US-12621067-B2 - Predictive channel modeling method based on generative adversarial network and long short-term memory artificial neural network

US12621067B2US 12621067 B2US12621067 B2US 12621067B2US-12621067-B2

Abstract

Disclosed in the present disclosure is a predictive channel modeling method based on a generative adversarial network and a long short-term memory artificial neural network, which method effectively achieves a channel prediction function in different frequency bands and scenarios, and generates a large number of channel data sets for simulation experiments. The method comprises: firstly, inputting channel measurement data for existing frequency bands and scenarios for training; then, learning true channel data using a long short-term memory artificial neural network, and acquiring a channel time sequence feature; by means of adversarial learning of a generative adversarial network, greatly eliminating redundant information of the channel data, and on the basis of the measurement data, generating accurate channel data, and acquiring massive channel information; and finally, achieving the balance between a generative model and a discriminative model during the continuous iteration of the generative adversarial network, and then outputting a trained predictive channel model. A statistical channel feature obtained by means of prediction by a model can clearly specify the predictive learning for a channel distribution feature in the present disclosure, and real-time and complex prediction problems in wireless communication can be solved.

Inventors

Chengxiang WANG
Zheao LI
Jie Huang
Wenqi Zhou
Chen Huang

Assignees

SOUTHEAST UNIVERSITY

Dates

Publication Date: 20260505
Application Date: 20230319
Priority Date: 20220307

Claims (5)

1 . A predictive channel modeling method based on a generative adversarial network and a long short-term memory artificial neural network, wherein the method comprises following steps: Step 1, determining environment where wireless channels are positioned and physical environment parameters for an antenna position; Step 2, determining a frequency band, a line-of-sight condition and a non-line-of-sight condition that are used for channel measurements in current environment, and collecting, by using channel measurement instruments, a channel impulse response to channel measurement data; Step 3, preprocessing the collected channel measurement data, normalizing, in a case where a neural network is used to train and learn channel data, an input channel impulse response data set as needed; Step 4, constructing a predictive channel model based on the generative adversarial network and the long short-term memory artificial neural network; firstly, constructing a part of a generative model for the generative adversarial network in the predictive channel model, and a function of the generative model is to generate the channel data with specific properties by utilizing noise vectors z; Step 5, constructing a part of a discriminative model for the generative adversarial network in the predictive channel model; Step 6, training, after constructing the predictive channel model based on the generative adversarial network and the long short-term memory artificial neural network, the predictive channel model; and Step 7, obtaining trained predictive channel model parameters and analyzing channel statistical characteristics.
2 . The predictive channel modeling method based on the generative adversarial network and the long short-term memory artificial neural network according to claim 1 , wherein Step 3 specifically includes following steps: Step 3.1, calculating a mean value μ for channel measurement data examples and a standard deviation σ of the channel measurement data examples; and Step 3.2, normalizing, by using a z-score standardization means, the channel impulse response to the channel measurement data: X = x - μ σ , where X is normalized data and x is original measurement data.
3 . The predictive channel modeling method based on the generative adversarial network and a long short-term memory artificial neural network according to claim 2 , wherein Step 4 specifically includes following steps: Step 4.1, taking three convolutional layers as a main structure of the generative model for the generative adversarial network; Step 4.2, adding a batch normalization layer after each of the convolutional layers, and normalizing an output of each nodes in the neural network to enhance a generalization and a robustness of the generative model; Step 4.3, reprocessing, by using an activation function of a parametric rectified linear unit; and Step 4.4, adding the artificial neural network into the constructed generative model for the generative adversarial network as an input port, to construct a generative model framework for the generative adversarial network and the long short-term memory artificial neural network; wherein a predictive issue of the long short-term memory artificial neural network is to predict a subsequent channel state by using states previous to j data in continuous channel data in a space-time domain, and an output h t of the long short-term memory artificial neural network includes obtained time step update information from the space-time domain sequence: h t = f LSTM ( h ⁡ ( n - j + 1 ) , h ⁡ ( n - j + 2 ) , … , h ⁡ ( n ) ) , where f LSTM represents a function of the long short-term memory artificial neural network for a sequence prediction on sequence data, h(n) is the channel impulse response to the channel measurement data, and n is a n-th sampling point in a latency the continuous channel impulse response sequence.
4 . The predictive channel modeling method based on the generative adversarial network and a long short-term memory artificial neural network according to claim 3 , wherein Step 5 specifically includes following steps: Step 5.1, taking four convolutional layers as a main structure of the discriminative model for the generative adversarial network, for extracting multi-dimensional characteristics in a time domain, to facilitate identifying and correcting the generated channel data; Step 5.2, adding a batch normalization layer after each of the convolutional layers to enhance a generalization and a robustness of the model; Step 5.3, using an activation function of a parametric rectified linear unit to improve an accuracy and an efficiency of a training; Step 5.4, adding a Dropout layer after the activation function of the parametric rectified linear unit to improve an anti over-fitting ability of the discriminative model; Step 5.5, adding an additional fully connected layer after a network structure of a last convolutional layer to act as a classifier in an entire convolutional neural network; and Step 5.6, adding a Sigmoid growth curve activation function at an end of the discriminative model.
5 . The predictive channel modeling method based on the generative adversarial network and a long short-term memory artificial neural network according to claim 4 , wherein Step 6 specifically includes following steps: Step 6.1, inputting the channel impulse response obtained from measurements as a reference sample for an input of the model for training a predictive channel model; Step 6.2, achieving a balance between an identification accuracy rate of the generative model for the predictive channel model and an identification accuracy rate of the discriminative model for the predictive channel model, during an iterative adversarial learning process of model training, and Step 6.2 specifically includes following steps: Step 6.2.1, setting the generative model to process a random Gaussian noise vector z to obtain a generated data distribution P G , wherein an actual data distribution is P data ; Step 6.2.2, repeatedly using, in a case where the generated channel data are input into the discriminative model, these networks as needed in multiple training iterations, and minimizing, by dynamically adjusting a weight of this layer, a loss value for the discriminative model; and Step 6.2.3, enabling, during a cyclic training process, an ability of the discriminative model to tend towards a convergence value, thus obtaining an optimal value: D G * = P data P G + P data , where D G * is an optimal value for a discriminator.

Description

TECHNICAL FIELD The present disclosure belongs to the technical field of channel modeling, and specifically relates to a predictive channel modeling method based on a generative adversarial network (GAN) and a long short-term memory (LSTM) artificial neural network. BACKGROUND With the developments on new technologies and applications of the sixth generation wireless communication system (6G), the traditional passive channel characterizations bring some problems, such as the high channel measurement costs, the complex channel parameter estimations, and the lack of predictive capability to unknown information. The complex and various scenarios in the 6G standard require high-performance channel detectors for measurements, while the prices of these instruments are extremely high, and the channel measurements cannot exhaust all frequency bands and scenarios. In the case where the channel parameters are estimated, the channel data amount that needs to be processed is extremely large, and the algorithm complexity is high. Eventually, the traditional non-predictive channel models cannot predict future channel characteristics, as well as the unknown frequency bands or scenarios. In view of the above problems, a novel predictive channel modeling based on the 6G vision needs to be proposed, which actively identifies or controls the channels according to the physical environment. Artificial intelligence (AI) technology rapidly develops and becomes a hot field thanks to its powerful ability in solving real-time prediction problems in the channel modeling. AI can not only learn and extract the potential characteristics that the traditional modeling methods cannot describe and collect from the real channel measurement data, but also can predict the channel distribution characteristics of the future time, the unknown frequency bands, and the unknown scenarios from the known information. The channel parameter estimation is an essential process for channel analysis and modeling. The channel characteristic prediction based on AI can greatly improve the accuracy and the efficiency in extracting parameters, and equipped with the ability of predicting channel characteristics from the specific environment to the general channel environment. At present, the method for predicting and modeling the channel parameters based on AI generally relies on learning and training a plurality of independent channel characteristics in the data set, such as received power, delay spread (DS), and model angle information. The artificial intelligence algorithms used in these methods are generally the deep learning network algorithms such as feed-forward neural network (FNN), radial basis function neural network (RBF-NN), and convolutional neural network (CNN). However, these AI-based channel parameter prediction methods for the channel measurement data firstly increase the computational complexity of the channel modeling, and the independent channel parameter index values obtained after the prediction cannot effectively and directly reflect the continuous channel characteristics in the space-time domains. Channel measurement is the core and the foundation of the traditional channel modeling research, while it is quite difficult to conduct the channel measurement and collect the data. On the one hand, the high-precision channel detection instruments are required, and on the other hand, a large number of manpower are required to conduct channel measurement campaigns in different communication scenarios. Due to the limitations in the manpower and the costs, the measurement scenarios are commonly quite limited, and the channel measurement for all of the frequency bands and the scenarios can not be exhausted. In addition, in order to achieve the balance between the universality and accuracy, the traditional channel models still have disadvantages in accuracy for certain specific communication scenarios. And in the modeling process, the parameter estimation has the disadvantages of requiring a large amount of data with high algorithm complexity. Moreover, due to the changes or the limitations on the environment, the measurement errors caused by human subjective judgments, and improper or faulty operations on the measurement instruments, a part of the channel measurement data are lost or affected by the errors, which leads to that the collected measurement data sets are insufficient and the data are inaccurate. SUMMARY The objectives of the present disclosure are to provide a predictive channel modeling method based on a GAN and a LSTM artificial neural network, to solve the technical problems of insufficient data set, low qualities and diversities in the required channel data, and low parameter generation efficiency in the channel modeling. In order to solve the above-mentioned technical problems, the specific technical solutions of the present disclosure are as follows. Provided is a predictive channel modeling method based on a GAN and a LSTM artifici