CN110708129B - A method for acquiring wireless channel state information - Google Patents

Abstract

A wireless channel state information acquisition method belongs to the technical field of wireless and mobile communication, and is characterized by comprising the following steps: carrying out uniform format processing on a large number of channel information data samples obtained by channel estimation to form training samples; inputting the sample data into a prediction neural network to obtain a group of optimal parameters aiming at the data set; judging to obtain the actual state of the current prediction precision by the aid of the prediction effect, and training further optimization parameters if the prediction precision does not meet the requirement; the channel state information obtained by predicting the network will be applied to the originating terminal. The second characteristic is that the proposed Adaptive Structure Extreme Learning Machine (ASELM) algorithm can adaptively adjust the Structure of the network to match the changing characteristics of the channel data.

Description

A method for acquiring wireless channel state information

technical field

The invention relates to a method for acquiring wireless channel state information, which belongs to the technical field of wireless and mobile communication.

Background technique

Taking the drone communication scene as an example, in recent years, drones have been widely used in many fields of social production and life. In some joint detection missions, in order to make the UAV group cooperate more effectively, some data needs to be shared between UAVs. Therefore, there are more communication requirements between UAVs. Channel state information is a key parameter for communication between UAVs. However, the channel state changes rapidly in the mobile state. Due to the feedback delay in the channel estimation process, the obtained channel state information has the problem of being out of date, which will lead to communication interruption. Therefore, high dynamics pose great challenges to channel estimation, while channel prediction offers potential possibilities to solve such problems. In order to solve the above problems, researchers have invested a lot of energy in channel prediction. In traditional transmission schemes, channel prediction based on historical channel state information obtained from channel estimation has the effect of improving system performance.

Researchers have designed some channel prediction algorithms. In the classic autoregressive (AR) algorithm, the channel impulse response is represented by a linear combination of historical channel state information. The effect of Le frequency offset on channel prediction is obviously decreased. In addition, the echo state network (Echo State Networks, ESN) is also used to predict the state change of the Rayleigh channel. However, from the perspective of simulation results, the performance of ESN algorithm in UAV communication scenarios is still difficult to guarantee the reliability of communication. Different from the AR and ESN algorithms, the ASELM algorithm can adaptively adjust its own prediction network structure during the prediction process to deal with the rapid change of the channel state in the UAV communication process.

Contents of the invention

In view of this, the present invention considers the adaptive adjustment of the structure of the prediction neural network, and proposes the ASELM channel prediction algorithm based on the ELM algorithm. It is characterized in that the algorithm contains the following process, and the improved ELM algorithm for channel information acquisition mainly includes the following two steps:

Training process: set the demand forecasting accuracy and the length of the forecasting window. The algorithm achieves good forecasting results by setting a certain length of forecasting window and the sliding step of the forecasting window. After training, the algorithm can automatically Adaptively adjust its internal neural network structure, so as to realize the joint optimization of prediction accuracy and prediction time. Since the number of hidden neurons that meets the demand is close to the prediction error within a certain interval; therefore, it can be considered that within a certain interval of the number of hidden neurons, The prediction result at any hidden neuron number value can be approximately considered as the corresponding prediction effect of all hidden neuron number values in the whole group; therefore, the method of group search can be adopted (here, the number of adjacent hidden neurons in a certain interval is set to be A group, the number of hidden neurons in each group is randomly selected as a typical value) to quickly search for the optimal number of hidden neurons, and determine by comparing the NMSE values of the test sets corresponding to the typical values of the number of hidden neurons. Whether the prediction effect of the typical value reaches the demand prediction accuracy; if it reaches the demand prediction accuracy, stop searching for new parameters immediately, that is, find the smallest typical value that meets the accuracy requirements. Experiments have proved that the training process takes less than 1 second.

Prediction process: Set the number of parameters of the algorithm to the typical values obtained in the above training process, and output the channel prediction results of the communication scene between UAV platforms; it has been verified that the prediction time of each node is generally less than 0.1 second.

According to the method of the present invention, the channel state information data is acquired as the training set of the proposed algorithm by performing channel estimation on the channel insertion pilot frequency, and the relevant features of the channel state data are extracted. In the actual communication process, combined with the way of sending pilots, channel prediction and channel estimation can be effectively combined, so as to ensure that the sender can grasp the channel state during the data transmission process. At the same time, due to the full use of the historical data of the channel state information, problems caused by the expiration of the channel state can be avoided, and the difficulty of rapid and changeable channel states caused by changes in the communication environment can be effectively overcome.

Description of drawings

Figure 1 is the algorithm flow chart of the proposed wireless channel state information acquisition method

Fig. 2 is a scene model diagram of UAV communication, which shows the situation that the communication channel between UAV platforms changes continuously with the movement of UAVs.

Figure 3 is a basic algorithm block diagram of the ELM algorithm.

Fig. 4 is a frame structure diagram of channel prediction. Shows the relationship between channel prediction and channel estimation during data transmission. Channel estimation is performed by means of pilot insertion to obtain historical channel state information. These data are used to train the prediction neural network. After training, the channel prediction network provides the predicted value of the channel state information during data transmission, thereby ensuring the The understanding of the channel state when sending data can effectively reduce the problem of packet loss and error and even communication interruption caused by the change of channel state during data transmission.

Figure 5 is a MATLAB simulation comparison chart of the prediction accuracy of the ASELM algorithm and the AR and ESN algorithms.

Fig. 6 is a flow chart of the complete working process of the proposed method for obtaining wireless channel state information during the communication process.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings.

Theoretically, the propagation path and time of radio waves determine the channel state information. If both the transmitting end and the receiving end are in a moving state, a considerable Doppler frequency deviation will be generated, and the channel state changes quickly. The amount of data generated by the change is large, and the prediction of these data is essentially a mathematical regression problem. When dealing with regression problems, ELM has the characteristics of strong generalization ability and fast learning. The prediction network learns these channel state data The characteristics of the data can be extracted, and then the trend of data changes can be quickly and effectively predicted. Based on these prior knowledge, we optimized the ELM algorithm and proposed a wireless channel state information acquisition method.

The following example illustrates the application of the present invention in the channel prediction of UAV platform communication. The complex sinusoidal channel model is used to equivalent the broadband wireless channel model. The UAV at the transceiver end is in 3D relative motion, and its channel state information can be expressed as

Among them, t is time, α is the complex channel gain, f _d is the Doppler frequency offset, the change of Doppler frequency offset depends on the change of the transmitting end and the receiving end, a _R is the normalized receiving matrix navigation vector, a _T is the normalized send matrix navigation vector. (θ _R ,φ _R ) is the arrival angle, (θ _T ,φ _T ) represents the launch angle, (θ,φ) represents the beam pointing, and the navigation vector a(θ,φ) can be expressed as

表示的是克罗内克积。N_y和N_x分别表示的是2×2的MIMO平面阵列上的y方向和x方向上的天线阵元的数量。d_x＝d_y＝λ/2，分别表示y方向和x方向上天线阵元的间隔，此处间隔设置为半波长。G_ant(·)表示的是辐射方向图，Where Ω _y =kd _y sin(θ)sin(φ), Ω _x =kd _x sin(θ)cos(φ), wave number k=2π/λ,

represents the Kronecker product. N _y and N _x respectively represent the number of antenna elements in the y direction and the x direction on the 2×2 MIMO planar array. d _x =d _y =λ/2, which represent the intervals of the antenna elements in the y direction and the x direction respectively, where the interval is set to half a wavelength. G _ant (·) represents the radiation pattern,

where θ and φ can be expressed as

By inserting pilots, we can obtain the above channel state information h(t). By organizing the data format of channel state information, the training set and test set required for algorithm training and testing are obtained.

The ASELM algorithm is an improved algorithm based on the ELM algorithm. The ELM algorithm is a single-hidden layer feedforward neural network (SLFNs). The algorithm has a clear structure and a clear hierarchy. The basic ELM algorithm process is as follows :

激励函数g(x)，隐藏神经元个数N,Suppose given a training set

The activation function g(x), the number of hidden neurons N,

Step 1: Arbitrarily input weight w _i and bias b _i ,

Step 2: Calculate the hidden layer output matrix H.

Step 3: Calculate the output weight β:

Among them, the hidden layer neural network output matrix is

The output weight is

The output matrix is

Based on the above ELM algorithm, ASELM algorithm is proposed. It is characterized in that the algorithm contains the following process, and the improved ELM algorithm for channel information acquisition mainly includes the following two steps:

Training process: set the demand forecasting accuracy and the length of the forecasting window. The algorithm achieves good forecasting results by setting a certain length of forecasting window and the sliding step of the forecasting window. After training, the algorithm can automatically Adaptively adjust its internal neural network structure, so as to realize the joint optimization of prediction accuracy and prediction time, because the number of hidden neurons that meets the demand is similar to the prediction error within a certain interval. Therefore, it can be considered that within a certain hidden neuron number interval, the prediction result at any hidden neuron number value can be approximately considered as the corresponding prediction effect of all hidden neuron number values in the entire group. Therefore, the method of group search can be adopted (here, the number of adjacent hidden neurons in a certain interval is set as a group, and the number of hidden neurons in each group is randomly selected as a typical value) to quickly determine the number of optimal hidden neurons. Search, by comparing the NMSE values of the test sets corresponding to different typical values of the number of hidden neurons, to determine whether the prediction effect of the typical value has reached the demand prediction accuracy. If the demand prediction accuracy is reached, the search for new parameters is immediately stopped, that is, the minimum typical value that meets the accuracy requirements is found. Experiments have proved that the training process takes less than 1 second.

Prediction process: Set the number of parameters of the algorithm to the typical values obtained in the above training process, and output the channel prediction results of the communication scene between UAV platforms. It has been verified that the prediction time of each node is generally less than 0.1 second.

(1) The specific implementation process of the ASELM algorithm is shown in Figure 1.

(2) Referring to Figure 2, in the UAV communication model, there are two flying UAVs at the receiving and receiving end. This model is equivalent to the end-to-end channel model; the dotted line with arrows indicates the direction of UAV movement and traces, the stick connections represent the channels established by the UAV's beamalignment.

The goal of our design is to ensure efficient prediction of the channel state during communication between UAV platforms. Since the UAV is doing high-speed 3D movement, the channel state changes rapidly, and it is difficult to effectively predict the channel state information. This not only requires the algorithm to have good generalization ability, but also requires the algorithm to have the ability to learn Fast features, overly complex prediction algorithms will not be able to adapt to the needs of UAV communication scenarios.

(3) Referring to Fig. 3, in this method, it can be seen from the ELM algorithm block diagram that ELM is a lightweight single-layer neural network. The improved ASELM algorithm based on the ELM algorithm is also a lightweight algorithm, which can avoid consuming a lot of computing and training time, save computing resources and computing time, and adapt to the fast and changeable channels between UAVs. The state lays the groundwork.

(4) See Figure 4. In the UAV communication system, after the communication link is established, the communication system will first perform channel estimation by inserting pilots, and then feedback the status of the channel to the originator, but the channel is rapidly changing The characteristics of the feedback channel state have expired, so it cannot be directly applied to the current data transmission. At this time, it is necessary to predict the current channel state based on the estimated historical channel state information. The channel prediction frame structure is shown in Figure 4, where the channel estimation value is the training data set, and the predicted channel state information is used as the basis for the current sending end to send signals.

(5) In order to demonstrate the practical performance of various mechanisms in the present invention, the applicant has conducted multiple simulation tests. The results of the MATLAB simulation test are shown in Figure 5.

(6) Referring to FIG. 6, in the UAV communication system, FIG. 6 shows a complete acquisition process of wireless channel state information.

The above descriptions are only preferred examples of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (5)

1. A wireless channel state information acquisition method adopts an adaptive structure overrun learning machine ASELM algorithm; the method is characterized in that the number of hidden neurons in the ELM algorithm of the ultralimit learning machine is optimized, so that the network structure of the ELM can effectively adapt to the change characteristics of channel state data, and a better prediction effect is obtained;

the improved ELM algorithm training and prediction process is as follows:

training process: setting demand prediction precision, setting the length of a prediction window, obtaining a good prediction effect by an algorithm through setting the prediction window with a certain length and the sliding step length of the prediction window, and adaptively adjusting the internal neural network structure of the algorithm according to the predicted precision demand after training so as to realize combined optimization of the prediction precision and the prediction time, wherein the prediction errors are similar in a certain interval due to the number of hidden neurons meeting the demand; in a certain hidden neuron number interval, the prediction result at any hidden neuron number value is approximately regarded as the corresponding prediction effect of all hidden neuron number values in the whole group, the fast search of the optimal hidden neuron number is carried out in a grouping search mode, and whether the prediction effect of the typical value reaches the required prediction precision or not is judged by comparing the normalized mean square error NMSE values of test sets corresponding to different hidden neuron number typical values; if the demand prediction precision is achieved, immediately stopping searching new parameters, namely finding the minimum typical value meeting the precision demand;

and (3) prediction process: and setting the parameters of the algorithm to be typical values obtained in the training process, and outputting a channel prediction result.

2. The method of claim 1, wherein parameters are rapidly filtered and optimized by means of group search, so as to obtain a channel prediction result meeting a required prediction accuracy.

3. The method as claimed in claim 1, wherein the method for obtaining status information of radio channel performs calibration of prediction accuracy by means of retraining.

4. The method of claim 1, wherein the method is used for any one of forward channel prediction, backward channel prediction, precoding, pre-equalization, antenna selection, originating scheduling, and adaptive transmission.

5. The method of claim 2, wherein the method is used for any one of forward channel prediction, backward channel prediction, precoding, pre-equalization, antenna selection, originating scheduling, and adaptive transmission.

Images