layer 3. Skip connections are introduced to prevent gradient disappearance and maintain information transmission.

[0075] Pooling layer 1: downsample the output feature map to reduce the spatial dimension;

[0076] Convolutional layer 4: further extract features;

[0077] Residual block 2: two convolutional layers: convolutional layer 5 and convolutional layer 6, extracting complex features while keeping the gradient stable;

[0078] Pooling layer 2: further downsampling;

[0079] Convolutional layer seven: continue to extract features;

[0080] Fully connected layer: maps high-dimensional features into low-dimensional vectors and outputs the final feature vector K D ;

[0081] Regularization layer: reduce the risk of overfitting;

[0082] The output ends of the lightweight path and the complex path are both connected to the gating unit, and the gating unit is connected to the scene recognition module and the frequency mapping layer FML respectively.

[0083] The lightweight path uses fewer parameters, is suitable for fast calculations, and can effectively handle low-complexity inputs. The complex path is deeper, can handle complex scenes, has stronger feature extraction capabilities, and is especially suitable for highly dynamic environments.

[0084] Further, the frequency mapping layer FML includes an FCV generation module, a splicing module, a first fully connected layer, a nonlinear activation layer and a second fully connected layer connected in sequence; the FCV generation module is used to generate a vector from the data of the target domain through a feature extraction network, representing the global feature information of the target frequency;

[0085] The concatenation module is used to concatenate the source frequency feature vector [B, F] and the FCV, and the first fully connected layer is used to perform a linear transformation on the concatenated feature vector to generate a higher-dimensional feature representation;

[0086] The nonlinear activation layer uses the ReLU activation function to perform a nonlinear transformation on the features output by the first fully connected layer, and the second fully connected layer maps the activated features to the target frequency feature space, and finally generates the target frequency features.

[0087] The frequency mapping layer FML has a simple structure and is easy to use. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 It is a schematic diagram of the flow of the cross-frequency channel characteristic parameter migration method of the present invention;

[0089] Figure 2 It is a linear schematic diagram of a power delay profile PDP of the cross-frequency channel characteristic parameter migration method of the present invention;

[0090] Figure 3 It is a structural schematic diagram of a scene recognition module SRM of a cross-frequency channel characteristic parameter migration system of the present invention;

[0091] Figure 4 It is a structural schematic diagram of a dynamic path selection module DPS of a cross-frequency channel characteristic parameter migration system of the present invention;

[0092] Figure 5 It is a structural schematic diagram of a frequency mapping layer FML of a cross-frequency channel characteristic parameter migration system of the present invention;

[0093] Figure 6 It is a structural diagram of the output layer of the cross-frequency channel characteristic parameter migration system of the present invention. DETAILED DESCRIPTION

[0094] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0095] In the description of the present invention, it is necessary to understand that the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0096] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0097] The present invention discloses a method for migrating cross-frequency channel characteristic parameters, breaking the traditional fixed network structure, selecting appropriate subnet paths according to input channel characteristics to process different channel data, adapting to feature generation at different frequencies, and thus adapting to channel data in different scenarios; the scene recognition module can extract features according to the input time-frequency graph, and generate a scene recognition vector (SRV) in combination with the channel features, thereby guiding path selection and enhancing the accuracy of selection. Figure 1 As shown, the cross-frequency channel characteristic parameter migration method includes the following steps:

[0098] The original channel signal is collected and preprocessed to obtain three channel characteristics, namely power delay spectrum PDP, root mean square delay spread RMS and Rice K factor K. factor ;

[0099] The channel time-frequency response is regarded as an image, and the time-frequency graph is input into the scene recognition module to extract image features, which are combined with the three channel features extracted above to generate a scene recognition vector SRV;

[0100] According to the scene recognition vector SRV and the scene complexity, the weight of the lightweight path or the complex path is determined, the feature representation is optimized, and the source frequency feature vector is obtained; the weights of the two paths are determined according to the output of the gating unit, and the output of the gating unit is determined by SRV and gradient feedback.

[0101] The source frequency feature vector and the frequency control vector from the target domain are combined, and then subjected to nonlinear transformation to generate a target frequency feature;

[0102] Deconvolution is performed on the target frequency features to restore the low-dimensional feature vector to a high-dimensional time-frequency graph.

[0103] In a preferred embodiment of the present invention, the establishment of a data set is the basis of the cross-frequency channel migration model. However, due to the roughness of the original channel data, it is necessary to preprocess it and extract the channel characteristics. The method of collecting the original channel signal and preprocessing it to obtain the three channel characteristics is:

[0104] Clean and slice the original channel signal, including removing invalid and redundant data and slicing periodic data;

[0105] When extracting channel characteristics (extracting channel-related characteristics, namely, power delay spectrum, root mean square delay spread, and Rice K factor), the power delay spectrum PDP is obtained by analyzing the channel impulse response. The PDP can reflect the power size of the direct-line path and the power, relative delay, and number of other multipath components, such as Figure 2 As shown in the figure, the horizontal axis is the relative delay and the vertical axis is the power spectrum delay. Calculate the root mean square delay spread τ of each sample rms and Rice factor Kfactor :

[0106]

[0107]

[0108] Among them, τ k ,a k ,p(τ k ) represent the channel delay, amplitude and power of the kth path respectively, P Los In the case of a direct line of sight, it represents the power of the direct line of sight path; in the case of a scene without a direct line of sight path, it represents the power of the strongest multipath component. mpc Represents the power of the remaining multipath components; is the average relative delay, defined as the first-order moment of PDP, which represents the average propagation delay of the multipath component relative to the first path:

[0109]

[0110] In a preferred embodiment of the present invention, the steps of inputting the time-frequency graph into the scene recognition module, extracting image features, and combining them with the three channel features to generate a scene recognition vector SRV are as follows:

[0111] The format of the source channel time-frequency response is [B,C,H,W], where B represents the batch, C represents the number of channels, H represents the length, and W represents the width;

[0112] The input time-frequency image is divided into blocks of size nxn to generate a local block set B S ;

[0113] Perform convolution operation on each local block, with the convolution kernel size of nxn and the step size consistent with the convolution kernel size to ensure no overlapping convolution;

[0114] Based on the feature maps within the local blocks, the similarities between them are calculated to form a similarity matrix. According to the similarities in the similarity matrix, a weighted matrix is ​​constructed to represent the weighted ratio of each block to its adjacent blocks. The features of adjacent blocks are weightedly aggregated using the weighted matrix to obtain an interactive feature map;

[0115] The interaction feature map [B, C', H', W'] is compressed to [B, C'] by global average pooling, where B represents the batch, C' represents the number of channels of the interaction feature map, H' represents the length of the interaction feature map, and W' represents the width of the interaction feature map. The fully connected layer is used to map the features to a low-dimensional space to obtain Img, and the extracted three channel features are compared with the power delay spectrum PDP, root mean square delay spread RMS and Rice K factor K. factorThe scene recognition vector (SRV) is generated by combining the following formula, and the overall characteristics of the channel scene are obtained through layer-by-layer feature extraction and compression:

[0116] SRV=[Img,PDP,RMS,K factor ].

[0117] In a preferred embodiment of the present invention, the weights of the lightweight path and the complex path are determined according to the scene recognition vector SRV and the gradient information returned by the path, the feature representation is optimized, and the source frequency feature vector is obtained. The specific method is:

[0118] The simple path (i.e., lightweight path) uses a shallower network structure to process simple scenarios, such as input data with less variation. The complex path is used to process complex scenarios, with a deeper network structure that includes more advanced feature extraction and attention mechanisms.

[0119] According to SRV, a gating unit (e.g., Gating Unit) is used to perform a weighted summation of the outputs of the lightweight path and the complex path to generate the final feature representation:

[0120] F=αK L +(1-α)K D ,

[0121] Among them, K L The feature vector representing the output of the lightweight path, K D A feature vector representing the output of a complex path; SRV is the scene recognition vector; gradient information and It is calculated by the network during the forward propagation and back propagation processes, corresponding to the gradients of simple paths and complex paths respectively; the gradient information and SRV are not obtained at the same time. The gradient information that passes through the gating unit each time is obtained in the previous round of propagation. Finally, the Gating function dynamically generates the weight α based on this information to optimize the feature representation.

[0122] The gradient information is a vector, and its acquisition process is:

[0123] The loss function is calculated through forward propagation, and the chain rule is used in back propagation. By calculating the gradient of the loss function to the output and the gradient of the output to the intermediate layer, the gradient is calculated backward layer by layer until the input layer is reached. The gradient of each layer can be expressed as:

[0124]

[0125] Among them, L is the loss function, a is the output of the current layer, z i is the result of linear transformation, ω i is the weight of the i-th layer.

[0126] The process of obtaining the gate control unit:

[0127] The gate control unit is composed of three vectors: input gate vector, forget gate vector and output gate vector, which are combined and multiplied by the weight matrix of the gate control unit. The weight matrix can be obtained through training.

[0128] Through gradient information, path selection can be dynamically adjusted during training and inference, computing resource allocation can be optimized, model convergence can be accelerated, and adaptability, generalization, and robustness can be improved. This makes the network more efficient and intelligent in a changing channel environment.

[0129] In a preferred embodiment of the present invention, the source frequency feature vector and the frequency control vector from the target domain are combined, and then the target frequency feature is generated through nonlinear transformation. The specific steps are as follows:

[0130] Input the time-frequency diagram of the target domain [B, C', H', W'], which represents the channel characteristics at the target frequency;

[0131] Use the convolution layer to extract the characteristics of the target frequency to ensure that multi-scale and multi-dimensional frequency domain features are extracted from the time-frequency diagram; the convolution layer parameters are the number of convolution kernels k (such as 64), the convolution kernel size n*n, and the input target frequency time-frequency diagram [B, C', H', W'] to obtain the output convolution feature diagram [B, k, H', W'];

[0132] Perform global pooling on the convolutional features to generate global features of the target domain, that is, input the convolutional features [B, k, H', W'], and output the target domain global feature vector [B, k], where B represents the batch, C' represents the number of channels of the interaction feature map, H' represents the length of the interaction feature map, and W' represents the width of the interaction feature map;

[0133] The global feature vector of the target domain is mapped to FCV through a fully connected layer, that is, the input is the global feature vector [B, k], and the output is FCV [B, D f ], where D f is the dimension of the FCV (e.g. 2k), serving as a guide to the target frequency;

[0134] Concatenate the source frequency feature vector [B, F] and FCV to obtain the concatenated feature vector [B, F+D f ];

[0135] Perform linear transformation on the concatenated feature vectors to generate a higher-dimensional feature vector [B, 8k];

[0136] The feature vector [B,8k] is nonlinearly transformed using the ReLU activation function and then mapped to the target frequency feature space to finally generate the target frequency feature [B,F'], where F' is the feature dimension at the target frequency.

[0137] In a preferred embodiment of the present invention, a deconvolution operation is performed on the target frequency feature to restore the low-dimensional feature vector to a high-dimensional time-frequency diagram. The specific steps are:

[0138] The one-dimensional target frequency feature [B, F'] is converted into a four-dimensional tensor and upsampled multiple times to restore the target frequency feature to the same size as the original time-frequency graph;

[0139] Sigmoid activation is used for the target frequency feature to limit the output value range to [0,1] to obtain a high-dimensional time-frequency diagram.

[0140] In a preferred embodiment of the present invention, a loss function is introduced:

[0141] Frequency Consistency Loss Measure the difference between the generated colored time-frequency map and the true target frequency map to ensure that the generated time-frequency map meets the target frequency characteristics:

[0142]

[0143] Among them, N represents the batch size, that is, the number of samples input into the network during one training process, and x i is the source frequency image; FCV is the feature representation of the target extraction; G(x i ,FCV): Color time-frequency diagram generated by the frequency mapping layer; i : Colored time-frequency diagram of the true target frequency; |·|2 is the L2 norm;

[0144] In order to prevent the network from overfitting, L2 regularization term is introduced. reg , limiting the size of the model parameters and encouraging the model to find smoother and simpler solutions:

[0145]

[0146] Among them, λ: regularization coefficient, which controls the strength of the regularization term and usually needs to be tuned through experiments; θ: the set of trainable parameters in the model; w: each trainable parameter;

[0147] Total Loss Function L total Combining frequency map consistency loss and regularization term:

[0148] L total =L freq +L reg .

[0149] The present invention also provides a cross-frequency channel characteristic parameter migration system based on the method of the present invention, comprising an input layer, a scene recognition module SRM, a dynamic path selection module DPS, a frequency mapping layer FML and an output layer. The input layer is used to collect the original channel signal and perform preprocessing to obtain a channel characteristic data set.

[0150] The scene recognition module SRM and the dynamic path selection module DPS are both connected to the input layer. The scene recognition module SRM includes a local feature interaction layer, an intra-block feature extraction layer, an interactive feature aggregation layer, and a feature compression and scene representation layer connected in sequence, such as Figure 3 The scene recognition module extracts channel features from the input time-frequency diagram and generates a scene recognition vector (SRV). It obtains the overall characteristics of the channel scene through layer-by-layer feature extraction and compression from the input data.

[0151] In the scene recognition module SRM:

[0152] 1. Local feature interaction layer

[0153] Function: Divide the input time-frequency image into multiple non-overlapping local blocks, extract local features and interact between blocks.

[0154] Implementation method: Divide the image into blocks of size nxn to generate a local block set B S ,The information within the block is obtained through feature extraction.

[0155] 2. Intra-block feature extraction layer

[0156] Function: Perform convolution operation on each local block to extract local features.

[0157] Implementation: Apply convolution operation to each local block, with the convolution kernel size of nxn and the step size consistent with the convolution kernel size to ensure no overlapping convolution.

[0158] 3. Interactive Feature Aggregation Layer

[0159] Function: Realize feature interaction between local blocks and capture spatial changes between adjacent blocks.

[0160] Implementation method: Based on the intra-block feature map, an inter-block interaction matrix is ​​constructed, and weighted aggregation is performed by calculating the similarity between the block and its adjacent blocks (such as Euclidean distance or cosine similarity).

[0161] 4. Feature Compression and Scene Representation Layer

[0162] Function: Globally represent the interacted features and generate a low-dimensional scene recognition vector (SRV).

[0163] Implementation: Compress the interaction feature map [B, C', H', W'] to [B, C'] through global average pooling, where C' is the number of channels. Use a fully connected layer to map the features to a low-dimensional space and generate a scene recognition vector SRV.

[0164] The input end of the frequency mapping layer FML is connected to the output end of the dynamic path selection module DPS, the input end of the output layer is connected to the output end of the frequency mapping layer FML, and the output layer includes a fully connected layer, a first deconvolution layer (DeconvolutionLayer 1), a second deconvolution layer (Deconvolution Layer 2), a third deconvolution layer (Deconvolution Layer3) and a nonlinear activation layer (Tanh or Sigmoid), such as Figure 6 The task of the output layer is to deconvolve this feature vector into a time-frequency graph representing the target frequency. This operation is similar to the upsampling process in the generative network. Through a series of deconvolution operations (also called transposed convolution), the low-dimensional feature vector is restored to a high-dimensional time-frequency graph.

[0165] This system uses various modules and levels to obtain accurate characteristic parameter models through network-based training of measured source data, and then realizes the migration of channel characteristic parameters, which greatly saves the consumption of field measurement resources and improves the efficiency of channel measurement and simulation reproduction.

[0166] In a preferred embodiment of the present invention, Figure 4 As shown in the figure, the dynamic path selection module DPS includes a lightweight path and a complex path. The lightweight path uses fewer parameters, is suitable for fast calculation, and can effectively process low-complexity inputs. The complex path is deeper, can handle complex scenes, has stronger feature extraction capabilities, and is especially suitable for highly dynamic environments.

[0167] The lightweight path includes 2 convolutional layers, 2 pooling layers, 1 fully connected layer, and 1 regularization layer:

[0168] The first convolutional layer (Conv1) is used to initially extract features, and the spatial dimension is consistent with the input;

[0169] The first pooling layer (MaxPooling1) is used to reduce the spatial dimension and reduce the amount of calculation;

[0170] The second convolutional layer (Conv2) is used to extract more advanced features;

[0171] The second pooling layer (MaxPooling2) is used to reduce the spatial dimension;

[0172] Fully connected layer, used to output the final feature vector K L ;

[0173] Regularization layer (Dropout) to prevent overfitting;

[0174] The complex path includes 4 convolutional layers, 2 residual blocks, 2 pooling layers, and 1 fully connected layer:

[0175] Convolutional layer 1 (Conv1): preserves spatial dimensions;

[0176] Residual Block 1: Contains two convolutional layers: Convolutional Layer 2 (Conv2) and Convolutional Layer 3 (Conv3). Skip connections are introduced to prevent gradient disappearance and maintain information transfer.

[0177] Pooling layer 1 (MaxPooling1): downsample the output feature map to reduce the spatial dimension;

[0178] Convolutional layer 4 (Conv4): further extract features;

[0179] Residual Block 2: Two convolutional layers: Conv5 and Conv6, extracting complex features while keeping the gradient stable;

[0180] Pooling layer 2 (MaxPooling2): further downsampling;

[0181] Convolutional layer seven (Conv7): continue to extract features;

[0182] Fully Connected Layer: Maps high-dimensional features into low-dimensional vectors and outputs the final feature vector K D ;

[0183] Regularization layer (Dropout): reduces the risk of overfitting;

[0184] The output ends of the lightweight path and the complex path are both connected to the gating unit, which is connected to the scene recognition module and the frequency mapping layer FML respectively.

[0185] The lightweight path uses fewer parameters, is suitable for fast calculations, and can effectively handle low-complexity inputs. The complex path is deeper, can handle complex scenes, has stronger feature extraction capabilities, and is especially suitable for highly dynamic environments.

[0186] In a preferred embodiment of the present invention, Figure 5As shown, the frequency mapping layer FML includes an FCV generation module, a splicing module, a first fully connected layer, a nonlinear activation layer, and a second fully connected layer connected in sequence. The FCV generation module is used to generate a vector from the data in the target domain (i.e., the target frequency band) through a feature extraction network, representing the global feature information of the target frequency.

[0187] The concatenation module is used to concatenate the source frequency feature vector [B, F] and FCV. The first fully connected layer is used to perform a linear transformation on the concatenated feature vector to generate a higher-dimensional feature representation:

[0188] First fully connected layer parameters: Linear(in_features = K + D f ,out_features=8k)

[0189] Input: concatenated feature vector [B, F+D f ]

[0190] Output: Hidden layer feature vector [B,8k].

[0191] The nonlinear activation layer uses the ReLU activation function to perform nonlinear transformation on the features output by the first fully connected layer. The second fully connected layer maps the activated features to the target frequency feature space, and finally generates the target frequency features:

[0192] Second fully connected layer parameters: Linear(in_features=8k,out_features=K')

[0193] Input: Activated feature vector [B,8k].

[0194] Output: Target frequency feature vector [B, F'], where F' is the feature dimension at the target frequency.

[0195] The frequency mapping layer is the core module in the entire network. This layer receives the feature vector from the dynamic path selection module and then maps the source frequency channel feature to the target frequency feature. In order to achieve the conversion between the source frequency and the target frequency, this module combines the source frequency feature vector and the frequency control vector (FCV) from the target domain by concatenation, and then generates the target frequency feature through nonlinear transformation.

[0196] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0197] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for migrating characteristic parameters of a cross-frequency channel, characterized in that: The steps include: The original channel signal is collected and preprocessed to obtain three channel characteristics, namely power delay spectrum PDP, root mean square delay spread RMS and Rice K factor K. factor ; The channel time-frequency response is regarded as an image, and the time-frequency image is input into the scene recognition module to extract image features, and combined with the three channel features to generate a scene recognition vector SRV; According to the scene recognition vector SRV and the gradient information returned by the path, the weights of the lightweight path and the complex path are determined, the feature representation is optimized, and the source frequency feature vector is obtained; The source frequency feature vector and the frequency control vector from the target domain are combined, and then subjected to nonlinear transformation to generate a target frequency feature; Deconvolution is performed on the target frequency features to restore the low-dimensional feature vector to a high-dimensional time-frequency graph.

2. The method for migrating cross-frequency channel characteristic parameters according to claim 1, characterized in that: The method of collecting the original channel signal and preprocessing it to obtain the three channel characteristics is as follows: Clean and slice the original channel signal; When extracting channel characteristics, the power delay profile PDP is obtained by analyzing the channel impulse response, and the root mean square delay spread τ of each sample is calculated. rms and Rice factor K factor : Among them, τ k ,a k ,p(τ k ) represent the channel delay, amplitude and power of the kth path respectively, P Los In the case of a direct line of sight, it represents the power of the direct line of sight path; in the case of a scene without a direct line of sight path, it represents the power of the strongest multipath component. mpc Represents the power of the remaining multipath components; is the average relative delay, defined as the first-order moment of PDP, which represents the average propagation delay of the multipath component relative to the first path:

3. The cross-frequency channel characteristic parameter migration method according to claim 1, characterized in that: The steps of inputting the time-frequency image into the scene recognition module, extracting image features, and generating the scene recognition vector SRV are as follows: The format of the source channel time-frequency response image is [B, C, H, W], where B represents the batch, C represents the number of channels, H represents the length, and W represents the width; The input time-frequency image is divided into blocks of size nxn to generate a local block set B S ; Perform convolution operation on each local block, with the convolution kernel size of nxn and the step size consistent with the convolution kernel size to ensure no overlapping convolution; Based on the feature map within the local block, the interaction matrix between local blocks is constructed, and the interaction feature map is obtained by weighted aggregation by calculating the similarity between the local block and its adjacent blocks; The interaction feature map [B, C', H', W'] is compressed to [B, C'] through global average pooling, where C' represents the number of channels of the interaction feature map, H' represents the length of the interaction feature map, and W' represents the width of the interaction feature map. The fully connected layer is used to map the features to a low-dimensional space to obtain Img, which is combined with the extracted three channel features to generate a scene recognition vector SRV: SRV=[Img,PDP,RMS,K factor ]。 4. The method for migrating characteristic parameters of a cross-frequency channel according to claim 1, characterized in that: According to the scene recognition vector SRV and the gradient information returned by the path, the weights of the lightweight path and the complex path are determined, the feature representation is optimized, and the source frequency feature vector is obtained. The specific method is as follows: According to SRV, the outputs of the lightweight path and the complex path are weighted summed using the gating unit to generate the final feature representation: F=αK L +(1-α)K D , Among them, K L The feature vector representing the output of the lightweight path, K D A feature vector representing the output of a complex path; SRV is the scene recognition vector; gradient information and It is calculated by the network during the forward propagation and back propagation processes, corresponding to the gradients of simple paths and complex paths respectively; the gradient information and SRV are not obtained at the same time. The gradient information that passes through the gating unit each time is obtained in the previous round of propagation. Finally, the Gating function dynamically generates the weight α based on this information to optimize the feature representation.

5. The cross-frequency channel characteristic parameter migration method according to claim 1, characterized in that: The source frequency feature vector and the frequency control vector from the target domain are combined, and then the target frequency feature is generated through nonlinear transformation. The specific steps are as follows: Input the time-frequency diagram of the target domain [B, C', H', W'], which represents the channel characteristics at the target frequency; Use the convolution layer to extract the characteristics of the target frequency, ensuring that multi-scale and multi-dimensional frequency domain features are extracted from the time-frequency diagram; The convolution layer parameters are the number of convolution kernels k, the size of the convolution kernel n*n, and the input target frequency time-frequency map [B, C', H', W'] to obtain the output convolution feature map [B, k, H', W']; Perform global pooling on the convolutional features to generate global features of the target domain, that is, input the convolutional features [B, k, H', W'], and output the target domain global feature vector [B, k], where B represents the batch, C' represents the number of channels of the interaction feature map, H' represents the length of the interaction feature map, and W' represents the width of the interaction feature map; The global feature vector of the target domain is mapped to FCV through a fully connected layer, that is, the input is the global feature vector [B, k], and the output is FCV [B, D f ], where D f It is the dimension of FCV, which serves as a guide to the target frequency; Concatenate the source frequency feature vector [B, F] and FCV to obtain the concatenated feature vector [B, F+D f ]; Perform linear transformation on the concatenated feature vectors to generate a higher-dimensional feature vector [B, 8k]; The feature vector [B,8k] is nonlinearly transformed using the ReLU activation function and then mapped to the target frequency feature space to finally generate the target frequency feature [B,F'], where F' is the feature dimension at the target frequency.

6. The method for migrating characteristic parameters of a cross-frequency channel according to claim 1, characterized in that: Perform deconvolution on the target frequency feature to restore the low-dimensional feature vector to a high-dimensional time-frequency graph. The specific steps are as follows: The one-dimensional target frequency feature [B, F'] is converted into a four-dimensional tensor and upsampled multiple times to restore the target frequency feature to the same size as the original time-frequency graph; Sigmoid activation is used for the target frequency feature to limit the output value range to [0,1] to obtain a high-dimensional time-frequency diagram.

7. The cross-frequency channel characteristic parameter migration method according to claim 1, characterized in that: Introduce the loss function: Frequency consistency loss Measure the difference between the generated colored time-frequency map and the true target frequency map to ensure that the generated time-frequency map meets the target frequency characteristics: Among them, N represents the batch size, that is, the number of samples input into the network during one training process, and x i is the source frequency image; FCV is the feature representation of the target extraction; G(x i ,FCV): Color time-frequency diagram generated by the frequency mapping layer; i : Colored time-frequency diagram of the true target frequency; |·|2 is the L2 norm; Introducing L2 regularization reg , limiting the size of the model parameters and encouraging the model to find smoother and simpler solutions: Among them, λ: regularization coefficient, which controls the strength of the regularization term and usually needs to be tuned through experiments; θ: the set of trainable parameters in the model; w: each trainable parameter; The total loss function L total Combining frequency map consistency loss and regularization term: L total =L freq +L reg。 8. A system for migrating characteristic parameters of a cross-frequency channel based on the method of any one of claims 1 to 7, characterized in that: It includes input layer, scene recognition module SRM, dynamic path selection module DPS, frequency mapping layer FML and output layer; The input layer is used to collect the original channel signal and perform preprocessing to obtain a channel characteristic data set; The scene recognition module SRM and the dynamic path selection module DPS are both connected to the input layer, and the scene recognition module SRM includes a local feature interaction layer, an intra-block feature extraction layer, an interactive feature aggregation layer, and a feature compression and scene representation layer connected in sequence; The input end of the frequency mapping layer FML is connected to the output end of the dynamic path selection module DPS, the input end of the output layer is connected to the output end of the frequency mapping layer FML, and the output layer includes a fully connected layer, a first deconvolution layer, a second deconvolution layer, a third deconvolution layer and a nonlinear activation layer.

9. The cross-frequency channel characteristic parameter migration system according to claim 8, characterized in that: The dynamic path selection module DPS includes a lightweight path and a complex path; The lightweight path includes 2 convolutional layers, 2 pooling layers, 1 fully connected layer, and 1 regularization layer: The first convolutional layer is used to initially extract features, and the spatial dimension is consistent with the input; The first pooling layer is used to reduce the spatial dimension and reduce the amount of calculation; The second convolutional layer is used to extract higher-level features; The second pooling layer is used to reduce the spatial dimension; Fully connected layer, used to output the final feature vector K L ; Regularization layer Dropout is used to prevent overfitting; The complex path includes 4 convolutional layers, 2 residual blocks, 2 pooling layers, and 1 fully connected layer: Convolutional layer 1: preserves spatial dimensions; Residual block 1: contains two convolutional layers: convolutional layer 2 and convolutional layer 3. Skip connections are introduced to prevent gradient disappearance and maintain information transmission. Pooling layer 1: downsample the output feature map to reduce the spatial dimension; Convolutional layer 4: further extract features; Residual block 2: Two convolutional layers: convolutional layer 5 and convolutional layer 6, extracting complex features while keeping the gradient stable; Pooling layer 2: further downsampling; Convolutional layer seven: continue to extract features; Fully connected layer: maps high-dimensional features into low-dimensional vectors and outputs the final feature vector K D ; Regularization layer: reduce the risk of overfitting; The output ends of the lightweight path and the complex path are both connected to the gating unit, and the gating unit is connected to the scene recognition module and the frequency mapping layer FML respectively.

10. The cross-frequency channel characteristic parameter migration system according to claim 8, characterized in that: The frequency mapping layer FML includes an FCV generation module, a splicing module, a first fully connected layer, a nonlinear activation layer, and a second fully connected layer connected in sequence; the FCV generation module is used to generate a vector from the data in the target domain through a feature extraction network, representing the global feature information of the target frequency; The concatenation module is used to concatenate the source frequency feature vector [B, F] and the FCV, and the first fully connected layer is used to perform a linear transformation on the concatenated feature vector to generate a higher-dimensional feature representation; The nonlinear activation layer uses the ReLU activation function to perform a nonlinear transformation on the features output by the first fully connected layer, and the second fully connected layer maps the activated features to the target frequency feature space, and finally generates the target frequency features.

Images