A 6G Wireless Channel Feature Extraction Method Based on Dimensionality Reduction Complex Convolutional Network

Abstract

The present invention discloses a method for extracting the wireless channel characteristics of the sixth-generation (6G) mobile communication based on a dimensionality-reduced complex convolutional network. Aiming at the analysis and modeling of the wireless channel characteristics of 6G mobile communication, it solves the problem that it is difficult to achieve a compromise between accuracy and complexity in the existing wireless channel characteristic extraction. The implementation steps are as follows: 1) Use ray tracing simulation software to obtain channel simulation data under multiple typical scenarios; 2) Cluster the multipaths and calculate the intra-cluster / inter-cluster delay / angle spread statistics; 3) Use a stacked autoencoder to perform interference-free unsupervised data dimensionality reduction on the large-dimensional simulation data; 4) Propose a complex cross convolutional neural network to extract relevant statistics; 5) Use measured data for incremental learning to enhance the robustness of the constructed network; 6) Compare the complexity and accuracy performance of the proposed method with other networks. The wireless channel characteristic extraction method provided by the present invention has the characteristic of low complexity and can provide an effective solution for the processing of time-varying and large-dimensional measurement data.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communication and channel modeling, and particularly relates to a method for extracting the characteristics of a sixth generation (6G) wireless channel based on a dimensionality reduction complex convolutional network. Background Art

[0002] Driven by the diverse application requirements of users, the 6G wireless communication system will take "full coverage, full spectrum, full application, full senses, full digital, and strong security" as the ideal vision, and is committed to realizing the leap from the Internet of Everything to the Internet of Intelligent Things. However, the more complex and diverse communication scenarios have brought huge challenges to the design and optimal deployment of the 6G wireless communication system. As an important prerequisite for network optimization and network planning, wireless channel measurement, characteristic analysis, and modeling need to cope with larger data volumes, more complex data processing, higher accuracy requirements, and higher universality requirements under the impact of 6G. Among them, wireless channel characteristic analysis, as a key step connecting the physical and digital worlds, is of great significance for global performance optimization. However, with the sharp increase in the amount of measurement data, traditional methods, including high-resolution multipath parameter extraction, multipath clustering, correlation statistic calculation, etc., are facing increasing complexity and it is difficult to achieve a better compromise between complexity and accuracy.

[0003] The leapfrog development of a new generation of network information technologies such as artificial intelligence has promoted the development of mobile communication technologies towards the trend of intelligence, bringing all-round breakthroughs in storage, computing, and perception. As the core technology of artificial intelligence, machine learning has been widely used in the research of problems in the field of wireless communication to achieve ultra-high-efficiency simulation, decision-making, and prediction functions. Among them, deep learning simulates the working mechanism of the human brain, can better process massive data, and solves problems in a centralized manner to obtain the required answers. At present, some researchers have applied machine learning algorithms to optimize the extraction of wireless channel statistical characteristics. For example, the research group at Southeast University used a feedforward neural network and a neural network based on a radial basis function neural network to construct a big data-enabled channel model. By obtaining measurement data in an indoor office scenario and simulation data based on a geometry based stochastic channel model (GBSM), taking the transceiver coordinates and positions, and carrier frequency as the neural network inputs, and taking the received power, root mean square delay spread, and angular spread as the outputs, it was found after training and testing that the machine learning method can provide an effective analysis tool for future wireless channel modeling. The research group at Tsinghua University used a convolutional neural network to distinguish different wireless channels, taking the multipath parameters extracted by the high-resolution space-alternating generalized expectation-maximization (SAGE) algorithm as the input and the wireless channel category as the output.

[0004] By researching the existing work on wireless channel characteristic analysis based on machine learning / deep learning methods, it is found that generally shallow / deep networks are used to analyze a few characteristics, and there is still a lack of discussion on the problem of large-dimensional input, the extraction of multi-path cluster-related characteristics, the lack of global optimization of the network, and the robustness and universality of the constructed network are not considered. To address the above problems, the present invention is committed to developing a 6G adaptive wireless channel characteristic extraction method based on a cross-convolutional neural network (X-CNN) of the real and imaginary parts of complex numbers with a denoising stacked autoencoder, solving the problem that it is difficult to achieve a compromise between accuracy and complexity in existing solutions and improving the universality of the network. Summary of the Invention

[0005] Technical Problem:

[0006] To solve the problems of complex number input of large-dimensional data, joint extraction of multiple characteristics, and the robustness of the network, and further solve the problem that it is difficult to compromise the accuracy, complexity, and universality in channel characteristic extraction in 6G full-band, full-coverage, and full-application scenarios, the present invention provides a 6G wireless channel characteristic extraction method based on a dimensionality reduction complex convolutional network.

[0007] Technical Solution:

[0008] To achieve the above objectives, the present invention provides a 6G wireless channel characteristic extraction method based on a complex X-CNN with a denoising stacked autoencoder, including the following specific steps:

[0009] S1. Construct a twin scenario of the real communication environment, and use ray tracing (RT) software to obtain wireless channel simulation data in multiple typical scenarios;

[0010] S2. Calculate relevant statistics according to the simulation multi-path parameters and the results after clustering processing, and establish a wireless channel simulation database for typical scenarios;

[0011] S3. Input training data, use a denoising stacked autoencoder to perform feature dimensionality reduction on 6G large-dimensional data, and extract the encoder feature parameters;

[0012] S4. Input the encoder feature parameters into the constructed X-CNN for training to achieve channel characteristic extraction, and use the test data set to verify the network performance;

[0013] S5. Conduct channel measurements in different scenarios, process the measured data, and perform incremental learning on the network;

[0014] S6. Compare the accuracy and complexity of the method proposed by the present invention with methods based on other network structures.

[0015] The specific steps of step S1 are as follows:

[0016] Step S101: According to the actual 6G wireless communication environment, use Wireless Insite RT simulation software to construct a geometric environment in typical indoor office, corridor, indoor-to-outdoor, outdoor urban microcell and other scenarios using a simplified three-dimensional model. Taking the indoor office scenario as an example, a quadrilateral formed by four points is used as the ground, and a building is constructed with walls based on the quadrilateral as the basic unit. Each wall is determined by the coordinates of four vertices, and the building number to which each wall belongs is given. In addition, all walls use the same material, but windows, etc. are not considered.

[0017] Step S102: Set the transceiver positions, as well as the array scale, center frequency, bandwidth, etc. Use reverse ray tracing for radio wave propagation prediction to establish a virtual source point tree starting from the receiving source. Each child node of the tree represents a virtual source point. Establish a visible surface table and a visible wedge table for the virtual sources, and search for three-dimensional ray propagation paths such as direct, reflected, and diffracted rays based on the visible wedge table.

[0018] Step S103: Extract the relevant parameters of all paths, including the complex amplitude, delay, horizontal / pitch departure / arrival angle of each path, as well as the root mean square delay spread and angle spread, etc., so as to obtain the twin data of the real scenario.

[0019] The specific steps of step S2 are as follows:

[0020] Step S201: The statistical characteristics concerned by the present invention include the received power, as well as the root mean square delay spread (RMSDS) between multipaths and clusters, and the three-dimensional horizontal / pitch arrival / departure angle spread (azimuth / elevation arrival / departure angle spread, AAS / EAS / ADS / EDS). Among them, the multipath root mean square delay / angle spread can be directly obtained from the RT simulation software. The calculation methods of the corresponding statistics are given below. Considering that the transceiver is respectively configured with NR and NT root antennas, h qp (t, τ)(p = 1,..., N T , q = 1,..., N R ) represents the channel impulse response of the q-th receiving antenna and the p-th transmitting antenna at time t with a delay of τ. Then the signal power from the p-th transmitting antenna to the q-th receiving antenna can be calculated as follows:

[0021]

[0022] where L is the number of multipaths, β lis the complex amplitude of the l-th path. The RMSDS caused by different travel distances when reaching the receiving end can be calculated as follows:

[0023]

[0024] where τ l is the time delay of the l-th path, and μ τ is the mean time delay. For multipath RMS, it is denoted as σ τ,L ; for the subsequent inter-cluster delay spread, it is denoted as σ τ,C . Similarly, the RMS angular spread can be calculated as follows:

[0025]

[0026] where μ θ is the mean angle. It should be noted that θ l is the angle of the l-th path, which can be used to calculate the AAS, EAS, ADS, and EDS of multipath, denoted as respectively, and the AAS, EAS, ADS, and EDS between clusters, denoted as and

[0027] Step S202: For the inter-cluster statistical characteristics concerned by the present invention, it is necessary to first use the Mini Batch k-means clustering algorithm for clustering. Here, the clustering algorithm jointly considers the time delay, horizontal and pitch angle parameters of multipath, and the specific operations are as follows:

[0028] (1) Randomly select the parameter sets of some paths, and use the k-power-means algorithm for mini-batch clustering, iteratively update and calculate the Euclidean distance to obtain the initial centroids;

[0029] (2) Randomly select the parameter sets of other parts of the paths, and assign them to the nearest centroid by calculating the Euclidean distance from the centroids;

[0030] (3) Update the centroids according to the existing parameter sets;

[0031] (4) Update steps (2) and (3) until the change in the centroids is less than the set threshold or the maximum number of iterations.

[0032] Step S203: According to the clustering results of step S202, refer to the calculation method of step S201 to obtain the inter-cluster delay spread σ τ,C and the angular spread and

[0033] Step S204: Organize the RT simulation data and related statistics, and establish a simulation database for the next training and testing.

[0034] The specific steps of step S3 are as follows:

[0035] Step S301: To solve the problem of excessive RT simulation data volume, the present invention first uses a denoising stacked autoencoder for feature dimensionality reduction, and the principle is as Figure 2 shown. First, randomly generate interference samples contaminated by complex Gaussian distribution noise from the training dataset y where y is a set of channel impulse response data obtained by simulation in typical indoor office, corridor, indoor-to-outdoor and other scenarios.

[0036] Step S302: Since the training data is complex, the real part and the imaginary part of the data are respectively extracted and stacked into column vectors. Common activation functions include sigmoid, tanh, ReLU, maxout, etc. After passing through a fully connected network, the sigmoid activation function is used,

[0037]

[0038] and after normalization, it is input into the encoder e(·), where x is the input of each layer of the network. Here, the encoder includes an input layer and multiple hidden layers. Among them, the softplus is used as the activation function for the hidden layer, and the function expression is

[0039] f2(x) = ln(1 + e x ).

[0040] Step S303: Take the output of the last hidden layer of the encoder as the input d(·) of the decoder. Then, the goal of this denoising autoencoder is to train the network to obtain the network parameter values that minimize the cost function

[0041]

[0042] where the network optimization uses the gradient descent method. Next:

[0043] (1) Set the minimum loss function target threshold and the maximum number of iterations, and stop when the training reaches the target value or the maximum number of iterations; (2) Remove the decoder. At this time, the output {d n} of the encoder is the low-dimensional depth feature extracted after network processing, and its dimension is the same as the number of neurons in the last layer of the encoder, and can be used as the input of the X-CNN for the next prediction.

[0044] The specific steps of step S4 are as follows:

[0045] Step S401: Divide the established simulation database into training and test data sets according to the ratio of 70% and 30%. Extract the training data set after dimensionality reduction by the stacked autoencoder as the input of the X-CNN. Still, first perform parallel training according to the real part and the imaginary part, as Figure 3 shown. Use the z-score normalization method to perform data normalization preprocessing, eliminate the dimensionality influence between evaluation indicators, and use the corresponding characteristic parameters as labels.

[0046] Step S402: Since the deepening of the number of layers will cause local optimal solutions and gradient disappearance, consider constructing a CNN. The present invention adopts a hierarchical structure of convolution-pooling-convolution-pooling-convolution-pooling-long short-term memory (LSTM), which can better fit features, reduce information loss, and retain the correlation between data while reducing parameters and dimensionality. (1) Set the number, size, and stride of the convolution kernels of the convolutional layer to obtain the next-layer feature map.

[0047] Suppose the (u - 1)-th layer has convolution kernels of size M u-1 ×M u-1 Then the output of the feature map of the u-th layer can be expressed as

[0048]

[0049] where is the coefficient of the i,j-th neuron in the u-th layer, and b u is the bias of the u-th layer. g(·) is the activation function, which is used to map the input of each neural network to the next layer or the output layer.

[0050] (2) Since the dimension of the input data is large, when setting a small receptive field, the feature map after passing through the convolutional layer is still large. The present invention uses multiple pooling layers in this network to further reduce the dimensionality of the feature map.

[0051] (3) After separately processing the real part and the imaginary part of the complex number through the convolutional layer and the pooling layer, the features are merged and fitted through the fully connected layer.

[0052] (4) Finally, use the cascaded LSTM input gate and forget gate to extract the correlation between data.

[0053] (5) Define the training error using the cross-entropy function:

[0054]

[0055] where, Θ m is the statistic calculated in Step S2, is the value of the characteristic parameter output by the network.

[0056] (6) The optimization extreme value solution of the training error can be completed by using general gradient descent or Newton's method. When L does not reach the set threshold value or reaches the set number of iterations, the weights and biases of the neurons are changed at a certain learning rate.

[0057] (7) Select a suitable learning algorithm, such as the Levenberg–Marquardt algorithm (LM). The training process of the neural network is the continuous update of each weight and bias.

[0058] (8) After the learning process ends, substitute the verification test data into the network and observe the test performance of the network.

[0059] The specific steps of step S5 are as follows:

[0060] S501. Obtain the channel measurement data in the real scenario;

[0061] S502. Obtain the inter-cluster delay and angle spread parameter set in the same way as processing the RT data in step S2;

[0062] S503. Based on the DNN trained in step S4, use the obtained measurement data and the processed characteristic parameter set for incremental learning to enhance the plasticity and stability of the network for the measured data in different scenarios. The joint training algorithm is used for network training. Input the new measured data into the network trained in step S4, and retrain the model on the basis of the original network to obtain the joint optimal prediction for the new and old data.

[0063] The specific steps of step S6 are as follows:

[0064] S601. Use the RT simulation data to train networks such as the radial basis network and LSTM;

[0065] S602. Take the measured data as the input, respectively obtain the characteristic parameter estimation using the radial basis network and the LSTM network, calculate the root mean square error, and record the processing time;

[0066] S603. Use the denoising autoencoder complex X-CNN proposed by the present invention to predict the measured data, calculate the root mean square error, and record the processing time;

[0067] S604. Compare the accuracy and complexity of the denoising autoencoder X-CNN proposed by the present invention with the other two networks.

[0068] Beneficial effects:

[0069] Compared with the closest existing method, the beneficial effects of the technical solution provided by the present invention are:

[0070] Based on the generation of a large amount of RT simulation and measured data, the present invention proposes a method for extracting 6G wireless channel characteristics based on a denoising stacked autoencoder complex X-CNN. Compared with traditional methods and other neural networks, the present invention can well solve the problems of large-dimensional input of channel data and extraction of characteristics related to multipath clusters, providing lower complexity and better robustness while ensuring higher accuracy, and can provide an effective solution for channel modeling in 6G full-frequency band scenarios, full-coverage scenarios, and full-application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a schematic diagram of the process of the present invention;

[0072] Figure 2 It is a schematic diagram of the structure of the denoising autoencoder of the present invention;

[0073] Figure 3 It is a schematic diagram of the structure of the X-CNN of the present invention;

[0074] Figure 4 It is a schematic diagram of the process of the X-CNN window and side length of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0075] The present invention will be described in detail below with reference to the drawings and specific embodiments. This Embodiment 1 is implemented on the premise of the technical solution of the present invention, and gives a detailed implementation plan and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0076] Illustrate according to the content included in the claims

[0077] Embodiment 1:

[0078] The specific steps of step S1 are as follows:

[0079] Step S101, in a typical scenario, use the Wireless Insite RT simulation software to construct a geometric simulation environment of a real scenario. Adopt a simplified three-dimensional model to construct geometric environments in typical indoor office, corridor, indoor to outdoor, outdoor urban microcell and other scenarios. Taking the indoor office scenario as an example, a quadrilateral formed by four points constitutes the ground, and a building is constructed with the quadrilateral as the basic unit of the wall surface. Each wall surface is determined by four vertex coordinates, and the building number to which each wall surface belongs is given. In addition, all wall surfaces use the same material, but windows, etc. are not considered.

[0080] Step S102: Set the positions of the transmitter and receiver. Taking the transmitter as a 64 - element uniform linear array and the receiver as a 4 - element uniform linear array as an example, set the center frequency to 5.3 GHz, the bandwidth to 160 MHz, use reverse ray - tracing radio wave propagation prediction, establish a virtual source point tree starting from the receiving source, each sub - node of the tree represents a virtual source point, establish the visible surface table and visible wedge table of the virtual source, and search for three - dimensional ray propagation paths such as direct, reflected, and diffracted rays based on the visible surface wedge table.

[0081] Step S103: Extract the relevant parameters of all paths, including the complex amplitude, time delay, horizontal / pitch departure / arrival angles of each path, as well as the root - mean - square delay spread and angle spread, etc., so as to obtain the twin data of the real - world scenario.

[0082] The specific steps of step S2 are as follows:

[0083] Step S201: The statistical characteristics concerned in the present invention include received power, and the root - mean - square delay spread (RMSDS) between multipaths and clusters, and the three - dimensional horizontal / pitch arrival / departure angle spread (azimuth / elevation arrival / departure angle spread, AAS / EAS / ADS / EDS). Among them, the multipath root - mean - square delay / angle spread can be directly obtained from the RT simulation software. The calculation methods of the corresponding statistics are given below. Considering that the transmitter and receiver are respectively configured with NR and NT root antennas, h qp (t, τ)(p = 1,..., N T , q = 1,..., N R ) represents the channel impulse response of the q - th receiving antenna and the p - th transmitting antenna at time t with time delay τ. Then, the signal power magnitude from the p - th transmitting antenna to the q - th receiving antenna can be calculated as follows:

[0084]

[0085] where L is the number of multipaths, and β l is the complex amplitude of the l - th path. The RMS DS caused by different path lengths when reaching the receiver can be calculated as follows:

[0086]

[0087] where τ l is the time delay of the l - th path, and μ τ is the mean time delay. For the multipath RMS, we denote it as σ τ,L , and for the subsequent inter - cluster delay spread, we denote it as σ τ,C . Similarly, the RMS angle spread can be calculated as follows:

[0088]

[0089] where μ θ is the average angle. It should be noted that θ l is the angle of the l-th path, which can be used to calculate the AAS, EAS, ADS, and EDS of multipath, denoted as and the AAS, EAS, ADS, and EDS between clusters, denoted as and

[0090] Step S202: For the inter-cluster statistical characteristics concerned by the present invention, the Mini Batch k-means clustering algorithm needs to be used for clustering first. Here, the clustering algorithm jointly considers the delay, horizontal and pitch angle parameters of multipath. The specific operations are as follows:

[0091] (1) Randomly select the parameter sets of some paths, and use the k-power-means algorithm for mini-batch clustering, iteratively update and calculate the Euclidean distance to obtain the initial centroid;

[0092] (2) Randomly select the parameter sets of other parts of the paths, and assign them to the nearest centroid by calculating the Euclidean distance from the centroid;

[0093] (3) Update the centroid according to the existing parameter sets;

[0094] (4) Update (2) and (3) until the change of the centroid is less than the set threshold or the maximum number of iterations.

[0095] Step S203: According to the clustering results of Step S202, refer to the calculation method of Step S201 to obtain the inter-cluster delay spread σ τ,C and the angle spread and

[0096] Step S204: Organize the RT simulation data and related statistics, and establish a simulation database for the next training and testing.

[0097] The specific steps of Step S3 are as follows:

[0098] Step S301: To solve the problem of excessive RT simulation data volume, the present invention first uses a denoising stacked autoencoder for feature dimensionality reduction. The principle is as Figure 2 shown. First, randomly generate interference samples contaminated by complex Gaussian distributed noise from the training dataset y where y is a set of channel impulse response data obtained by simulation in typical indoor office, corridor, indoor-to-outdoor and other scenarios.

[0099] Step S302: Since the training data is complex, the real part and the imaginary part of the data are respectively extracted and stacked into column vectors. Common activation functions include sigmoid, tanh, ReLU, maxout, etc. After passing through the fully connected network, the sigmoid activation function is used, and after normalization, it is input into the encoder e(·), where x is the input of each layer of the network. Here, the encoder includes an input layer and multiple hidden layers. Among them, the softplus is used as the activation function for the hidden layer, and the function expression is

[0100]

[0101] f2(x) = ln(1 + e

[0102] x x ).

[0103] Step S303: The output of the last hidden layer of the encoder is used as the input d(·) of the decoder. Then, the goal of this denoising autoencoder is to train the network to obtain the network parameter values that minimize the cost function

[0104]

[0105] where the network optimization adopts the gradient descent method. Next:

[0106] (1) Set the minimum loss function target threshold and the maximum number of iterations, and stop when the training reaches the target value or the maximum number of iterations; (2) Remove the decoder. At this time, the output {d n} of the encoder is the low-dimensional depth feature extracted after passing through the network, and its dimension is the same as the number of neurons in the last layer of the encoder, which can be used as the input of the X-CNN for the next prediction.

[0107] The specific steps of step S4 are as follows:

[0108] Step S401: Divide the established simulation database into training and test data sets according to the ratio of 70% and 30%. Extract the training data set after dimensionality reduction by the stacked autoencoder as the input of the X-CNN. Still, parallel training is first performed according to the real part and the imaginary part, as Figure 3 shown. The z-score normalization method is used for data normalization preprocessing to eliminate the dimensionality influence between evaluation indicators, and the corresponding characteristic parameters are used as labels.

[0109] Step S402: Since local optimal solutions and gradient disappearance may occur as the number of layers deepens, consider constructing a CNN. The present invention adopts a hierarchical structure of convolution-pooling-convolution-pooling-convolution-pooling-long short-term memory (LSTM), which can better fit features, reduce information loss, and retain the correlation between data while reducing parameters and dimensionality.

[0110] (1) Set the number and size of convolutional kernels and the stride of the convolutional layer to obtain the next-layer feature map.

[0111] Assume that the (u - 1)-th layer has a convolutional kernel of size M u-1 ×M u-1 Then the output of the u-th layer feature map can be expressed as

[0112]

[0113] where is the coefficient of the i,j-th neuron in the u-th layer, and b u is the bias of the u-th layer. g(·) is the activation function, which is used to map the input of each neural network to the next layer or the output layer.

[0114] (2) Since the dimension of the input data is large, when setting a small receptive field, the feature map after passing through the convolutional layer is still large. The present invention uses multiple pooling layers in this network to further reduce the dimensionality of the feature map.

[0115] (3) After separately processing the real and imaginary parts of the complex number through the convolutional layer and the pooling layer, the features are merged and fitted through the fully connected layer.

[0116] (4) Finally, the concatenated LSTM input gate and forget gate are used to extract the correlation between data.

[0117] (5) The training error is defined using the cross-entropy function:

[0118]

[0119] where Θ m is the statistic calculated in step S2, is the characteristic parameter value output by the network.

[0120] (6) The optimization extreme value solution of the training error can be completed using the general gradient descent or Newton's method. When L does not reach the set threshold value or reaches the set number of iterations, the weights and biases of the neurons are changed at a certain learning rate.

[0121] (7) Select a suitable learning algorithm, such as the Levenberg–Marquardt algorithm (LM). The training process of the neural network is the continuous update of each weight and bias.

[0122] (8) After the learning process ends, substitute the verification test data into the network and observe the test performance of the network.

[0123] The specific steps of step S5 are as follows:

[0124] S501. Obtain channel measurement data in the real scenario;

[0125] S502. Obtain the inter-cluster delay and angle spread parameter set in the same way as processing RT data in step S2;

[0126] S503. Based on the DNN trained in step S4, use the obtained measurement data and the processed characteristic parameter set for incremental learning to enhance the plasticity and stability of the network for measured data in different scenarios. The joint training algorithm is used for network training. Input new measured data into the network trained in step S4, and retrain the model on the basis of the original network to obtain the joint optimal prediction for the new and old data.

[0127] The specific steps of step S6 are as follows:

[0128] S601. Use RT simulation data to train networks such as the radial basis network and LSTM;

[0129] S602. Take the measured data as the input, respectively obtain the characteristic parameter estimation using the radial basis network and the LSTM network, calculate the root mean square error, and record the processing time;

[0130] S603. Use the denoising autoencoder complex X-CNN proposed by the present invention to predict the measured data, calculate the root mean square error, and record the processing time;

[0131] S604. Compare the accuracy and complexity of the denoising autoencoder X-CNN proposed by the present invention with the other two networks.

[0132] Based on the generation of a large amount of RT simulation and measured data, the present invention proposes a method for extracting 6G wireless channel characteristics based on the denoising stacked autoencoder complex X-CNN. Compared with traditional methods and other neural networks, the present invention can well solve the problems of large-dimensional input of channel data and extraction of characteristics related to multipath clusters, provide lower complexity and better robustness while ensuring high accuracy, and can provide an effective solution for channel modeling in 6G full-frequency band scenarios, full-coverage scenarios, and full-application scenarios.

[0133] It can be understood that the present invention is described by means of some embodiments. Those skilled in the art will know that, without departing from the spirit and scope of the present invention, various changes or equivalent substitutions can be made to these features and embodiments. Additionally, under the teaching of the present invention, these features and embodiments can be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the scope protected by the present invention.

Claims

1. A method for extracting 6G wireless channel characteristics based on a dimensionality-reduced complex convolutional network, characterized in that It includes the following steps: S1. Construct a twin scenario of the real communication environment and obtain wireless channel simulation data under various typical scenarios; S2. Calculate relevant statistics according to the simulation multipath parameters and the results after clustering processing, and establish a wireless channel simulation database for typical scenarios; S3. Select training data from the simulation database, input the training data, use a denoising stacked autoencoder to perform feature dimensionality reduction on 6G large-dimensional data, and extract the encoder feature parameters; S4. Input the encoder feature parameters into the constructed X-CNN for training to achieve channel characteristic extraction, and use the test data set to verify the network performance; S5. Conduct channel measurements under different scenarios, process the measured data, and perform incremental learning on the network; The specific steps of S2 include: S201. The statistics include received power, and the root mean square delay spread, horizontal angle of arrival spread, elevation angle of arrival spread, horizontal angle of departure spread, and elevation angle of departure spread between multipaths and clusters. Among them, the multipath root mean square delay / angle spread is directly obtained from the RT simulation software; The calculation methods of the corresponding statistics are given below: The transceiver is respectively configured with N R and N T antenna elements. h qp (t, τ) represents the channel impulse response of the q-th receive antenna and the p-th transmit antenna at time t with a time delay of τ, where p = 1, …, N T , q = 1, …, N R . Then, the signal power from the p-th transmit antenna arriving at the q-th receive antenna is calculated as follows: where L is the number of multipaths, and β l is the complex amplitude of the l-th path; the root mean square delay spread caused by different distances when reaching the receiving end is calculated as follows: where τ l is the delay of the l-th path, and μ τ is the mean delay; for multipath RMS, denoted as σ τ,L , and for inter-cluster delay spread, denoted as σ τ,C ; the RMS angular spread is calculated as follows: where μ θ is the average angle; θ l is the angle of the l-th path, which is used to calculate the horizontal arrival angle spread, elevation arrival angle spread, horizontal departure angle spread, and elevation departure angle spread of the multipath, denoted as respectively, as well as the horizontal arrival angle spread, elevation arrival angle spread, horizontal departure angle spread, and elevation departure angle spread between clusters, denoted as and S202. For the inter-cluster statistics, use the Mini Batch k-means clustering algorithm for clustering; S203. Obtain the inter-cluster delay spread σ τ,C and the angular spread and S204. Organize the RT simulation data and related statistics to establish a simulation database; The specific steps of S3 include: S301. Feature dimension reduction using a denoising stacked autoencoder: First, randomly generate interference samples contaminated by complex Gaussian distribution noise from the training data set y where y is a set of channel impulse response data obtained by simulation in typical indoor office, corridor, and indoor-to-outdoor scenarios; S302. Extract the real part and the imaginary part of the data respectively and stack them into column vectors; after passing through the fully connected network, use the sigmoid activation function, and after normalization, input them into the encoder e(·), where x is the input of each layer of the network; the encoder includes an input layer and multiple hidden layers; among them, the softplus is used as the activation function for the hidden layer; ​ S303. Take the output of the last hidden layer of the encoder as the input d(·) of the decoder. Then the goal of this denoising autoencoder is to train the network to obtain the network parameter values that minimize the cost function where the network optimization uses the gradient descent method; S304. Set the minimum loss function target threshold and the maximum number of iterations, and stop when the training reaches the target value or the maximum number of iterations; remove the decoder. At this time, the output {d n} of the encoder is the low-dimensional depth feature extracted after being processed by the network. Its dimension is the same as the number of neurons in the last layer of the encoder and is used as the input of X-CNN for the next prediction; The specific steps of S4 include: S401. Divide the established simulation database into training and test data sets; Extract the training data set after dimensionality reduction by the stacked autoencoder as the input of the X-CNN. Still, first, perform parallel training according to the real and imaginary parts; Use the z-score normalization method for data normalization preprocessing to eliminate the dimensionality influence between evaluation indicators, and use the corresponding characteristic parameters as labels; S402. Construct an X-CNN, adopting a hierarchical structure of convolution-pooling-convolution-pooling-convolution-pooling-long short-term memory network.

2. The 6G wireless channel characteristic extraction method based on the dimensionality reduction complex convolution network according to claim 1, characterized in that, The specific steps of S1 include: S101. According to the actual 6G wireless communication environment, use the Wireless Insite RT simulation software to construct a geometric environment under typical indoor office, corridor, indoor-to-outdoor, and outdoor urban microcell scenarios using a three-dimensional model; S102. Set the transceiver positions, as well as the array scale, center frequency, and bandwidth. Use reverse ray tracing for radio wave propagation prediction to establish a virtual source point tree starting from the receiving source. Each child node of the tree represents a virtual source point. Establish a visible surface table and a visible split table of the virtual source, and search for the three-dimensional ray propagation path based on the visible surface split table; S103. Extract the relevant parameters of all paths, including the complex amplitude, delay, horizontal / elevation departure / arrival angle of each path, as well as the root mean square delay spread and angle spread, so as to obtain the twin data of the real scenario.

3. A method for extracting 6G wireless channel characteristics based on a dimensionality reduction complex convolutional network according to claim 1, characterized in that, The specific steps of step S202 include: (1) Randomly select the parameter sets of some paths, perform mini-batch clustering using the k-power-means algorithm, iteratively update the calculation of the Euclidean distance, and obtain the initial centroids; (2) Randomly select the parameter sets of other paths, and assign them to the nearest centroid by calculating the Euclidean distance from the centroids; (3) Update the centroids according to the existing parameter sets; Update steps (2) and (3) until the change in the centroids is less than the set threshold or the maximum number of iterations.

4. A method for extracting 6G wireless channel characteristics based on a dimensionality reduction complex convolutional network according to claim 1, characterized in that, The specific steps of S402 include: (1) Set the number, size, and stride of the convolutional kernels of the convolutional layer to obtain the next-layer feature map; Assume that the convolutional kernel in the (u - 1)-th layer has a size of M u-1 ×M u-1 , then the output of the feature map in the u-th layer is expressed as where is the coefficient of the \(i,j\)-th neuron \(x\) in the \(u\)-th layer i,j , \(b\) u is the bias of the \(u\)-th layer, and \(g(\cdot)\) is the activation function used to map the input of each neural network to the next layer or the output layer; (2) Use multiple pooling layers to reduce the dimension of the feature map; (3) After separately processing the real and imaginary parts of the complex numbers through the convolutional layer and the pooling layer, merge and fit the features through the fully connected layer; (4) Finally, use the cascaded LSTM input gate and forget gate to extract the correlation between data; (5) Define the training error using the cross-entropy function: Among them, Θ m is the statistic calculated in step S2, and is the characteristic parameter value output by the network; (6) Use gradient descent or Newton's method to solve the optimization extremum of the training error; when L does not reach the set threshold value or reaches the set number of iterations, change the weights and biases of the neurons at a certain learning rate; (7) Select a suitable learning algorithm to train the neural network; (8) After the learning process ends, substitute the verification test data into the network and observe the test performance of the network.

5. The 6G wireless channel characteristic extraction method based on a dimensionality reduction complex convolutional network according to claim 1, wherein The specific steps of S5 include: S501. Obtain the channel measurement data in the real scenario; S502. Obtain the inter-cluster delay and angle spread parameter sets in the same way as processing the RT data in S2; S503. Based on the X-CNN trained in S4, use the obtained measurement data and the processed characteristic parameter sets for incremental learning to enhance the plasticity and stability of the network for the measured data in different scenarios; use the joint training algorithm for network training; input the new measured data into the network trained in S4, and retrain the model on the basis of the original network to obtain the joint optimal prediction for the new and old data.

Images