阅读说明：本技术 一种毫米波亚毫米波频段非对称信道参数预测方法 (Millimeter wave and submillimeter wave frequency band asymmetric channel parameter prediction method ) 是由 张焱 袁萌 何遵文 李悦 张万成 宋九鹏 于 2021-11-12 设计创作，主要内容包括：本发公开的一种毫米波亚毫米波频段非对称信道参数预测方法,属于无线通信领域。本发明实现方法为：根据上行信道参数和环境特征预测下行信道参数,解决非对称的上下行信道之间不存在互易性,对称信道的仿真方法不能直接用于非对称信道的问题,实现非对称毫米波亚毫米波上下行无线信道准确、高效的联合生成；结合实例迁移算法预测新传播条件下的下行信道参数,解决新传播环境下训练样本不足的问题；本发明能够减少弱相关或不相关特征的计算量和对下行信道参数预测模型精度的负面影响,降低下行信道参数预测模型的复杂度。本发明应用于无线通信领域,能够支撑非对称毫米波亚毫米波通信系统设计、部署、优化,提高非对称信道下无线通信的效率和精度。(The invention discloses a millimeter wave and submillimeter wave frequency band asymmetric channel parameter prediction method, and belongs to the field of wireless communication. The implementation method of the invention comprises the following steps: predicting parameters of a downlink channel according to parameters of the uplink channel and environmental characteristics, solving the problems that reciprocity does not exist between asymmetric uplink and downlink channels and a simulation method of a symmetric channel cannot be directly used for the asymmetric channel, and realizing accurate and efficient combined generation of the asymmetric millimeter wave sub-millimeter wave uplink and downlink wireless channels; predicting downlink channel parameters under a new propagation condition by combining an example migration algorithm, and solving the problem of insufficient training samples under a new propagation environment; the method can reduce the calculated amount of weakly correlated or uncorrelated characteristics and the negative influence on the accuracy of the downlink channel parameter prediction model, and reduce the complexity of the downlink channel parameter prediction model. The invention is applied to the field of wireless communication, can support the design, deployment and optimization of an asymmetric millimeter wave and submillimeter wave communication system, and improves the efficiency and precision of wireless communication under an asymmetric channel.)

1. A millimeter wave and submillimeter wave frequency band asymmetric channel parameter prediction method is characterized by comprising the following steps: the method comprises the following steps:

determining importance degrees of uplink channel parameters and environmental characteristics on downlink channel parameter prediction by using a SHAP algorithm, sequencing according to the importance degrees, determining the optimal characteristics for the downlink channel parameter prediction by using an MDL algorithm, and selecting characteristics with strong correlation with the downlink channel parameters from the uplink channel parameters and the environmental characteristics, namely, reducing the calculated amount of weak correlation or irrelevant characteristics and the negative influence on the accuracy of a downlink channel parameter prediction model by using a SHAP and MDL-based characteristic selection algorithm, thereby effectively reducing the complexity of the downlink channel parameter prediction model; constructing a training set by the optimal characteristics, and inputting the training set constructed based on the optimal characteristics into a downlink channel parameter prediction model;

the channel parameters comprise path loss, multipath number and time delay expansion, and the environment characteristics comprise a terminal position, a propagation distance, an azimuth angle and the number of buildings;

secondly, the characteristic that the channel distribution environment presents asymmetry is fully considered, after the optimal uplink channel parameters and environment characteristics are selected in the first step, the downlink channel parameters are predicted according to the selected uplink channel parameters and the selected environment characteristics, the problem that reciprocity does not exist between asymmetric uplink and downlink channels, and a simulation method of a symmetric channel cannot be directly used for the asymmetric channel is solved, so that accurate and efficient combined generation of the asymmetric millimeter wave sub-millimeter wave uplink and downlink wireless channels is realized; in addition, the weak learners use the selected optimal uplink channel parameters, the environmental characteristics and the prediction target downlink channel parameters for training to obtain the prediction result and the corresponding weight of each weak learner, a downlink channel parameter prediction model based on ensemble learning is constructed by combining the weak learners, the downlink channel parameters under the same propagation condition are predicted based on the prediction model of the ensemble learning, and the performance of the downlink channel parameter prediction model is improved.

2. The method for predicting the asymmetric channel parameters in the millimeter wave and submillimeter wave frequency band according to claim 1, wherein: the method also comprises a third step of,

establishing a downlink channel parameter prediction model based on example migration to predict downlink channel parameters under new propagation conditions, is the total training set, initializes the source domainAnd a target domainWeight distribution W of₁R2 algorithm is adopted, the weight adjustment of a source domain example and a target domain example is divided into two stages, and in the first stage, calculation is only carried out in a target domainMean error e of_tAdjusting the target domainWeight of, source domainThe second stage is to calculate only on the source domainAverage error e of_tAdjusting the source domainWeight of, target domainWeight of (2)And keeping the prediction model unchanged to obtain a downlink channel parameter prediction model based on example migration, and rapidly predicting the downlink channel parameters in a new propagation environment.

3. The method for predicting the asymmetric channel parameters in the millimeter wave and submillimeter wave frequency band according to claim 1 or 2, characterized in that: and step four, according to the downlink channel parameters predicted in the step two or the step three, the method is applied to the field of wireless communication, supports design, deployment and optimization of the asymmetric millimeter wave and submillimeter wave communication system, and improves efficiency and precision of wireless communication under the asymmetric channel.

4. The method for predicting the asymmetric channel parameters in the millimeter wave and submillimeter wave frequency band according to claim 1, 2 or 3, characterized in that: the first implementation method comprises the following steps of,

step 1.1: determining importance degrees of uplink channel parameters and environmental characteristics to downlink channel parameter prediction by using a SHAP algorithm, and sequencing according to the importance degrees;

determining the importance degree of the uplink channel parameters and the environmental characteristics to the downlink channel parameter prediction by adopting a SHAP algorithm, wherein a median SHAP value is used for reflecting the relevance of the uplink channel parameters and the environmental characteristics to the downlink channel parameter prediction and is used for measuring the importance of the characteristics; considering the ith training sample, the basic prediction model used to calculate the SHAP value is denoted G, which is:

whereinIs the uplink channel parameter of the ith training sample,is the environmental characteristic of the ith training sample,SHAP value of the kth feature of the ith training sample; k is the total number of features;is a sample set;indicating the presence of the kth feature in the sample set,indicating its absence; phi is a₀Is the average predicted value of the training samples, calculated by:

SHAP value of k characteristic of i training sampleExpressed as:

wherein the content of the first and second substances,means fromRemovingIs the kth feature, S isIn order to calculate the importance of the kth feature to the predicted outcome, the modelByS, training; g_SIs a model trained with S only; each sample in the training set has its own SHAP value; to eliminate the effect of few extreme samples on the results, the median of the sample SHAP values was chosen to represent the overall importance of the features; median SHAP value of normalized features, e.g. s₁,s₂,…,s_KRepresenting the importance value of each feature in the model training;

step 1.2: according to the importance degree sequence obtained in the step 1.1, the optimal feature quantity for predicting the parameters of the downlink channel is determined through an MDL algorithm, and the features with strong correlation with the parameters of the downlink channel are selected from the parameters of the uplink channel and the environmental features, namely, the feature selection algorithm based on the SHAP and the MDL is adopted, so that the calculated quantity of the weakly correlated or uncorrelated features and the negative influence on the accuracy of a prediction model of the parameters of the downlink channel are reduced, and the complexity of the prediction model of the parameters of the downlink channel is effectively reduced;

after SHAP median values of all the features are calculated, the number of the selected features is calculated by adopting an MDL algorithm; the number of selected features calculated by the MDL isThe selected characteristic is expressed asThe objective function threshold of the MDL criteria used to compute the feature selection threshold is defined as:

where K is the total number of features, s_aIs the median SHAP value of the feature;

number of selected featuresEstimated as:

selecting the most importantA feature;

step 1.3: and (3) constructing a training set based on the optimal characteristics selected in the step (1.2), and inputting the training set constructed based on the optimal characteristics into a downlink channel parameter prediction model.

5. The method for predicting the asymmetric channel parameters in the millimeter wave and submillimeter wave frequency bands as claimed in claim 4, wherein: in the second step, the weak learner uses the selected optimal uplink channel parameters, the environmental characteristics and the prediction target downlink channel parameters to train so as to obtain the prediction result and the corresponding weight of each weak learner, and a downlink channel parameter prediction model based on ensemble learning is constructed by combining the weak learners so as to improve the performance of the downlink channel parameter prediction model,

weak learner using selected optimal uplink channel parametersEnvironmental characteristicsAnd predicting target downlink channel parametersTraining is carried out to obtain the prediction result p of each weak learner₁,...,p_LAnd their corresponding weights w₁,...,w_LAnd constructing a downlink channel parameter prediction model based on ensemble learning by combining the weak learners as shown in the following formula:

wherein the content of the first and second substances,is a downlink channel parameter prediction result; l is the number of weak learners, p_lIs the predicted result of the l weak learner, w_lIs the weight of the l-th weak learner.

6. The method for predicting the asymmetric channel parameters in the millimeter wave and submillimeter wave frequency band according to claim 5, wherein: in the third step, a downlink channel parameter prediction model based on example migration is established based on a two-stage TrAdaBoost. R2 algorithm to predict downlink channel parameters under a new propagation condition, and the prediction process is as follows:

source domain training set namedIt contains n samples, named as the target domain training setContains m samples;is the total training set;the uplink channel parameter of the v-th sample in (1) is namedEnvironmental feature of the v-th sampleIs named asThe predicted downlink channel parameter for the v-th sample is namedThe index of the sample in the original training set is represented, v is n +1, …, and n + m represents the index of the sample in the new training set; the name of a downlink channel parameter prediction model under a new propagation environment is M_T(ii) a First, initialization is performedAndweight distribution W of₁T represents the number of training iterations and is set to 1 during initialization;

W₁＝{w_1,1,...,w_1,i,...,w_1,n+m} (7)

wherein w_1,vThe weight representing the v-th sample in the first iteration is given by:

for each iteration T2, 3Training and constructing regression estimatorAnd is calculated only inMean error e of_tComprises the following steps:

then, willThe weight of (1) is updated as:

w_t,vrepresenting the weight of the v sample in the t iteration; z_tIs a normalization constant; beta is a_tIs given by

The weight of (d) remains unchanged; the process is a first stage of downlink channel parameter prediction in a new propagation environment; in the second stage of the prediction of the downlink channel parameters in the new propagation environment,error rate of_tThe calculation is as follows:

then, updateA weight of (2) andweight protection ofKeeping unchanged;

wherein gamma is_tGiven by:

obtaining an instance migration modelWhereinAs a downlink channel parameter prediction model under a new propagation condition; and inputting the selected characteristics of the new propagation environment into the trained downlink channel parameter prediction model to obtain a downlink channel parameter prediction model based on instance migration, and rapidly predicting the downlink channel parameters in the new propagation environment.

Technical Field

The invention belongs to the field of wireless communication, and particularly relates to a millimeter wave and submillimeter wave frequency band asymmetric channel parameter prediction method.

Background

In order to meet the requirements of high spectrum utilization and high energy utilization, large-scale multiple-input multiple-output (MIMO) antenna arrays are widely used in fifth-generation (5G) mobile networks. However, to meet the increasing demands of the sixth generation (6G) mobile networks, the number of antenna elements is increasing, resulting in the data processing burden and hardware cost becoming higher and higher. To solve this problem, asymmetric all-digital beamforming massive MIMO systems have been proposed. Different from the traditional symmetrical large-scale beam forming antenna array, the number of the receiving radio frequency links adopted by the asymmetrical large-scale MIMO system is less than that of the transmitting radio frequency links, so that the hardware cost and the energy consumption are reduced, and the asymmetrical large-scale MIMO system has high application potential in a future wireless communication system.

However, one problem that arises in asymmetric massive MIMO systems is that reciprocity between the uplink and downlink channels is no longer satisfied. This means that even in a TDD system, the measurement estimation result of the uplink channel cannot be directly used to predict the downlink channel. Therefore, both the Base Station (BS) and the terminal (UE) require pilot signals, which occupy more transmission and computation resources. This problem is more severe in the downlink because the number of transmit rf antennas on the base station side is very large. In fact, the non-reciprocity of the uplink and downlink channels is caused by different Tx and Rx beam patterns. I.e. the up/down signals do not propagate under the same propagation conditions. Analyzing the relationship between the beamwidth and the channel parameters helps to describe this non-reciprocity. Currently, some research work has been conducted on the aspect of studying the influence of the beam width on the channel parameters, and the research work can be divided into two categories. The first category is based on empirical models. However, they can only provide statistical results, which are of limited accuracy. Furthermore, it is only applicable to specific beamwidths and scenarios. The second category is that using ray tracing techniques, ray tracing based methods can provide accurate results but require a significant amount of time and computing power. Aiming at the millimeter wave sub-millimeter frequency band asymmetric channel parameter prediction model, a flow modeling method is formed, model accuracy and efficiency are considered, and the method has very important significance for design and deployment of future millimeter wave and sub-millimeter wave mobile communication systems.

Disclosure of Invention

In order to solve the problem that the asymmetry of a transmission environment is not fully considered in the existing wireless channel simulation technology, the invention aims to provide a method for predicting parameters of an asymmetric channel of a millimeter wave sub-millimeter wave frequency band, which predicts parameters of a downlink channel according to parameters of an uplink channel and environmental characteristics, solves the problems that the reciprocity does not exist between asymmetric uplink and downlink channels and a symmetric channel simulation method cannot be directly used for the asymmetric channel, and further realizes the accurate and efficient combined generation of the asymmetric millimeter wave sub-millimeter wave uplink and downlink wireless channels; and predicting the downlink channel parameters under the new propagation condition by combining an example migration algorithm, solving the problem of insufficient training samples under the new propagation environment, and rapidly predicting the downlink channel parameters under the new propagation environment. The invention is applied to the field of wireless communication, can support the design, deployment and optimization of an asymmetric millimeter wave and submillimeter wave communication system, and improves the efficiency and precision of wireless communication under an asymmetric channel.

The purpose of the invention is realized by the following technical scheme:

the invention discloses a millimeter wave and submillimeter wave frequency band asymmetric channel parameter prediction method, which comprises the following steps:

the method comprises the steps of firstly, determining importance degrees of uplink channel parameters and environmental characteristics on downlink channel parameter prediction by using a SHAP algorithm, sequencing according to the importance degrees, determining the optimal characteristics for the downlink channel parameter prediction by using an MDL algorithm, selecting characteristics with strong correlation with the downlink channel parameters from the uplink channel parameters and the environmental characteristics, namely, reducing the calculated amount of weak correlation or irrelevant characteristics and the negative influence on the accuracy of a downlink channel parameter prediction model by using a SHAP and MDL-based characteristic selection algorithm, and effectively reducing the complexity of the downlink channel parameter prediction model. And constructing a training set by using the optimal characteristics, and inputting the training set constructed based on the optimal characteristics into a downlink channel parameter prediction model.

The channel parameters comprise path loss, multipath number and time delay expansion, and the environment characteristics comprise terminal position, propagation distance, azimuth angle and building number.

Step 1.1: and determining the importance degree of the uplink channel parameters and the environmental characteristics to the downlink channel parameter prediction by using a SHAP algorithm, and sequencing according to the importance degree.

And determining the importance degree of the uplink channel parameters and the environmental characteristics to the downlink channel parameter prediction by adopting a SHAP algorithm, wherein the median SHAP value is used for reflecting the correlation of the uplink channel parameters and the environmental characteristics to the downlink channel parameter prediction and is used for measuring the importance of the characteristics. Considering the ith training sample, the basic prediction model used to calculate the SHAP value is denoted G, which is:

whereinIs the uplink channel parameter of the ith training sample,is the environmental characteristic of the ith training sample,the SHAP value of the k-th feature of the i-th training sample. K is the total number of features.Is a sample set.Indicating the presence of the kth feature in the sample set,indicating that it is not present. Phi is a₀Is the average predicted value of the training samples, calculated by:

SHAP value of k characteristic of i training sampleExpressed as:

wherein the content of the first and second substances,means fromRemoving Is the kth feature, S isIn order to calculate the importance of the kth feature to the predicted outcome, the modelByAnd S, training. G_SIs a model trained with S only. Each sample in the training set has its own SHAP value. To eliminate the effect of a few extreme samples on the results, the median of the sample SHAP values was chosen to represent the overall importance of the feature. After normalizationMedian SHAP value of feature such as s₁,s₂,…,s_KAnd representing the importance value of each feature in the model training.

Step 1.2: according to the importance degree sequence obtained in the step 1.1, the optimal feature quantity for predicting the parameters of the downlink channel is determined through an MDL algorithm, and the features with strong correlation with the parameters of the downlink channel are selected from the parameters of the uplink channel and the environmental features, namely, the feature selection algorithm based on the SHAP and the MDL is adopted, so that the calculation amount of the weakly correlated or uncorrelated features and the negative influence on the accuracy of the prediction model of the parameters of the downlink channel are reduced, and the complexity of the prediction model of the parameters of the downlink channel is effectively reduced.

After the SHAP median of all features is calculated, the number of selected features is calculated using the MDL algorithm. The number of selected features calculated by the MDL isThe selected feature is represented as f₁′,f₂′,…,The objective function threshold of the MDL criteria used to compute the feature selection threshold is defined as:

where K is the total number of features, s_aIs the median SHAP value of the feature.

Number of selected featuresEstimated as:

selecting the most importantAnd (4) a feature.

Step 1.3: and (3) constructing a training set based on the optimal characteristics selected in the step (1.2), and inputting the training set constructed based on the optimal characteristics into a downlink channel parameter prediction model.

Secondly, the characteristic that the channel distribution environment presents asymmetry is fully considered, after the optimal uplink channel parameters and environment characteristics are selected in the first step, the downlink channel parameters are predicted according to the selected uplink channel parameters and the selected environment characteristics, the problem that reciprocity does not exist between asymmetric uplink and downlink channels, and a simulation method of a symmetric channel cannot be directly used for the asymmetric channel is solved, so that accurate and efficient combined generation of the asymmetric millimeter wave sub-millimeter wave uplink and downlink wireless channels is realized; in addition, the weak learners use the selected optimal uplink channel parameters, the environmental characteristics and the prediction target downlink channel parameters for training to obtain the prediction result and the corresponding weight of each weak learner, a downlink channel parameter prediction model based on ensemble learning is constructed by combining the weak learners, the downlink channel parameters under the same propagation condition are predicted based on the prediction model of the ensemble learning, and the performance of the downlink channel parameter prediction model is improved.

The weak learners are trained by using the selected optimal uplink channel parameters, the environmental characteristics and the prediction target downlink channel parameters to obtain the prediction result and the corresponding weight of each weak learner, and the weak learners are combined to construct a downlink channel parameter prediction model based on ensemble learning, so that the performance of the downlink channel parameter prediction model is improved, and the specific implementation method comprises the following steps:

weak learner using selected optimal uplink channel parametersEnvironmental characteristicsAnd predicting target downlink channel parametersTraining is carried out to obtain the prediction result p of each weak learner₁,...,p_LAnd their corresponding weights w₁,...,w_LAnd constructing a downlink channel parameter prediction model based on ensemble learning by combining the weak learners as shown in the following formula:

wherein the content of the first and second substances,is a downlink channel parameter prediction result. L is the number of weak learners, p_lIs the predicted result of the l weak learner, w_lIs the weight of the l-th weak learner.

Predicting the downlink channel parameters under the same propagation condition based on the prediction model based on ensemble learning established in the step two, but when the propagation environment changes rapidly, enough training samples are difficult to obtain in a short time, the prediction precision of the ensemble learning model is influenced, and in order to solve the problem that the training samples are insufficient under the new propagation environment, the method also comprises the step three of establishing a downlink channel parameter prediction model based on example migration to predict the downlink channel parameters under the new propagation environment,is the total training set, initializes the source domainAnd a target domainWeight distribution W of₁R2 algorithm is adopted, the weight adjustment of a source domain example and a target domain example is divided into two stages, and in the first stage, calculation is only carried out in a target domainMean error e of_tAdjusting the target domainWeight of, source domainThe second stage is to calculate only on the source domainAverage error e of_tAdjusting the source domainWeight of, target domainThe weight of the downlink channel is kept unchanged, a downlink channel parameter prediction model based on instance migration is obtained, and the downlink channel parameters are rapidly predicted in a new propagation environment.

Establishing a downlink channel parameter prediction model based on example migration based on a two-stage TrAdaBoost. R2 algorithm to predict downlink channel parameters under a new propagation condition, wherein the prediction process is described as follows:

source domain training set namedIt contains n samples, named as the target domain training setContaining m samples.Is the total training set.The uplink channel parameter of the v-th sample in (1) is namedThe environmental characteristics of the v-th sample are namedThe predicted downlink channel parameter for the v-th sample is namedv-1, …, n denotes the index of the sample in the original training set, v-n +1, …, n + m denotes the index of the sample in the new training set. The name of a downlink channel parameter prediction model under a new propagation environment is M_T. First, initialization is performedAndweight distribution W of₁And t represents the number of training iterations and is set to 1 at initialization.

W₁＝{w_1,1,...,w_1,i,...,w_1,n+m} (7)

Wherein w_1,vThe weight representing the v-th sample in the first iteration is given by:

for each iteration T2, 3Training and constructing regression estimatorAnd is calculated only inMean error e of_tComprises the following steps:

then, willThe weight of (1) is updated as:

w_t,vrepresenting the weight of the v-th sample in the t-th iteration. Z_tIs a normalization constant. Beta is a_tIs given by

The weight of (c) remains unchanged. This process is the first stage of downlink channel parameter prediction in a new propagation environment.

In the second stage of the prediction of the downlink channel parameters in the new propagation environment,error rate of_tThe calculation is as follows:

then, updateA weight of (2) andthe weight of (c) remains unchanged.

Wherein gamma is_tGiven by:

obtaining an instance migration modelWhereinAs a prediction model of the parameters of the downlink channel under the new propagation condition. And inputting the selected characteristics of the new propagation environment into the trained downlink channel parameter prediction model to obtain a downlink channel parameter prediction model based on instance migration, and rapidly predicting the downlink channel parameters in the new propagation environment.

The method also comprises the following four steps: and (4) applying the downlink channel parameters predicted in the second step or the third step to the field of wireless communication, supporting design, deployment and optimization of the asymmetric millimeter wave and submillimeter wave communication system, and improving the efficiency and precision of wireless communication under the asymmetric channel.

Has the advantages that:

1. the invention discloses a millimeter wave sub-millimeter wave frequency band asymmetric channel parameter prediction method, which predicts parameters of a downlink channel according to parameters of the uplink channel and environmental characteristics, solves the problems that no reciprocity exists between asymmetric uplink and downlink channels, and a symmetric channel simulation method cannot be directly used for the asymmetric channels, and further realizes accurate and efficient combined generation of asymmetric millimeter wave sub-millimeter wave uplink and downlink wireless channels.

2. The invention discloses a millimeter wave and submillimeter wave frequency band asymmetric channel parameter prediction method, which is used for predicting a downlink channel parameter under a new propagation condition by combining an example migration algorithm, solving the problem of insufficient training samples under a new propagation environment and rapidly predicting the downlink channel parameter under the new propagation environment.

3. The invention discloses a millimeter wave sub-millimeter wave frequency band asymmetric channel parameter prediction method, which adopts a SHAP algorithm to determine the importance degree of uplink channel parameters and environmental characteristics to downlink channel parameter prediction, determines the optimal characteristics for the downlink channel parameter prediction through an MDL algorithm according to the importance degree sequence, and selects the characteristics with strong correlation with the downlink channel parameters from the uplink channel parameters and the environmental characteristics, namely adopts a characteristic selection algorithm based on the SHAP and the MDL, reduces the calculated amount of weak correlation or irrelevant characteristics and the negative influence on the accuracy of a downlink channel parameter prediction model, and effectively reduces the complexity of the downlink channel parameter prediction model.

4. The millimeter wave sub-millimeter wave frequency band asymmetric channel parameter prediction method disclosed by the invention can be applied to the field of wireless communication, supports the design, deployment and optimization of an asymmetric millimeter wave sub-millimeter wave communication system, and improves the efficiency and precision of wireless communication under an asymmetric channel.

Drawings

Fig. 1 is a flowchart of a millimeter wave and submillimeter wave frequency band asymmetric channel parameter prediction method according to the present invention.

Fig. 2 is a case simulation environment, the ottawa urban distribution environment.

Fig. 3 shows the result of the MDL-based feature selection algorithm when the prediction parameter is path loss.

Fig. 4 shows the prediction effect of the prediction model of the downlink channel parameter based on the Adaboost algorithm under the same propagation condition when the prediction parameter is the path loss.

Fig. 5 is a comparison of prediction accuracy of a downlink channel parameter prediction model based on the Adaboost algorithm before and after migration under a new propagation condition when a prediction parameter is a path loss.

Detailed Description

The following describes in detail a method for predicting parameters of an asymmetric millimeter wave downlink channel according to an embodiment of the present invention with reference to the accompanying drawings.

As shown in fig. 1, the method for predicting parameters of an asymmetric channel in a millimeter wave and submillimeter wave frequency band disclosed in this embodiment includes the following specific steps: the parameters are generated by simulation of the Wireless Insite software, and the simulation area is a typical urban area of Ottawa in Canada as shown in FIG. 2. The simulation area is 1000m × 630 m. The base station is fixed at the center, and the height from the ground is 30 m. The propagation environment is divided into three parts centered on the base station location. The receiving beam width of the base station is 60 degrees/120 degrees, and the transmitting beam width is 10 degrees/30 degrees. All terminals adopt omnidirectional antennas for transmission and reception and are distributed along roads. Each terminal has a height of 1.5 meters and there are a total of 5635 different terminal locations at each of which channel parameters are calculated.

The method comprises the following steps: the method comprises the steps of determining importance degrees of uplink channel parameters and environmental characteristics on downlink channel parameter prediction by using a SHAP algorithm, sequencing according to the importance degrees, determining the optimal characteristics for the downlink channel parameter prediction by using an MDL algorithm, and selecting the characteristics with strong correlation with the downlink channel parameters from the uplink channel parameters and the environmental characteristics, namely, reducing the calculated amount of weak correlation or irrelevant characteristics and the negative influence on the accuracy of a downlink channel parameter prediction model by using a SHAP and MDL-based characteristic selection algorithm, thereby effectively reducing the complexity of the downlink channel parameter prediction model. And constructing a training set by using the optimal characteristics, and inputting the training set constructed based on the optimal characteristics into a downlink channel parameter prediction model. The channel parameters and environmental characteristics are shown in the following table.

Step 1.1 case consideration of the downstream channel parameter pathloss PL_DLIs a parameter to be predicted, and adopts SHAP algorithm to determine the path loss PL of the uplink channel parameter and the environment characteristic to the downlink channel parameter_DLThe predicted importance degree, the median SHAP value is used for reflecting the path loss PL of the uplink channel parameters and the environment characteristics to the downlink channel parameters_DLThe predicted relevance is used to measure the importance of the features. Considering the ith training sample, the basic predictive model used to calculate the SHAP value is denoted G, which can be expressed as:

whereinIs the uplink channel parameter of the ith training sample,is the environmental characteristic of the ith training sample,the SHAP value of the k-th feature of the i-th training sample. K is the total number of features.Is a sample set.Indicating the presence of the kth feature in the sample set,indicating that it is not present. Phi is a₀Is the average predicted value of the training samples, calculated by:

SHAP value of k characteristic of i training sampleExpressed as:

wherein the content of the first and second substances,means fromRemoving Is the kth feature, S isIn order to calculate the importance of the kth feature to the predicted outcome, the modelByAnd S, training. G_SIs a model trained with S only. Each sample in the training set has its own SHAP value. To eliminate the effect of a few extreme samples on the results, the median of the sample SHAP values was chosen to represent the overall importance of the feature.

And 1.2, calculating the quantity of the selected features by adopting an MDL algorithm after the SHAP median values of all the features are calculated. The number of selected features calculated by the MDL isThe selected feature is represented as f₁′,f₂′,…,The objective function threshold of the MDL criteria used to compute the feature selection threshold is defined as:

wherein the number of selected featuresEstimated as:

selecting the most importantAnd (4) a feature.

And 1.3, constructing a training set based on the selected optimal features as shown in the following table based on the selected optimal features, and inputting the training set constructed based on the optimal features into a downlink channel parameter prediction model.

Step two, fully considering the characteristic that the channel distribution environment presents asymmetry, after selecting the optimal uplink channel parameter and environment characteristic in the step one, predicting the path loss PL of the downlink channel parameter according to the selected uplink channel parameter and environment characteristic_DLThe problem that reciprocity does not exist between asymmetric uplink and downlink channels and a simulation method of a symmetric channel cannot be directly used for the asymmetric channel is solved, and accurate and efficient combined generation of the asymmetric millimeter wave sub-millimeter wave uplink and downlink wireless channels is further realized; in addition, the weak learner uses the selected optimal uplink channel parameters and environmental characteristics and the predicted target downlink channel parameters path loss PL_DLTraining is carried out to obtain a prediction result and corresponding weight of each weak learner, an integrated learning downlink channel parameter prediction model based on an Adaboost algorithm is built by combining the weak learners, the integrated learning downlink channel parameter prediction model based on the Adaboost algorithm predicts downlink channel parameters under the same propagation condition, and the performance of the downlink channel parameter prediction model is improved.

The Adaboost algorithm is to reasonably combine weak learners to create a strong learner. In the Adaboost algorithm, there are two weights. One is the weight of the weak learner. The weights of the weak learners trained in each iteration will be adjusted according to their prediction accuracy. Weak learners with smaller prediction errors will be assigned higher weights. The other is the weight of the training sample. In this iteration, the weight of the sample with large prediction error will increase. More attention is paid to these samples in the next iteration. Thus, each weak learner only focuses on a portion of the data, and the weak learners can be applied in combination to the entire data set.

Weak learner using selected optimal uplink channel parametersEnvironmental characteristicsAnd predicting the path loss of the target downlink channel parameterTraining is carried out to obtain the prediction result p of each weak learner₁,...,p_LAnd their corresponding weights w₁,...,w_LAnd constructing a downlink channel parameter prediction model based on ensemble learning by combining the weak learners as shown in the following formula:

wherein the content of the first and second substances,is the downlink channel parameter pathloss PL_DLAnd predicting the result. L is the number of weak learners, p_lIs the predicted result of the l weak learner, w_lIs the weight of the l-th weak learner.

In the case, 2518 data samples at different terminal locations in sector 1 are collected. In order to verify the prediction capability of the proposed integrated learning downlink channel parameter prediction model based on the Adaboost algorithm, collected data are randomly divided into two groups. 80% of the data is used as a training set, the rest of the data is used as a test set, and the prediction effect of the Adaboost method is shown in FIG. 4 by taking the MAE as an evaluation index.

Predicting the downlink channel parameters under the same propagation condition based on the integrated learning prediction model based on the Adaboost method established in the step two, but when the propagation environment changes rapidly, enough training samples are difficult to obtain in a short time, the prediction precision of the integrated learning model is influenced, in order to solve the problem that the training samples are insufficient under the new propagation environment, the method also comprises the step three of establishing a downlink channel parameter prediction model based on example migration to predict the downlink channel parameters under the new propagation condition, is the total training set, initializes the source domainAnd a target domainWeight distribution W of₁R2 algorithm is adopted, the weight adjustment of a source domain example and a target domain example is divided into two stages, and in the first stage, calculation is only carried out in a target domainMean error e of_tAdjusting the target domainWeight of, source domainThe second stage is to calculate only in the source domainMean error of_tAdjusting the source domainWeight of, target domainThe weight of the downlink channel is kept unchanged, a downlink channel parameter prediction model based on instance migration is obtained, and the downlink channel parameters are rapidly predicted in a new propagation environment.

Establishing a downlink channel parameter prediction model based on example migration based on a two-stage TrAdaBoost. R2 algorithm to predict downlink channel parameters under a new propagation condition, wherein the prediction process is described as follows:

source domain training set namedIt contains n samples, named as the target domain training setContaining m samples.Is the total training set.The uplink channel parameter of the v-th sample in (1) is namedThe environmental characteristics of the v-th sample are namedThe predicted downlink channel parameter for the v-th sample is namedv-1, …, n denotes the index of the sample in the original training set, v-n +1, …, n + m denotes the index of the sample in the new training set.

Downlink channel parameter prediction in new propagation environmentsThe name of the model is M_T. First, initialization is performedAndweight distribution W of₁And t represents the number of training iterations and is set to 1 at initialization.

W₁＝{w_1,1,...,w_1,i,...,w_1,n+m} (7)

Wherein w_1,vThe weight representing the v-th sample in the first iteration is given by:

for each iteration T2, 3Training and constructing regression estimatorAnd is calculated only inMean error e of_tComprises the following steps:

then, willThe weight of (1) is updated as:

w_t,vrepresenting the weight of the v-th sample in the t-th iteration. Z_tIs a normalization constant. Beta is a_tGiven by:

the weight of (c) remains unchanged. The process is the first stage of downlink channel parameter prediction in a new propagation environment. In the second stage of the prediction of the downlink channel parameters in the new propagation environment,error rate of_tThe calculation is as follows:

then, updateA weight of (2) andthe weight of (c) remains unchanged.

Wherein gamma is_tGiven by:

obtaining an instance migration modelWhereinAs a prediction model of the parameters of the downlink channel under the new propagation condition. And inputting the selected characteristics of the new propagation environment into the trained downlink channel parameter prediction model to obtain a downlink channel parameter prediction model based on instance migration, and rapidly predicting the downlink channel parameters in the new propagation environment.

In this case, the source domain is a channel having a reception beam width of 120 ° and a transmission beam width of 30 ° in sector 1, and the target domain is a channel having a reception beam width of 120 ° and a transmission beam width of 10 ° in the same sector. The prediction target is the downlink channel parameter pathloss PL with the new transmit beamwidth. The total number of samples in the target domain is 2518, and only 10 to 250 samples are selected for training. The training set consists of 2518 samples in the source domain and selected samples in the target domain. Adaboost algorithm model based on example migration learning is compared with Adaboost algorithm model based on non-migration learning. The result is shown in fig. 5, and the accuracy of the downlink channel parameter prediction model is improved by the example migration algorithm. The prediction accuracy of the model gradually improves as the number of target domain samples increases.

In the embodiment, the parameters of the downlink channel are predicted according to the parameters of the uplink channel and the environmental characteristics, so that the problems that reciprocity does not exist between asymmetric uplink and downlink channels and a simulation method of a symmetric channel cannot be directly used for the asymmetric channel are solved, and accurate and efficient combined generation of the asymmetric millimeter wave sub-millimeter wave uplink and downlink wireless channels is further realized; and predicting the downlink channel parameters under the new propagation condition by combining an example migration algorithm, solving the problem of insufficient training samples under the new propagation environment, and rapidly predicting the downlink channel parameters under the new propagation environment.

In summary, the above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.