A wireless channel propagation loss prediction method driven by measured data

Abstract

The present invention discloses a method for predicting wireless channel propagation loss driven by measured data. For a scene to be tested that has measured conditions and sufficient measured data, an artificial neural network is built as an initial prediction network, and the initial prediction network is trained using measured data; for a scene to be tested that has measured conditions but insufficient measured data, an artificial neural network with dual-weight neurons is used as the initial prediction network, and network training simulation data is first used for pre-training, and then the network training measured data is used for optimization training; for a scene to be tested that does not have measured conditions, an inter-scene network parameter migration module is executed, and an artificial neural network with dual-weight neurons is used as the initial prediction network, and measured data, simulation data of a reference scene, and simulation data of the scene to be tested are used for training. The present invention can effectively achieve accurate prediction of propagation loss in three situations: sufficient, insufficient, and missing measured data.

Description

Technical Field

[0001] The present invention relates to the field of wireless information transmission technology, in particular to the application of propagation loss prediction based on neural networks in different communication scenarios, and specifically to a method for predicting wireless channel propagation loss driven by measured data. Background Art

[0002] With the rapid development of wireless communication technology, the number of various communication terminals has increased dramatically, and the demand for low-latency, high-speed wireless communication networks has become increasingly urgent. In addition, communication scenarios have become more and more diverse, covering various scenarios such as cities, rural areas, suburbs, indoors, and moving vehicles. There are significant differences in the scatterers and multipath effects that signals experience when propagating in different scenarios, which makes the wireless channel diverse and complex, and the propagation loss of the signal shows different characteristics. Therefore, in order to establish a stable and efficient wireless communication link, it is increasingly important to accurately predict the propagation loss in the wireless channel, an important large-scale fading parameter.

[0003] At present, traditional propagation loss prediction methods are mainly divided into deterministic methods and empirical methods. Deterministic methods usually model propagation loss based on the ray tracing (RT) method combined with a specific propagation environment. This method has high accuracy, but it is usually accompanied by problems such as time-consuming calculations and high complexity. Empirical methods usually combine radio wave propagation theory to model propagation loss as a multi-parameter expression related to channel parameters such as delay and carrier frequency. Although this method is simple and effective, it has poor scene dependence and limited measured scenes. In addition, the cost of this method when fitting with measured data is high. As the scene changes and the channel parameters increase, it is sometimes difficult to give an accurate closed expression.

[0004] Machine learning has attracted more and more attention in the field of parameter prediction due to its powerful learning ability. At present, there are few studies on propagation loss prediction based on machine learning, and machine learning algorithms such as neural networks require a large amount of measured data for training to achieve high prediction accuracy. The acquisition cost of measured data is high, and the learning and training cycle is long. In some non-cooperative scenarios, there is even no data available for training and learning. Considering the high manpower and material costs brought about by the actual measurement work and the limitations of conditions such as measurement equipment, its applicability will be greatly reduced when the measured data is insufficient or missing. Therefore, how to improve the structure of the neural network and reasonably use the known conclusions in other similar scenarios to reduce the demand for measured data of the neural network and expand its scope of application is an urgent problem to be solved. Summary of the invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes a method for predicting wireless channel propagation loss driven by measured data. Based on the powerful learning and data analysis prediction capabilities of neural networks, the method comprehensively considers the impact of path delay, azimuth and elevation angles of signal arrival angle on propagation loss. In addition, the impact of different amounts of measured data on the training effect of the neural network is also considered. By optimizing the neural network structure and using the prediction network migration between scenes, the accurate prediction of propagation loss can be effectively achieved in three situations: sufficient, insufficient and missing measured data.

[0006] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0007] A method for predicting wireless channel propagation loss driven by measured data, the method comprising the following steps:

[0008] S1, the user sets the scene to be tested and determines whether the scene to be tested meets the actual measurement conditions; if the scene to be tested does not meet the actual measurement conditions, proceed to step S4; if the scene to be tested meets the actual measurement conditions, proceed to step S2;

[0009] S2, using the measured data acquisition module of the scene to be tested to build a network training measured data set, determine the threshold value of the number of acquisitions according to the scene to be tested, and judge whether the number of receiver positions is less than the threshold value. If it is less than, it is determined that the amount of measured data of the scene to be tested is insufficient, and the process goes to step S3. Otherwise, it is determined that the amount of measured data of the scene to be tested is sufficient, and the measured data network training module is executed to build an artificial neural network as an initial prediction network. The initial prediction network is trained using the network training measured data set, and the weight matrix and bias matrix of the trained initial prediction network are output, and the process goes to step S5;

[0010] S3, execute the simulation data acquisition module of the scene to be tested, obtain simulation data, and construct a network training simulation data set; execute the hybrid data network training module, use the artificial neural network of dual-weight neurons as the initial prediction network, and the initial prediction network is composed of a pre-training module and a measured optimization module. During training, the pre-training module first uses the network training simulation data set for pre-training to learn the preliminary relationship between propagation loss and other channel parameters, and then the measured optimization module uses the network training measured data set for optimization training, corrects the training results of the pre-training module, outputs the total weight matrix and total bias matrix of the trained initial prediction network, and proceeds to step S5;

[0011] S4, select one or more similar scenes with actual measurement conditions as reference scenes, and obtain the actual measurement data, simulation data of the reference scene and the simulation data of the scene to be measured, respectively construct the network training actual measurement data set of the reference scene, the network training simulation data set of the reference scene and the network training simulation data set of the scene to be measured; execute the network parameter migration module between scenes, use the artificial neural network of dual-weight neurons as the initial prediction network, as the neural network core of the scene migration, use the network training simulation data set of the reference scene and the network training simulation data set of the scene to be measured to train the neural network core respectively, and calculate the migration matrix between the weights and bias parameters of the neural network in the two scenes; then use the network training actual measurement data set of the reference scene to optimize the neural network core, and obtain the law of change of propagation loss; finally, based on the migration matrix, migrate the weight matrix and bias matrix of the reference scene after actual measurement optimization to the scene to be measured, and output the weight matrix and bias matrix of the trained initial prediction network;

[0012] S5, build a propagation loss prediction module with the same structure as the initial prediction network, assign the weight matrix and bias matrix of the initial prediction network to the propagation loss prediction module, generate the final propagation loss prediction module, input the new delay and angle parameters in the test scenario into the propagation loss prediction module, and predict the propagation loss.

[0013] Furthermore, in step S2, the process of using the measured data acquisition module of the scene to be tested to construct the measured data set for network training includes the following steps:

[0014] Fixed channel measurement transmitter position, uniformly or randomly select N receiver positions;

[0015] Collect and record the delay, azimuth angle of signal arrival, elevation angle of signal arrival, and propagation loss data under the line-of-sight (LoS) and non-line-of-sight (NLoS) paths at each receiver location;

[0016] Construct the measured data set for network training. The measured data sets under LoS path and NLoS path are expressed as:

[0017]

[0018]

[0019] in, and They respectively represent the delay and propagation loss data of the LoS / NLoS path at each receiver location obtained by actual measurement in the scenario to be tested; and They respectively represent the azimuth and elevation angles of the signal arrival angle of the NLoS path at each receiver position obtained by actual measurement in the scenario to be tested; i=1, 2, ..., N represents different receiver positions, and j represents different NLoS paths.

[0020] Furthermore, in step S2, the measured data network training module is executed, an artificial neural network (ANN) is built as an initial prediction network, the initial prediction network is trained using the network training measured data set, and the weight matrix and bias matrix of the trained initial prediction network are outputted. The process includes the following steps:

[0021] S21, build an artificial neural network as the initial prediction network. Under the LoS path, the time delay is used as the input of the initial prediction network; under the NLoS path, the time delay, the azimuth of the arrival angle, and the elevation angle of the arrival angle are used as the input of the initial prediction network. In both cases, the propagation loss is used as the output of the initial prediction network; the predicted values ​​of the propagation loss under the LoS path and the NLoS path are expressed as:

[0022]

[0023]

[0024] Among them, F ANN (·) is the transfer function of the neural network determined by the network structure; Represents the delay data of the LoS path input by the predicted network; and They represent the delay of the NLoS path, the azimuth angle of the signal arrival angle and the elevation angle data of the prediction network input respectively; w, b and σ(·) represent the weight matrix, bias matrix and activation function of the neurons in the neural network respectively;

[0025] The network structure is further expanded as follows:

[0026]

[0027]

[0028] in, represents the connection weight between the qth neuron in the kth layer of the network and the pth neuron in the previous layer, k=1, 2 represents the hidden layer and the output layer respectively, and the value range of p and q is [1, m]; represents the bias value of the qth neuron in the kth layer of the network; σ H / O (·)∈σ(·) represent the activation functions of the hidden layer and output layer in the network respectively;

[0029] S22, the number of neurons in the hidden layer M is calculated using the following formula:

[0030]

[0031] Among them, M In represents the number of neurons in the input layer; M Out represents the number of neurons in the output layer; a M is a constant in the range of [1, 10];

[0032] S23, set the activation function of the network. The hidden layer uses the rectified linear unit (ReLU) function, and the output layer uses the linear identity function, which are expressed as:

[0033] σ ReLU (x) = max(x, 0) (8)

[0034] σ Identity (x) = x (9)

[0035] S24, set the network loss function and add the L2 regularization term. The loss function is expressed as:

[0036]

[0037] in, represents the predicted value of the artificial neural network; represents the true value of the propagation loss obtained through actual measurement; H represents the number of samples in the training set; μ represents the regularization factor of L2 regularization; K represents the number of neural network layers; w (k) Represents the network weight matrix of the kth layer;

[0038] S25, use the network training measured data set to train the initial prediction network to make the network stable. The update of the network weight and bias parameters at each iteration is expressed as:

[0039]

[0040]

[0041] in, Represents partial derivative; η represents the learning rate of the neural network; ω represents the momentum factor, which is used to prevent the network parameters from falling into the local minimum when updating; and Represent the updated changes of network weights and bias values ​​respectively;

[0042] S26, after the training is completed, the weight matrix and bias matrix of the initial prediction network are output.

[0043] Furthermore, in step S3, the process of executing the simulation data acquisition module of the tested scene to acquire simulation data and constructing a network training simulation data set includes the following steps:

[0044] Reconstruct the three-dimensional scene of the scene to be tested, reproduce the scatterers in the scene including roads, buildings, trees, and water areas, and set the location of the transceiver;

[0045] Based on the reconstructed three-dimensional digital map, ray tracing simulation is performed on the scene to be tested, the propagation process of radio waves is simulated, and the time delay, azimuth angle of arrival, elevation angle of arrival and propagation loss channel parameters are calculated;

[0046] Construct a network training simulation data set. The simulation data sets under the LoS path and the NLoS path are expressed as follows:

[0047]

[0048]

[0049] in, and They respectively represent the delay and propagation loss data of the LoS / NLoS path at each receiver location obtained by simulation in the scenario to be tested; and They respectively represent the azimuth and elevation angles of the signal arrival angle of the NLoS path at each receiver position obtained by simulation in the scenario to be tested.

[0050] Furthermore, in step S3, the process of executing the hybrid data network training module includes the following steps:

[0051] S31, a Double-Weight-Neurons-based Artificial Neural Network (DWN-ANN) is used as the initial prediction network, which consists of a pre-training module (PTM) and a measurement optimization module (MOM); the predicted values ​​of propagation loss under LoS path and NLoS path are expressed as:

[0052]

[0053]

[0054] Among them, F DWN-ANN (·) is the transfer function of the neural network determined by the network structure; Represents the delay data of the LoS path input by the pre-training module / measured optimization module; and They represent the delay, azimuth and elevation angle data of the NLoS path input by the pre-training module / measured optimization module respectively; w Total With b Total Represent the total weight matrix and the total bias matrix of the artificial neural network with dual-weight neurons respectively;

[0055] S32, set the network loss function and add the L2 regularization term; the loss functions of the pre-training module and the measured optimization module are respectively expressed as:

[0056]

[0057]

[0058] in, and They represent the predicted values ​​of the pre-training module and the measured optimization module in DWN-ANN respectively; and Represent the true values ​​of propagation loss obtained through simulation and measurement respectively; w PTM(k) With w MOM(k) Represent the network weight matrices of the kth layer in the pre-training module and the measured optimization module respectively;

[0059] S33, a pre-training module of an artificial neural network with dual-weight neurons is trained using simulation data. The predicted values ​​under the LoS path and the NLoS path are respectively expressed as:

[0060]

[0061]

[0062] Among them, w PTM With b PTM They represent the weight matrix and bias matrix of the DWN-ANN pre-training module respectively;

[0063] S34, after the pre-training is completed, the measured data is used to train the measured optimization module of the artificial neural network with dual-weight neurons. The predicted values ​​under the LoS path and the NLoS path are respectively expressed as:

[0064]

[0065]

[0066] Among them, w MOM With b MOMThey represent the weight matrix and bias matrix of the DWN-ANN measured optimization module respectively; the final structure of the network is expanded as follows:

[0067]

[0068]

[0069] in, and They represent the connection weights and bias values ​​of neurons in the pre-training module, respectively. k=1 and 2 represent the hidden layer and the output layer, respectively. and They represent the connection weights and bias values ​​of neurons in the hidden layer and output layer in the measured optimization module respectively;

[0070] S35, after the training is completed, the total weight matrix and the total bias matrix of the artificial neural network with dual weight neurons are output.

[0071] Furthermore, in step S4, the process of respectively constructing a network training measured data set of the reference scene, a network training simulation data set of the reference scene, and a network training simulation data set of the scene to be tested includes the following steps:

[0072] Obtain the measured data of the reference scene and construct the measured data set for network training of the reference scene; assuming there are L reference scenes, the measured data sets under the LoS and NLoS paths in the lth reference scene are expressed as:

[0073]

[0074]

[0075] in, and They represent the delay and propagation loss data of the LoS / NLoS path at each receiver location obtained by actual measurement in the lth reference scenario; and represent the azimuth and elevation angles of the signal arrival angle of the NLoS path at each receiver position obtained by actual measurement in the lth reference scenario; l = {1, 2, ...L} represents different reference scenarios;

[0076] The simulation data under each reference scenario is obtained, and a network training simulation data set under the reference scenario is constructed; the simulation data sets under the LoS and NLoS paths under the lth reference scenario are respectively expressed as:

[0077]

[0078]

[0079] in, and represent the delay and propagation loss data of the LoS / NLoS path at each receiver location obtained by simulation in the lth reference scenario; and represent the azimuth and elevation angles of the signal arrival angle of the NLoS path at each receiver position obtained by simulation in the lth reference scenario respectively;

[0080] Construct a network training simulation dataset for the scenario to be tested.

[0081] Furthermore, in step S4, the process of executing the network parameter migration module between scenarios includes the following steps:

[0082] S41, an artificial neural network with dual-weight neurons is selected as the neural network core of scene migration. Different training data are used in different scenes, but each neural network core has the same structure;

[0083] S42, using the simulation data of each reference scene to pre-train the neural network kernel under the corresponding reference scene, and save the weight matrix and bias matrix of the pre-trained network, which are recorded as the reference original matrix; the reference original matrices of the neural network weight and bias under the lth reference scene are recorded as and

[0084] S43, using the measured data of each reference scene to optimize the neural network kernel under the corresponding reference scene, and saving the weight matrix and bias matrix of the network after optimization training, which are recorded as reference optimization matrices; the reference optimization matrices of network weights and biases under the lth reference scene are recorded as and

[0085] S44, in the test scenario, use the simulation data of the test scenario to train the neural network kernel, and save the weight matrix and bias matrix of the trained network, which are recorded as the original matrix w to be tested. Targ With b Targ ;

[0086] S45, based on the original matrix to be tested {w Targ , b Targ}, combined with the reference original matrix obtained by training L reference scenes The migration matrices between the reference scenario and the tested scenario are:

[0087]

[0088]

[0089] in, and Respectively represent the weight transfer matrix and bias transfer matrix between the neural network core of the scene to be tested and the neural network core of the lth reference scene;

[0090] S46, based on the obtained migration matrix Combined with the reference optimization matrix obtained through measured optimization training under L reference scenarios Calculate the final weight and bias matrix for the test scenario; the final weight matrix w Final With the bias matrix b Final for:

[0091]

[0092]

[0093] Among them, I l Represents the dataset size in the lth reference scenario;

[0094] S47, output the final weight matrix w after migration Final And the final bias matrix b Final .

[0095] Compared with the prior art, the present invention has the following beneficial effects:

[0096] First, the measured data-driven wireless channel propagation loss prediction method of the present invention solves the prediction accuracy problem of the neural network by optimizing and improving the traditional neural network structure and considering the influence of multiple variable factors such as time delay, azimuth and elevation angle of signal arrival angle, and can realize accurate prediction of propagation loss.

[0097] Second, the measured data-driven wireless channel propagation loss prediction method of the present invention comprehensively considers the impact of different measured data amounts on the network prediction performance, and proposes prediction schemes for three cases where the measured data is sufficient, insufficient, and missing. This solves the problem that traditional neural networks require a large amount of measured data during training, expands the scope of application of neural networks when training data is sparse or missing, and saves the manpower and material costs required for measurements. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 This is a flow chart of a method for predicting wireless channel propagation loss driven by measured data according to an embodiment of the present invention;

[0099] Figure 2 It is a schematic diagram of the structure of the initial prediction network built using artificial neural network;

[0100] Figure 3A schematic diagram of the structure of the initial prediction network built using an artificial neural network with dual-weight neurons;

[0101] Figure 4 The following is a diagram of the migration process. DETAILED DESCRIPTION

[0102] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0103] The embodiment of the present invention discloses a method for predicting wireless channel propagation loss driven by measured data, and the method for predicting wireless channel propagation loss comprises the following steps:

[0104] S1, the user sets the scene to be tested and determines whether the scene to be tested meets the actual measurement conditions; if the scene to be tested does not meet the actual measurement conditions, proceed to step S4; if the scene to be tested meets the actual measurement conditions, proceed to step S2;

[0105] S2, using the measured data acquisition module of the scene to be tested to build a network training measured data set, determine the threshold value of the number of acquisitions according to the scene to be tested, and judge whether the number of receiver positions is less than the threshold value. If it is less than, it is determined that the amount of measured data of the scene to be tested is insufficient, and the process goes to step S3. Otherwise, it is determined that the amount of measured data of the scene to be tested is sufficient, and the measured data network training module is executed to build an artificial neural network as an initial prediction network. The initial prediction network is trained using the network training measured data set, and the weight matrix and bias matrix of the trained initial prediction network are output, and the process goes to step S5;

[0106] S3, execute the simulation data acquisition module of the scene to be tested, obtain simulation data, and construct a network training simulation data set; execute the hybrid data network training module, use the artificial neural network of dual-weight neurons as the initial prediction network, and the initial prediction network is composed of a pre-training module and a measured optimization module. During training, the pre-training module first uses the network training simulation data set for pre-training to learn the preliminary relationship between propagation loss and other channel parameters, and then the measured optimization module uses the network training measured data set for optimization training, corrects the training results of the pre-training module, outputs the total weight matrix and total bias matrix of the trained initial prediction network, and proceeds to step S5;

[0107] S4, select one or more similar scenes with actual measurement conditions as reference scenes, and obtain the actual measurement data, simulation data of the reference scene and the simulation data of the scene to be measured, respectively construct the network training actual measurement data set of the reference scene, the network training simulation data set of the reference scene and the network training simulation data set of the scene to be measured; execute the network parameter migration module between scenes, use the artificial neural network of dual-weight neurons as the initial prediction network, as the neural network core of the scene migration, use the network training simulation data set of the reference scene and the network training simulation data set of the scene to be measured to train the neural network core respectively, and calculate the migration matrix between the weights and bias parameters of the neural network in the two scenes; then use the network training actual measurement data set of the reference scene to optimize the neural network core, and obtain the law of change of propagation loss; finally, based on the migration matrix, migrate the weight matrix and bias matrix of the reference scene after actual measurement optimization to the scene to be measured, and output the weight matrix and bias matrix of the trained initial prediction network;

[0108] S5, build a propagation loss prediction module with the same structure as the initial prediction network, assign the weight matrix and bias matrix of the initial prediction network to the propagation loss prediction module, generate the final propagation loss prediction module, input the new delay and angle parameters in the test scenario into the propagation loss prediction module, and predict the propagation loss.

[0109] The overall process of the program is as follows Figure 1 As shown, the following steps are included:

[0110] Step 1: The user sets the scene to be tested and determines whether the scene to be tested meets the actual measurement conditions. If the scene to be tested does not meet the actual measurement conditions, jump directly to step 5; if the scene to be tested meets the actual measurement conditions, execute the actual measurement data acquisition module of the scene to be tested. The specific implementation steps are as follows:

[0111] 1.1) Fixed channel measurement transmitter position, uniformly or randomly select N receiver positions.

[0112] 1.2) Collect and record the delay, azimuth angle of signal arrival, elevation angle of signal arrival, and propagation loss data under the line-of-sight (LoS) and non-line-of-sight (NLoS) paths at each receiver location.

[0113] 1.3) Construct the measured data set for network training. The measured data sets under LoS and NLoS paths can be expressed as

[0114]

[0115]

[0116] in, and They respectively represent the delay and propagation loss data under the LoS / NLoS path at each receiver location obtained by actual measurement in the scenario to be tested; and They respectively represent the azimuth and elevation angles of the signal arrival angle of the NLoS path at each receiver position obtained by actual measurement in the scenario to be tested; i=1, 2, ..., N represents different receiver positions, and j represents different NLoS paths.

[0117] Step 2: Determine the threshold value of the number of acquisitions according to the scene to be tested. When N < the threshold value, it is considered that the measured data is insufficient, and jump directly to step 3; when N > the threshold value, it is determined that the measured data volume is sufficient, and execute the measured data network training module. The specific implementation steps are as follows:

[0118] 2.1) The present invention builds an artificial neural network (ANN) as the initial prediction network, and the network structure is as follows Figure 2 As shown. In the case of LoS, the time delay is used as the input of the ANN; in the case of NLoS, the time delay, the azimuth of the arrival angle, and the elevation angle of the arrival angle are used as the input of the ANN. At the same time, both use the propagation loss as the output of the ANN. At this time, the predicted values ​​of the propagation loss in the LoS and NLoS cases can be expressed as

[0119]

[0120]

[0121] Among them, F ANN (·) is the transfer function of the neural network determined by the network structure; Represents the delay data of the LoS path input by the predicted network; and Respectively represent the delay of the NLoS path, the azimuth and elevation angle data of the signal arrival angle of the prediction network input; w, b, σ(·) respectively represent the weight matrix, bias matrix and activation function of the neurons in the neural network. Taking an ANN with an input layer, a hidden layer containing M neurons and an output layer as an example, the network structure can be further expanded as follows:

[0122]

[0123]

[0124] in, represents the connection weight between the qth neuron in the kth layer of the network and the pth neuron in the previous layer, k=1, 2 represents the hidden layer and the output layer respectively, and the value range of p and q is [1, m]; represents the bias value of the qth neuron in the kth layer of the network; σ H / O (·)∈σ(·) represent the activation functions of the hidden layer and the output layer in the network respectively.

[0125] 2.2) The number of neurons in the hidden layer M can be calculated as

[0126]

[0127] Among them, M In represents the number of neurons in the input layer; M Out represents the number of neurons in the output layer; a M is a constant in the range of [1, 10]. The number of network layers and neurons can be adjusted according to specific application scenarios.

[0128] 2.3) Set the activation function of the network. The hidden layer uses the rectified linear unit (ReLU) function, and the output layer uses the linear identity function. The two are expressed as

[0129] σ ReLU (x) = max(x, 0) (40)

[0130] σ Identity (x) = x (41)

[0131] 2.4) Set the network loss function and add the L2 regularization term for subsequent network training. The loss function can be expressed as

[0132]

[0133] in, Represents the predicted value of the neural network; represents the true value of the propagation loss obtained through actual measurement; H represents the number of samples in the training set; μ represents the regularization factor of L2 regularization; K represents the number of neural network layers; w (k) Represents the network weight matrix of the kth layer.

[0134] 2.5) Use the measured data set obtained in the first step to train the ANN initial prediction network to make the network stable. The update of the network weight and bias parameters at each iteration can be expressed as

[0135]

[0136]

[0137] in, Represents partial derivative; η represents the learning rate of the neural network; ω represents the momentum factor, which can prevent the network parameters from falling into the local minimum when updating; and Represent the updated changes of network weights and bias values ​​respectively.

[0138] 2.6) After training is completed, output the ANN weight matrix and bias matrix and jump to step 7.

[0139] Step 3: If the measured data is insufficient, execute the simulation data acquisition module of the tested scene to obtain simulation data for pre-training of the neural network. The specific implementation steps are as follows:

[0140] 3.1) Reconstruct the three-dimensional scene of the scene to be tested, reproduce the scatterers such as roads, buildings, trees, water areas, etc. in the scene, and set the location of the transceiver.

[0141] 3.2) Based on the reconstructed three-dimensional digital map, ray tracing simulation is performed on the scene to be tested, the propagation process of the radio waves is simulated, and channel parameters such as delay, azimuth of arrival angle, elevation angle of arrival angle, and propagation loss are calculated.

[0142] 3.3) Constructing network training simulation data set. The simulation data sets under LoS and NLoS paths can be expressed as

[0143]

[0144]

[0145] in, and They respectively represent the delay and propagation loss data under the LoS / NLoS path at each receiver location obtained by simulation in the scenario to be tested; and They respectively represent the azimuth and elevation angles of the signal arrival angle of the NLoS path at each receiver position obtained by simulation in the scenario to be tested.

[0146] Step 4: Execute the hybrid data network training module when the measured data is insufficient. The specific implementation steps are as follows:

[0147] 4.1) The present invention adopts a double-weighted artificial neural network (DWN-ANN) as the initial prediction network. The network consists of a pre-training module (PTM) and a measurement optimization module (MOM). The network structure is as follows: Figure 3 As shown. The pre-training module uses simulation data for pre-training to learn the approximate relationship between propagation loss and other channel parameters, reducing the network's demand for measured data. The measured optimization module uses a small amount of measured data for optimization training, accurately corrects the preliminary training results, and improves the network's prediction accuracy. In the LoS case, the time delay is used as the input of the neural network; in the NLoS case, the time delay, the azimuth of the arrival angle, and the pitch angle of the arrival angle are used as the input of the neural network. At the same time, both use the propagation loss as the output of the neural network. At this time, the predicted values ​​of the propagation loss in the LoS and NLoS cases can be expressed as

[0148]

[0149]

[0150] Among them, F DWN-ANN (·) is the transfer function of the neural network determined by the network structure; Represents the delay data of the LoS path input by the pre-training module / measured optimization module; and They represent the delay, azimuth and elevation angle data of the NLoS path input by the pre-training module / measured optimization module respectively; w Total With b Total Represent the total weight matrix and total bias matrix of DWN-ANN respectively. The calculation of the number of hidden layer neurons M and the setting of activation function are the same as in the second step.

[0151] 4.2) Set the network loss function and add the L2 regularization term. The loss functions of the pre-training module and the measured optimization module can be expressed as

[0152]

[0153]

[0154] in, and They represent the predicted values ​​of the pre-training module and the measured optimization module in DWN-ANN respectively; and Represent the true values ​​of propagation loss obtained through simulation and measurement respectively; w PTM(k) With w MOM(k) They represent the network weight matrices of the kth layer in the pre-training module and the measured optimization module respectively.

[0155] 4.3) Using the simulation data in the third step to train the pre-training module of DWN-ANN, the predicted values ​​in the LoS and NLoS cases can be expressed as

[0156]

[0157]

[0158] Among them, w PTM With b PTM They represent the weight matrix and bias matrix of the DWN-ANN pre-training module respectively.

[0159] 4.4) After pre-training, a small amount of measured data from the first step is used to train the measured optimization module of DWN-ANN. The predicted values ​​in the LoS and NLoS cases can be expressed as

[0160]

[0161]

[0162] Among them, w MOM With b MOM They represent the weight matrix and bias matrix of the DWN-ANN measured optimization module. Taking a DWN-ANN with one input layer, one hidden layer containing M neurons, and one output layer as an example, the final structure of the network can be specifically expanded as follows:

[0163]

[0164]

[0165] in, and They represent the connection weights and bias values ​​of neurons in the pre-training module, respectively. k=1 and 2 represent the hidden layer and the output layer, respectively. and They represent the connection weights and bias values ​​of the hidden layer and output layer neurons in the measured optimization module respectively.

[0166] 4.5) After the training is completed, the total weight matrix and total bias matrix of the trained DWN-ANN are output, and jump to step 7.

[0167] Step 5: When the scene to be tested does not meet the actual measurement conditions, select one or more similar scenes that meet the actual measurement conditions as reference scenes, and obtain the actual measurement data, simulation data of the reference scenes and simulation data of the scene to be tested. The specific implementation steps are as follows:

[0168] 5.1) Execute the reference scene measured data acquisition module, adopt the method of the measured scene measured data acquisition module in the first step, obtain the measured data of the reference scene, and construct the network training measured data set of the reference scene. The reference scene of the present invention can be one or more. Assuming there are L reference scenes, the measured data sets under the LoS and NLoS paths in the lth reference scene are respectively expressed as

[0169]

[0170]

[0171] in, and They represent the delay and propagation loss data of the LoS / NLoS path at each receiver location obtained by actual measurement in the lth reference scenario; and They respectively represent the azimuth and elevation angles of the signal arrival angle of the NLoS path at each receiver position obtained by actual measurement in the lth reference scenario; l = {1, 2, ...L} represents different reference scenarios.

[0172] 5.2) Execute the reference scenario simulation data acquisition module, use the method of the test scenario simulation data acquisition module in the third step to obtain the simulation data under each reference scenario, and construct the network training simulation data set under the reference scenario. The simulation data sets under the LoS and NLoS paths under the lth reference scenario can be expressed as

[0173]

[0174]

[0175] in, and They respectively represent the delay and propagation loss data under the LoS / NLoS path at each receiver location obtained by simulation in the lth reference scenario. and They respectively represent the azimuth and elevation angles of the signal arrival angle under the NLoS path at each receiver position obtained by simulation in the lth reference scenario.

[0176] 5.3) Execute the simulation data acquisition module for the test scenario in the third step to construct a network training simulation data set for the test scenario.

[0177] Step 6: Execute the network parameter migration module between scenarios when measured data is missing, build a neural network kernel for migration, use the simulated data of the scenario to be tested and the reference scenario to train the neural network kernel respectively, and calculate the migration matrix between the weights and bias parameters of the neural network in the two scenarios; then, use the measured data of the reference scenario to optimize the neural network kernel and obtain the law of propagation loss change; finally, based on the migration matrix, migrate the weight matrix and bias matrix of the reference scenario after measured optimization to the scenario to be tested. The migration process is as follows: Figure 4 As shown, the specific implementation steps are as follows:

[0178] 6.1) Select DWN-ANN in the fourth step as the neural network core for scene migration. Different scenes use different training data, but each neural network core has the same structure.

[0179] 6.2) In each reference scenario, use the simulation data in step 5.2 to pre-train the neural network kernel in that scenario, and save the weight matrix and bias matrix of the pre-trained network, which are recorded as the reference original matrix. The reference original matrices of the neural network weight and bias in the lth reference scenario are recorded as and

[0180] 6.3) In each reference scenario, use the measured data in step 5.1 to optimize the neural network kernel in that scenario, and save the weight matrix and bias matrix of the network after optimization training, which are recorded as reference optimization matrices. The reference optimization matrices of network weights and biases in the lth reference scenario are recorded as and

[0181] 6.4) In the test scenario, use the simulation data in step 5.3 to train the neural network kernel, and save the weight matrix and bias matrix of the trained network, which are recorded as the original matrix w to be tested. Targ With b Targ .

[0182] 6.5) Based on the original matrix {w Targ , b Targ}, combined with the reference original matrix obtained by training L reference scenes The migration matrix between the reference scene and the scene to be tested can be calculated as

[0183]

[0184]

[0185] in, and They respectively represent the weight transfer matrix and bias transfer matrix between the neural network kernel of the tested scene and the neural network kernel of the lth reference scene.

[0186] 6.6) Based on the obtained migration matrix Combined with the reference optimization matrix obtained through measured optimization training under L reference scenarios The final weight and bias matrix under the test scenario can be calculated to achieve the migration of the propagation loss prediction network. The final weight matrix w Final With the bias matrix b Final It can be calculated as

[0187]

[0188]

[0189] Among them, I1 represents the dataset size in the lth reference scenario.

[0190] 6.7) Output the final weight matrix w after migration Final And the final bias matrix b Final , and jump to step 7.

[0191] Step 7: Execute the propagation loss prediction module at any transmitting and receiving position in the test scenario to predict the propagation loss. The specific implementation steps are as follows:

[0192] 7.1) Build an initial ANN or DWN-ANN with the same network structure as the previous training.

[0193] 7.2) Receive the neural network weight matrix and bias matrix outputted in the second step, the fourth step, or the sixth step, and assign them to the corresponding initial ANN or DWN-ANN to generate the final propagation loss prediction network.

[0194] 7.3) The user inputs the new delay and angle parameters of the scenario to be tested into the prediction network, and the network outputs the corresponding predicted value of the propagation loss.

[0195] Examples

[0196] The following is a clear and complete description of the technical solution in conjunction with a specific implementation example. This implementation example includes the following steps:

[0197] Step 1: Select an area of ​​100m×100m as the scene to be tested, and assume that the scene to be tested has two conditions: one with actual measurement conditions and one without actual measurement conditions. If the scene to be tested does not have actual measurement conditions, jump directly to step 5; if the scene to be tested has actual measurement conditions, execute the actual measurement data acquisition module of the scene to be tested. The specific implementation steps are as follows:

[0198] 1.1) Fixed channel measurement The transmitter position is located at the center of the scene to be measured, and 64 receiver positions are evenly selected.

[0199] 1.2) Collect and record the delay, azimuth of the signal arrival angle, elevation of the signal arrival angle, and propagation loss data under the LoS and NLoS paths at each receiver location. Since each NLoS path has similar propagation characteristics, this case only uses the ground reflection path as an example to verify the correctness of the prediction method.

[0200] 1.3) Construct the measured data set for network training. The measured data sets under LoS and NLoS paths are shown in Table 1 and Table 2 respectively.

[0201] Table 1 Measured data sets under LoS path

[0202]

[0203] Table 2 Measured datasets under NLoS paths

[0204]

[0205] Step 2: According to the scene to be tested, the threshold value of the number of acquisitions is determined to be 50. When N < the threshold value, it is considered that the measured data is insufficient, and jump directly to step 3; when N > the threshold value, it is determined that the measured data is sufficient. At this time, there are 64 sets of measured data, which is greater than the set threshold value, and the measured data network training module is executed. The specific implementation steps are as follows:

[0206] 2.1) Build ANN as the initial prediction network. The predicted values ​​of propagation loss in LoS and NLoS cases can be expressed as

[0207]

[0208]

[0209] The network structure can be further expanded into

[0210]

[0211]

[0212] 2.2) The number of neurons in the hidden layer M can be calculated as

[0213]

[0214] Since the number of parameters of the input layer in the LoS and NLoS cases is 1 and 3 respectively, and the number of parameters of the output layer is 1, the number of neurons in the hidden layer can be set to 5 according to the above formula.

[0215] 2.3) Set the activation function of the network. The hidden layer uses the ReLU function and the output layer uses the linear identity function. The two are expressed as

[0216] σ ReLU (x) = max(x, 0) (70)

[0217] σ Identity (x) = x (71)

[0218] 2.4) Set the network loss function and add the L2 regularization term for subsequent network training. The loss function can be expressed as

[0219]

[0220] Among them, the number of samples H in the training set is set to 64; the number of neural network layers K is set to 2; and the initial value of the regularization factor μ of L2 regularization is set to 0.01.

[0221] 2.5) Use the measured data set obtained in the first step to train the ANN initial prediction network to make the network stable. The update of the network weight and bias parameters at each iteration can be expressed as

[0222]

[0223]

[0224] Among them, the learning rate η of the neural network is set to 0.01; the momentum factor ω is set to 0.9, which can prevent the network parameters from falling into the local minimum when updating.

[0225] 2.6) After the training is completed, the weight matrix and bias matrix of the ANN are output. The results are shown in Table 3, and jump to step 7.

[0226] Table 3 Network parameters after ANN training

[0227]

[0228] Step 3: In the case of insufficient measured data, assume that only 32 sets of measured data are obtained, which is less than the set threshold value of 50. At this time, the simulation data acquisition module of the scene to be tested is executed to obtain simulation data for pre-training of the neural network. The specific implementation steps are as follows:

[0229] 3.1) Reconstruct the scene to be tested in three dimensions, reproduce the scatterers such as roads, buildings, trees, water areas, etc. in the scene, and set the location of the transceiver. In this case, a total of 1 transmitter and 100 receivers are set.

[0230] 3.2) Based on the reconstructed three-dimensional digital map, ray tracing simulation is performed on the scene to be tested, the propagation process of the radio waves is simulated, and channel parameters such as delay, azimuth of arrival angle, elevation angle of arrival angle, and propagation loss are calculated.

[0231] 3.3) Construct network training simulation data sets. The simulation data sets under LoS and NLoS paths are shown in Table 4 and Table 5 respectively.

[0232] Table 4 Simulation data set under LoS path

[0233]

[0234] Table 5 Simulation data set under NLoS path

[0235]

[0236] Step 4: Execute the hybrid data network training module when the measured data is insufficient. The specific implementation steps are as follows:

[0237] 4.1) DWN-ANN is used as the initial prediction network. The predicted values ​​of propagation loss in LoS and NLoS cases can be expressed as

[0238]

[0239]

[0240] Among them, the number of neurons in the hidden layer is set to 5, and the hidden layer and output layer use ReLU function and linear identity function as activation functions respectively.

[0241] 4.2) Set the network loss function and add the L2 regularization term. The loss functions of the pre-training module and the measured optimization module can be expressed as

[0242]

[0243]

[0244] Among them, the number of training set samples H of the pre-training module and the measured optimization module is set to 100 and 32 respectively; the number of neural network layers K is set to 2; the initial value of the regularization factor μ of L2 regularization is set to 0.01; the learning rate is set to 0.01; and the momentum factor during parameter update is set to 0.9.

[0245] 4.3) Using the simulation data in the third step to train the pre-training module of DWN-ANN, the predicted values ​​in the LoS and NLoS cases can be expressed as

[0246]

[0247]

[0248] 4.4) After pre-training, a small amount of measured data from the first step is used to train the measured optimization module of DWN-ANN. The predicted values ​​in the LoS and NLoS cases can be expressed as

[0249]

[0250]

[0251] Taking a DWN-ANN with one input layer, one hidden layer containing M neurons and one output layer as an example, the final structure of the network can be specifically expanded as follows:

[0252]

[0253]

[0254] 4.5) After the training is completed, the total weight matrix and total bias matrix of the trained DWN-ANN are output. The results are shown in Table 6, and jump to step 7.

[0255] Table 6 Network parameters after DWN-ANN training

[0256]

[0257]

[0258] Step 5: When the scene to be tested does not meet the actual measurement conditions, select one or more similar scenes that meet the actual measurement conditions as reference scenes, and obtain the actual measurement data, simulation data of the reference scenes and simulation data of the scene to be tested. The specific implementation steps are as follows:

[0259] 5.1) Execute the reference scene measured data acquisition module, adopt the method of the measured data acquisition module of the scene to be tested in the first step, obtain the measured data of the reference scene, and construct the network training measured data set of the reference scene. The reference scene of the present invention can be one or more. In this case, only one reference scene is selected for migration prediction, and the measured data sets of the LoS and NLoS paths under the reference scene are shown in Table 7 and Table 8 respectively.

[0260] Table 7 Measured data set under reference scenario LoS ​​path

[0261]

[0262] Table 8 Measured datasets under reference scenario NLoS path

[0263]

[0264] 5.2) Execute the reference scenario simulation data acquisition module, use the method of the test scenario simulation data acquisition module in the third step to obtain the simulation data under each reference scenario, and construct the network training simulation data set under the reference scenario. The simulation data sets of the LoS and NLoS paths under the reference scenario are shown in Tables 9 and 10 respectively.

[0265] Table 9 Simulation data set under reference scenario LoS ​​path

[0266]

[0267] Table 10 Simulation dataset under reference scenario NLoS path

[0268]

[0269] 5.3) Execute the simulation data acquisition module for the test scenario in the third step to construct a network training simulation data set for the test scenario.

[0270] Step 6: Execute the network parameter migration module between scenarios when measured data is missing, build a neural network kernel for migration, use the simulation data of the scenario to be tested and the reference scenario to train the neural network kernel respectively, and calculate the migration matrix between the weights and bias parameters of the neural network in the two scenarios; then, use the measured data of the reference scenario to optimize the neural network kernel and obtain the law of propagation loss change; finally, based on the migration matrix, migrate the weight matrix and bias matrix of the reference scenario after measured optimization to the scenario to be tested. The specific implementation steps are as follows:

[0271] 6.1) Select DWN-ANN in the fourth step as the neural network kernel for scene migration.

[0272] 6.2) Use the simulation data in step 5.2 to pre-train the neural network kernel in the reference scenario, and save the weight matrix and bias matrix of the pre-trained network, which are recorded as the reference original matrix. The reference original matrices of the neural network weights and biases in the reference scenario are recorded as w Base With b Base .

[0273] 6.3) Use the measured data in step 5.1 to optimize the neural network kernel under the reference scenario, and save the weight matrix and bias matrix of the optimized network after training, which are recorded as reference optimization matrices. The reference optimization matrices of network weights and biases under the reference scenario are recorded as w Opti With b Opti .

[0274] 6.4) In the test scenario, use the simulation data in step 5.3 to train the neural network kernel, and save the weight matrix and bias matrix of the trained network, which are recorded as the original matrix w to be tested. TargWith b Targ .

[0275] 6.5) Based on the original matrix {w Targ , b Targ}, combined with the reference scene training obtained reference original matrix {w Base , b Base The migration matrix between the reference scene and the scene to be tested can be calculated as

[0276]

[0277]

[0278] 6.6) Based on the obtained migration matrix {T w , T b}, combined with the reference optimization matrix {w Opti , b Opti}, the final weight and bias matrix under the test scenario can be calculated to achieve the migration of the propagation loss prediction network. The final weight matrix w Final With the bias matrix b Final It can be calculated as

[0279]

[0280]

[0281] Since only one reference scenario is selected in this case, the above formula can be simplified to

[0282] w Final =w Opti ·T w (89)

[0283] b Final =b Opti ·T b (90)

[0284] 6.7) Output the final weight matrix w after migration Final And the final bias matrix b Final , the results are shown in Table 11, and jump to step 7.

[0285] Table 11 Final network parameters after migration

[0286]

[0287] Step 7: Execute the propagation loss prediction module at any transmitting and receiving position in the test scenario to predict the propagation loss. The specific implementation steps are as follows:

[0288] 7.1) Build an initial ANN or DWN-ANN with the same network structure as the previous training.

[0289] 7.2) Receive the neural network weight matrix and bias matrix outputted in the second step, the fourth step, or the sixth step, and assign them to the corresponding initial ANN or DWN-ANN to generate the final propagation loss prediction network.

[0290] 7.3) The user inputs 50 sets of new delay and angle parameters under the three scenarios into the prediction network, and the network outputs the corresponding predicted values ​​of propagation loss. The predicted values ​​of propagation loss under the conditions of sufficient, insufficient and missing measured data are shown in Table 12-1, Table 12-2, Table 13-1, Table 13-2, Table 14-1 and Table 14-2 respectively.

[0291] Table 12-1 LoS path prediction results when there is sufficient measured data

[0292]

[0293] Table 12-2 NLoS path prediction results when there is sufficient measured data

[0294]

[0295] Table 13-1 LoS path prediction results when measured data is insufficient

[0296]

[0297] Table 13-2 NLoS path prediction results when measured data is insufficient

[0298]

[0299]

[0300] Table 14-1 LoS path prediction results when measured data is missing

[0301]

[0302] Table 14-2 NLoS path prediction results when measured data is missing

[0303]

[0304] It can be seen from the results in Table 12-1, Table 12-2, Table 13-1, Table 13-2, Table 14-1 and Table 14-2 that the proposed method for predicting wireless channel propagation loss driven by measured data can achieve accurate prediction of propagation loss in the cases of sufficient, insufficient and missing measured data, verifying the correctness of the proposed method.

[0305] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for predicting wireless channel propagation loss driven by measured data, characterized in that: The wireless channel propagation loss prediction method comprises the following steps: S1, the user sets the scene to be tested and determines whether the scene to be tested meets the actual measurement conditions; if the scene to be tested does not meet the actual measurement conditions, proceed to step S4; if the scene to be tested meets the actual measurement conditions, proceed to step S2; S2, using the measured data acquisition module of the scene to be tested to build a network training measured data set, determine the threshold value of the number of acquisitions according to the scene to be tested, and judge whether the number of receiver positions is less than the threshold value. If it is less than, it is determined that the amount of measured data of the scene to be tested is insufficient, and the process goes to step S3. Otherwise, it is determined that the amount of measured data of the scene to be tested is sufficient, and the measured data network training module is executed to build an artificial neural network as an initial prediction network. The initial prediction network is trained using the network training measured data set, and the weight matrix and bias matrix of the trained initial prediction network are output, and the process goes to step S5; S3, execute the simulation data acquisition module of the scene to be tested, obtain simulation data, and construct a network training simulation data set; execute the hybrid data network training module, use the artificial neural network of dual-weight neurons as the initial prediction network, and the initial prediction network is composed of a pre-training module and a measured optimization module. During training, the pre-training module first uses the network training simulation data set for pre-training to learn the preliminary relationship between propagation loss and other channel parameters, and then the measured optimization module uses the network training measured data set for optimization training, corrects the training results of the pre-training module, outputs the total weight matrix and total bias matrix of the trained initial prediction network, and proceeds to step S5; S4, select one or more similar scenes with actual measurement conditions as reference scenes, and obtain the actual measurement data, simulation data of the reference scene and the simulation data of the scene to be measured, respectively construct the network training actual measurement data set of the reference scene, the network training simulation data set of the reference scene and the network training simulation data set of the scene to be measured; execute the network parameter migration module between scenes, use the artificial neural network of dual-weight neurons as the initial prediction network, as the neural network core of the scene migration, use the network training simulation data set of the reference scene and the network training simulation data set of the scene to be measured to train the neural network core respectively, and calculate the migration matrix between the weights and bias parameters of the neural network in the two scenes; then use the network training actual measurement data set of the reference scene to optimize the neural network core, and obtain the law of change of propagation loss; finally, based on the migration matrix, migrate the weight matrix and bias matrix of the reference scene after actual measurement optimization to the scene to be measured, and output the weight matrix and bias matrix of the trained initial prediction network; S5, build a propagation loss prediction module with the same structure as the initial prediction network, assign the weight matrix and bias matrix of the initial prediction network to the propagation loss prediction module, generate a final propagation loss prediction module, input the new delay and angle parameters under the test scenario to the propagation loss prediction module, and predict the propagation loss; In step S2, the process of using the measured data acquisition module of the scene to be tested to construct the measured data set for network training includes the following steps: Fixed channel measurement transmitter position, uniformly or randomly select N receiver positions; Collect and record the delay, azimuth angle of signal arrival angle, elevation angle of signal arrival angle and propagation loss data under line-of-sight path and non-line-of-sight path at each receiver position; Construct network training measured data sets, measured data sets under LoS paths and NLoS paths and Respectively expressed as: in, and They respectively represent the delay and propagation loss data of the LoS / NLoS path at each receiver location obtained by actual measurement in the scenario to be tested; and They represent the azimuth and elevation angles of the signal arrival angle of the NLoS path at each receiver position obtained by actual measurement in the scenario to be tested; i=1,2,...,N represents different receiver positions, and j represents different NLoS paths; In step S2, the measured data network training module is executed, an artificial neural network is built as an initial prediction network, the initial prediction network is trained using the network training measured data set, and the weight matrix and bias matrix of the trained initial prediction network are outputted. The process includes the following steps: S21, build an artificial neural network as the initial prediction network. Under the LoS path, the time delay is used as the input of the initial prediction network; under the NLoS path, the time delay, the azimuth of the arrival angle, and the elevation angle of the arrival angle are used as the input of the initial prediction network. In both cases, the propagation loss is used as the output of the initial prediction network; the predicted values ​​of the propagation loss under the LoS path and the NLoS path are expressed as: in, and represent the predicted values ​​of propagation loss under LoS path and NLoS path respectively; F ANN (·) is the transfer function of the neural network determined by the network structure; Represents the delay data of the LoS path input by the predicted network; and They represent the delay of the NLoS path, the azimuth angle of the signal arrival angle and the elevation angle data of the prediction network input respectively; w, b and σ(·) represent the weight matrix, bias matrix and activation function of the neurons in the neural network respectively; The network structure is further expanded as follows: in, represents the connection weight between the qth neuron in the kth layer of the network and the pth neuron in the previous layer, k=1,2 represents the hidden layer and the output layer respectively, and the value range of p and q is [1,m]; represents the bias value of the qth neuron in the kth layer of the network; σ H (·) and σ O (·) represent the activation functions of the hidden layer and the output layer in the network, σ H (·)∈(·),σ O (·)∈σ(·); m=1,2,...,M, M is the number of neurons in the hidden layer; S22, the number of neurons in the hidden layer M is calculated using the following formula: Among them, M In represents the number of neurons in the input layer; M Out represents the number of neurons in the output layer; a M is a constant in the range of [1,10]; S23, set the activation function of the network, the hidden layer uses the rectified linear unit function, and the output layer uses the linear identity function, which are expressed as: σ ReLU (x)=max(x,0) s Identity (x)=x In the formula, σ ReLU (x) is the activation function of the hidden layer; σ Identity (x) is the activation function of the output layer; x represents the input value of the activation function, which is the calculation result output by each neuron; S24, set the loss function of the artificial neural network and add the L2 regularization term. The loss function of the artificial neural network is E LoS / NLoS,j It is expressed as: in, represents the predicted value of the artificial neural network; represents the true value of the propagation loss obtained through actual measurement; h = 1, 2, ..., H, H represents the number of samples in the network training measured data set; μ represents the regularization factor of L2 regularization; K represents the number of neural network layers; w (k) Represents the network weight matrix of the kth layer; S25, use the network training measured data set to train the initial prediction network to make the network stable. The update of the network weight and bias parameters at each iteration is expressed as: in, Represents partial derivative; η represents the learning rate of the neural network; ω represents the momentum factor, which is used to prevent the network parameters from falling into the local minimum when updating; and Represent the updated changes of network weights and bias values ​​respectively; S26, after the training is completed, the weight matrix and bias matrix of the initial prediction network are output; In step S3, the process of executing the simulation data acquisition module of the tested scene to acquire simulation data and constructing a network training simulation data set includes the following steps: Reconstruct the three-dimensional scene of the scene to be tested, reproduce the scatterers in the scene including roads, buildings, trees, and water areas, and set the location of the transceiver; Based on the reconstructed three-dimensional digital map, ray tracing simulation is performed on the scene to be tested, the propagation process of radio waves is simulated, and the time delay, azimuth angle of arrival, elevation angle of arrival and propagation loss channel parameters are calculated; Construct network training simulation data set, simulation data set under LoS path and NLoS path and Respectively expressed as: in, and They respectively represent the delay and propagation loss data of the LoS / NLoS path at each receiver location obtained by simulation in the scenario to be tested; and represent the azimuth and elevation angles of the signal arrival angle of the NLoS path at each receiver position obtained by simulation in the scenario to be tested; i=1,2,...,N represents different receiver positions, and j represents different NLoS paths; In step S3, the process of executing the hybrid data network training module includes the following steps: S31, using an artificial neural network with dual-weight neurons as the initial prediction network, which consists of a pre-training module and a measured optimization module; the predicted values ​​of propagation loss under LoS path and NLoS path and Respectively expressed as: Among them, F DWN-ANN (·) is the transfer function of the artificial neural network composed of dual-weight neurons; Represents the delay data of the LoS path input by the pre-training module / measured optimization module; and They represent the delay, azimuth and elevation angle data of the NLoS path input by the pre-training module / measured optimization module respectively; w Total With b Total Represent the total weight matrix and the total bias matrix of the artificial neural network with dual-weight neurons respectively; σ(·) represents the activation function of the neurons in the neural network; S32, set the network loss function, add L2 regularization term; loss function of pre-training module and measured optimization module and Respectively expressed as: in, and They represent the predicted values ​​of the pre-training module and the measured optimization module in DWN-ANN respectively; and Represent the true values ​​of propagation loss obtained through simulation and measurement respectively; h=1,2,...,H, H represents the number of samples in the measured data set for network training; μ represents the regularization factor of L2 regularization; K represents the number of neural network layers; w PTM(k) With w MOM(k) Represent the network weight matrices of the kth layer in the pre-training module and the measured optimization module respectively; S33, a pre-training module of an artificial neural network with dual-weight neurons is trained using simulation data. The predicted values ​​under the LoS path and the NLoS path are respectively expressed as: Among them, w PTM With b PTM The weight matrix and bias matrix of the pre-trained module of the artificial neural network representing the dual-weight neurons respectively; S34, after the pre-training, the measured optimization module of the artificial neural network with dual-weight neurons is trained using measured data, and the predicted values ​​under the LoS path and NLoS path and Respectively expressed as: Among them, w MOM With b MOM They represent the weight matrix and bias matrix of the DWN-ANN measured optimization module respectively; S35, after the training is completed, output the total weight matrix and the total bias matrix of the artificial neural network with dual weight neurons; In step S4, the process of respectively constructing a network training measured data set for the reference scenario, a network training simulation data set for the reference scenario, and a network training simulation data set for the scenario to be tested includes the following steps: Obtain the measured data of the reference scene and construct the measured data set for network training of the reference scene; assuming there are L reference scenes, the measured data set under the LoS and NLoS paths in the lth reference scene and Respectively expressed as: in, and They represent the delay and propagation loss data of the LoS / NLoS path at each receiver location obtained by actual measurement in the lth reference scenario; and represent the azimuth and elevation angles of the signal arrival angle of the NLoS path at each receiver position obtained by actual measurement in the lth reference scenario; l = {1, 2, …L} represents different reference scenarios; Get the simulation data for each reference scenario and build the network training simulation data set for the reference scenario; the simulation data set for the LoS and NLoS paths for the lth reference scenario and Respectively expressed as: in, and represent the delay and propagation loss data of the LoS / NLoS path at each receiver location obtained by simulation in the lth reference scenario; and represent the azimuth and elevation angles of the signal arrival angle of the NLoS path at each receiver position obtained by simulation in the lth reference scenario respectively; Construct a network training simulation data set for the scenario to be tested; In step S4, the process of executing the network parameter migration module between scenarios includes the following steps: S41, an artificial neural network with dual-weight neurons is selected as the neural network core of scene migration. Different training data are used in different scenes, but each neural network core has the same structure; S42, using the simulation data of each reference scene to pre-train the neural network kernel under the corresponding reference scene, and save the weight matrix and bias matrix of the pre-trained network, which are recorded as the reference original matrix; the reference original matrices of the neural network weight and bias under the lth reference scene are recorded as and S43, using the measured data of each reference scene to optimize the neural network kernel under the corresponding reference scene, and saving the weight matrix and bias matrix of the network after optimization training, which are recorded as reference optimization matrices; the reference optimization matrices of network weights and biases under the lth reference scene are recorded as and S44, in the test scenario, use the simulation data of the test scenario to train the neural network kernel, and save the weight matrix and bias matrix w of the trained network Targ With b Targ , denoted as the original matrix to be tested {w Targ ,b Targ }; S45, based on the original matrix to be tested {w Targ ,b Targ }, combined with the reference original matrix obtained by training L reference scenes The migration matrices between the reference scenario and the tested scenario are: in, and Respectively represent the weight transfer matrix and bias transfer matrix between the neural network core of the scene to be tested and the neural network core of the lth reference scene; S46, based on the obtained migration matrix Combined with the reference optimization matrix obtained through measured optimization training under L reference scenarios Calculate the final weight and bias matrix for the test scenario; the final weight matrix w Final With the bias matrix b Final for: Among them, I l Represents the dataset size in the lth reference scenario; S47, output the final weight matrix w after migration Final And the final bias matrix b Final .

Images