Channel Scenario Classification Method, System and Storage Medium Based on Deep Learning

Abstract

The present invention provides a channel scenario classification method, system and storage medium based on deep learning. The method includes: obtaining channel data, where the channel data includes the I-channel signal, Q-channel signal, pilot signal corresponding to the channel, and the channel impulse response estimated based on the obtained pilot signal; constructing a channel scenario classification model, where the channel scenario classification model includes an input layer, a bidirectional long short-term memory network layer, a fully connected layer, a softmax layer, and an output layer; inputting the obtained channel data into the bidirectional long short-term memory network layer to obtain the feature data of the channel; inputting the feature data into the fully connected layer to obtain a target feature vector; inputting the target feature vector into the softmax layer to obtain the probability distribution of various channel scenarios; and obtaining a channel scenario classification result based on the probability distribution. This channel scenario classification method can accurately identify the scenario category of the channel, thereby improving the technical performance of subsequent adaptive modulation and coding.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and in particular, to a channel scenario classification method, system, and storage medium based on deep learning. Background Art

[0002] ITU-T (International Telecommunication Union Telecommunication Standardization Sector) clearly points out that low-earth orbit satellite communication systems will become an important part of 6G communication network systems. In low-earth orbit satellite systems, the ACM (Adaptive Coding and Modulation) technology is the core technology for realizing adaptive adjustment of link transmission modes according to channel states; the ACM technology not only improves the original channel transmission capacity but also greatly expands the service scope. In a low-earth orbit satellite communication system adopting the ACM transmission scheme, an important limitation is that the transmitting end needs to obtain real-time and accurate channel state information (CSI). When performing modulation and coding strategy (MCS) switching, the transmitting end must rely on the real-time and accurate CSI sequence periodically fed back by the receiving end.

[0003] For a complete low-earth orbit satellite multi-beam satellite link-level communication system including a gateway station - satellite - receiving terminal, multiple code rates and modulation modes are supported. First, different code rates and modulation modes are combined, sorted according to spectral efficiency, and the MCS number is obtained. Secondly, an MCS strategy is selected, and a channel model is selected as the communication scenario. For the SNR (Signal-to-Noise Ratio) range traversing [-20:20] dB, with a fixed target block error rate (BLER) value of 10 -3 as the limit standard, the switching threshold, that is, the demodulation threshold, under each modulation and coding method is obtained. Then, all MCS numbers are traversed, and the switching threshold of the MCS under this channel model is obtained according to the target bit error rate algorithm. When multiple MCSs simultaneously meet the target bit error rate, the MCS with the highest spectral efficiency is selected. Finally, an MCS mapping criterion table under this channel model is formed.

[0004] In a communication system adopting the ACM transmission scheme, after obtaining accurate and effective CSI, it is necessary to use a codebook lookup table to quantify the interval in which the estimated SNR (Signal-to-Noise Ratio) value is located to select the best MCS strategy. The selection of the MCS strategy is not only highly sensitive to SNR, and with different channel models used in simulation or real communication or different channel conditions experienced during testing, the mapping criterion between MCS and CSI will be different, and different channel types will cause changes in the MCS mapping criterion of ACM; the wrong modulation and coding strategy mapping criterion will lead to resource waste and seriously affect the throughput performance of the system. Therefore, the existing ACM technology has problems such as high resource consumption, invalidation of the MCS judgment criterion during scenario switching, and low system throughput.

[0005] The applicant's research has discovered that if the current channel scenario of the terminal's beam can be obtained before determining the MCS mapping criteria, the computational complexity of the MCS selection process can be greatly reduced, saving resources and improving system throughput. Therefore, accurately classifying the current channel scenario of the terminal's beam is a technical problem that needs to be solved urgently. Summary of the Invention

[0006] In view of this, the present invention provides a channel scene classification method, system and storage medium based on deep learning to solve one or more problems existing in the prior art.

[0007] According to one aspect of the present invention, the present invention discloses a channel scene classification method based on deep learning, the method comprising:

[0008] Acquire channel data of a beam where the terminal is located, the channel data including an I-channel signal, a Q-channel signal, a pilot signal corresponding to the channel, and a channel impulse response estimated based on the acquired pilot signal;

[0009] Constructing a channel scene classification model, wherein the channel scene classification model includes an input layer, a bidirectional long short-term memory network layer, a fully connected layer, a softmax layer, and an output layer;

[0010] Inputting the acquired channel data into the bidirectional long short-term memory network layer to obtain feature data of the channel;

[0011] Inputting the feature data into a fully connected layer to obtain a target feature vector;

[0012] The target feature vector is input into the softmax layer to obtain the probability distribution of various channel scenarios; and the channel scenario classification result is obtained based on the probability distribution.

[0013] In some embodiments of the present invention, the method comprises:

[0014] Construct loss function;

[0015] The network model is pre-trained to obtain a trained channel scene classification model.

[0016] In some embodiments of the present invention, pre-training a network model to obtain a trained channel scene classification model includes:

[0017] Acquire a sample data set, where the sample data in the sample data set includes channel data and corresponding channel scene labels;

[0018] Generate a training set and a test set from the sample data set according to a preset ratio;

[0019] Input the training set and the test set into the network model for pre-training to obtain the trained channel scene classification model.

[0020] In some embodiments of the present invention, the loss function is:

[0021]

[0022] where N is the number of channel scene categories, x represents the input channel data, and label represents the output channel scene label.

[0023] In some embodiments of the present invention, the scene categories of the channel include: rural scene, urban scene, and suburban scene.

[0024] In some embodiments of the present invention, the activation function adopted by the forget gate in the bidirectional long short-term memory network layer is the sigmoid function, and the activation functions adopted by the input gate and the output gate are both the tanh function.

[0025] In some embodiments of the present invention, the output result of the input gate is:

[0026]

[0027]

[0028] where represents the input at time t, ω il is the iteration matrix of size i×l of the input gate, I represents the number of rows of the input gate iteration matrix, represents the output of the memory cell at time t - 1, ω cl represents the iteration matrix of size c×l of the input gate core unit, C represents the number of rows of the input gate core unit iteration matrix, F g represents the input gate activation function, F f represents the forget gate activation function, represents the output of the input gate, Iy(m) represents the I-channel signal, Qy(m) represents the Q-channel signal, and x(m) represents the pilot signal, represents the channel impulse response estimated based on the pilot signal.

[0029] The output result of the forget gate is:

[0030]

[0031]

[0032] where represents the input at time t, ωiφ The iterative matrix representing the size of the forget gate as i×φ, where I represents the number of rows of the forget gate iterative matrix, represents the output of the memory cell at time t - 1, ω cφ The iterative matrix representing the size of the core unit of the forget gate as c×φ, where C represents the number of rows of the forget gate core unit iterative matrix, F f represents the forget gate activation function, represents the output of the forget gate.

[0033] The output result of the output gate is:

[0034]

[0035]

[0036] Among them, represents the input at time t, ω iω The iterative matrix representing the size of the output gate as i×ω, where I represents the number of rows of the output gate iterative matrix, represents the output of the memory cell at time t, ω cω The iterative matrix representing the size of the core unit of the output gate as c×ω, where C represents the number of rows of the output gate core unit iterative matrix, F h represents the output gate activation function, F f represents the forget gate activation function, represents the output of the output gate, and c′ is the output of the core unit in the bidirectional long short - term memory network layer,

[0037] In some embodiments of the present invention, the bidirectional long short - term memory network layer includes two hidden layers.

[0038] According to another aspect of the present invention, there is also disclosed a channel scenario classification system based on deep learning, which includes a processor and a memory. Computer instructions are stored in the memory, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method described in any one of the above embodiments.

[0039] According to still another aspect of the present invention, there is also disclosed a computer - readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps of the method described in any one of the above embodiments.

[0040] The channel scenario classification method, system, and storage medium based on deep learning according to the embodiments of the present invention can accurately classify channel scenarios by using a channel scenario classification model based on a bidirectional long short-term memory network, that is, can effectively distinguish the communication scenarios where the terminal is located, so as to quickly obtain the correct MCS mapping criterion, thereby reducing resource consumption and improving the throughput performance of the system.

[0041] Additional advantages, objects, and features of the present invention will be partly described below, and will partly become apparent to those of ordinary skill in the art after studying the following description, or may be learned from the practice of the present invention. The objects and other advantages of the present invention can be realized and obtained by the structure specifically pointed out in the written description, claims, and drawings.

[0042] Those skilled in the art will understand that the objects and advantages that can be achieved by the present invention are not limited to the above specifically described, and the above and other objects that the present invention can achieve will be more clearly understood according to the following detailed description. Brief Description of the Drawings

[0043] The drawings described herein are used to provide a further understanding of the present invention, form a part of this application, and do not limit the present invention. The components in the drawings are not drawn to scale, but are only for showing the principles of the present invention. For the convenience of showing and describing some parts of the present invention, the corresponding parts in the drawings may be enlarged, that is, may become larger relative to other components in the exemplary device actually manufactured according to the present invention. In the drawings:

[0044] Figure 1 It is a schematic flowchart of a channel scenario classification method based on deep learning according to an embodiment of the present invention.

[0045] Figure 2 It is a schematic architecture diagram of a channel scenario classification model according to an embodiment of the present invention.

[0046] Figure 3 It is a curve graph showing the change of the loss function calculation value during the update and iteration process of a channel scenario classification model according to an embodiment of the present invention.

[0047] Figure 4 It is a curve graph showing the change of the accuracy function calculation value during the update and iteration process of a channel scenario classification model according to an embodiment of the present invention.

[0048] Figure 5 It is a first comparison schematic diagram of the system throughput performance with channel scenario classification and the system throughput performance without channel scenario classification.

[0049] Figure 6It is a second comparison schematic diagram of the system throughput performance with channel scenario classification and the system throughput performance without channel scenario classification. Detailed implementation manners

[0050] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer and more understandable, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Herein, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but not to limit the present invention.

[0051] Herein, it should be noted that in order to avoid obscuring the present invention due to unnecessary details, only the structures and / or processing steps closely related to the solution of the present invention are shown in the drawings, while other details less related to the present invention are omitted.

[0052] It should be emphasized that the terms "comprising / including / having" when used herein refer to the presence of features, elements, steps or components, but do not exclude the presence or addition of one or more other features, elements, steps or components.

[0053] Different from the terrestrial communication scenario, the trajectory of the satellite communication scenario of the same satellite in the same communication situation is relatively fixed, and the terminal environment information is also relatively fixed, which is more suitable for classifying the receiving terminal scenario. At the same time, channel classification is also beneficial to the subsequent judgment criteria for switching thresholds of adaptive coding and modulation technologies. In the scenario of low-earth orbit satellite communication, the entire process from when the satellite is locked and accessed to when it leaves the communication range is only about 11.5 minutes; as described in the international standard ITU-R M.1225: For the speed of the receiving terminal, two standard values should usually be considered. The operating speed of the pedestrian's handheld access terminal environment is 3 km / h, and the operating speed of the vehicle's access terminal environment is 70 km / h; therefore, when the satellite moves throughout the communication process, the communication environment on the ground will not change drastically during the entire communication process. Thus, accurately classifying the receiving end environment scenario can effectively improve the performance of the ACM technology. Based on this, the present invention provides a channel scenario classification method, system and storage medium based on deep learning. This channel scenario classification method can effectively distinguish the communication scenario where the receiving end is located, which is beneficial to the subsequent acquisition of CSI information and the judgment criteria for switching thresholds of adaptive coding and modulation technologies, and can improve the performance of the ACM technology.

[0054] In the following, embodiments of the present invention will be described with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0055] Figure 1 It is a flowchart of a channel scenario classification method based on deep learning according to an embodiment of the present invention, as Figure 1As shown, the deep learning-based channel scenario classification method at least includes steps S10 to S50.

[0056] Step S10: Obtain the channel data of the beam where the terminal is located. The channel data includes the I-channel signal, Q-channel signal, pilot signal corresponding to the channel, and the channel impulse response estimated based on the obtained pilot signal.

[0057] In this step, the I-channel signal and Q-channel signal can be obtained by cutting the signals collected at the receiver end. The I-channel signal can be denoted as Iy(m), and the Q-channel signal can be denoted as Qy(m). The pilot signal can also be obtained from the receiver end and can be denoted as x(m). In a low-earth orbit satellite adaptive communication system, according to the channel state estimator, the estimated channel impulse response information can be obtained. Then, the channel impulse response estimated based on the obtained pilot signal can be further denoted as

[0058] Exemplarily, the signals collected at the receiver end are cut into two channels, I and Q, for processing. Iy(m) is the I-channel signal, Qy(m) is the Q-channel signal, and the estimated channel impulse response is obtained based on the channel state estimator. At this time, the channel data

[0059] Step S20: Construct a channel scenario classification model. The channel scenario classification model includes an input layer, a bidirectional long short-term memory network layer, a fully connected layer, a softmax layer, and an output layer.

[0060] In this step, the channel scenario classification model is obtained by training the network model. The channel scenario classification model adopts a bidirectional long short-term memory network structure. Refer to Figure 2 , specifically, the bidirectional long short-term memory network layer may include two hidden layers.

[0061] Step S30: Input the obtained channel data into the bidirectional long short-term memory network layer to obtain the feature data of the channel.

[0062] In this step, specifically, the channel data obtained in step S10 is input into the bidirectional long short-term memory network layer through the input layer to obtain the feature data of the channel. Exemplarily, each hidden layer in the LSTM layer contains 100 hidden nodes. And in the hidden layer, the bidirectional long short-term memory network layer realizes the automatic extraction of features and selects different filters through the input gate, output gate, and forget gate to extract different features. Specifically, where ω iω represents the iteration matrix of the output gate with a size of i×ω, represents the iteration matrix of the forget gate with a size of i×φ, ω ilAn iterative matrix representing the size of the input gate as i×l, ω ic An iterative matrix representing the size of the core unit as i×c, Iy(m) is the I-channel signal, Qy(m) is the Q-channel signal, and x(m) is the pilot signal. Is the channel impulse response estimated based on the pilot signal.

[0063] Step S40: Input the feature data into the fully connected layer to obtain the target feature vector.

[0064] The fully connected layer is responsible for converting the two-dimensional feature map output by the bidirectional long short-term memory network layer into a one-dimensional feature vector. In the present invention, the input of the fully connected layer receives the output of the bidirectional long short-term memory network layer.

[0065] Step S50: Input the target feature vector into the softmax layer to obtain the probability distribution of various channel scenarios, and obtain the channel scenario classification result based on the probability distribution.

[0066] In this step, the channel scenario classification result corresponding to the channel is obtained based on the target feature vector output by the fully connected layer. As Figure 2 shown, the channel scenario classification result can be output through the output layer of the channel scenario classification model. When the channel scenario category is set to 3 categories (rural scenario, suburban scenario, urban scenario), the channel scenario classification result output by the output layer is one of rural, suburban, and urban. It should be understood that limiting the channel scenario category to 3 categories is only an example. In other embodiments, the channel scenario category can also be set to more categories.

[0067] In one embodiment, constructing the channel scenario classification model specifically includes the following steps: constructing a loss function; pre-training the network model to obtain the trained channel scenario classification model. The loss function is used to evaluate the difference between the predicted value and the true value of the model. The smaller the loss function, the better the robustness of the model. When pre-training the model, specifically include: obtaining a sample data set, where the sample data in the sample data set includes channel data and corresponding channel scenario labels; generating a training set and a test set from the sample data set according to a preset ratio; inputting the training set and the test set into the network model for pre-training to obtain the trained channel scenario classification model.

[0068] In this embodiment, the channel data of the sample data includes the I-channel signal, Q-channel signal, pilot signal corresponding to the channel, and the channel impulse response estimated based on the obtained pilot signal, and the corresponding channel scene label is specifically one of the rural scene, suburban scene, and urban scene. Exemplarily, the number of sample data in the selected training set can be 8,000 groups, and the number of sample data in the test set can be 1,000 groups, and the length of a single sample data is 500 bit. When training the model, the number of iterations of the batch data is selected to be 100 times; the feature of a single sample data is a two-dimensional feature, and the number of target classifications is 3 categories (rural scene, suburban scene, and urban scene). The bidirectional long short-term memory network layer in the network model has two hidden layers, the deep learning network uses the Adam neural network optimizer, and the loss function in the entire deep learning network is selected as the cross-entropy loss function. The cross-entropy loss function is specifically:

[0069]

[0070] where N is the number of channel scene categories, x represents the input channel data, and label represents the output channel scene label.

[0071] In an embodiment of the present invention, the activation function adopted by the forget gate in the bidirectional long short-term memory network layer is the sigmoid function, and the activation functions adopted by the input gate and the output gate are both the tanh function. In the LSTM structure, the hidden layer at each moment contains multiple memory blocks, each memory block corresponds to a memory cell, and each memory block structure can output a corresponding vector during training. In the network model of the present invention, the structural processes of each part in the LSTM layer are as follows:

[0072] For the input gate, it is composed of the pilot signal passed in through the output layer at the current moment and the input of the moment conversion matrix parameters of each order and all the memory cells in the same memory block at the previous moment The output result of the input gate obtained from these two input vectors is:

[0073]

[0074]

[0075] where represents the input at time t, ω il is the iteration matrix of size i×l of the input gate, I represents the number of rows of the input gate iteration matrix, represents the output of the memory cell at time t - 1, ω clAn iterative matrix representing the size of the input gate core unit as c×l, where C represents the number of rows of the iterative matrix of the input gate core unit, and F g Represents the input gate activation function, F f Represents the forget gate activation function Represents the output of the input gate Iy(m) represents the I-channel signal, Qy(m) represents the Q-channel signal, and x(m) represents the pilot signal Represents the channel impulse response estimated based on the pilot signal

[0076] For the forget gate, the input consists of the pilot signal passed through the output layer at the current moment and the input of the moment conversion matrix parameters And all the memory cells within the same memory block at the previous moment The output result of the forget gate obtained from these two input vectors is:

[0077]

[0078]

[0079] Among them, Represents the input at time t, ω iφ Represents an iterative matrix with the size of the forget gate as i×φ, where I represents the number of rows of the iterative matrix of the forget gate Represents the output of the memory cell at time t - 1, ω cφ Represents an iterative matrix with the size of the forget gate core unit as c×φ, where C represents the number of rows of the iterative matrix of the forget gate core unit, and F f Represents the forget gate activation function Represents the output of the forget gate, Iy(m) represents the I-channel signal, Qy(m) represents the Q-channel signal, and x(m) represents the pilot signal Represents the channel impulse response estimated based on the pilot signal

[0080] For the core unit, it is the product of the input gate and the input data (F f (Z i )*F g (Z)) and the product of the forget gate and the output of the corresponding core unit at the previous moment (c*F f (Z f ). The final output result of the core unit is:

[0081]

[0082]

[0083] Among them, Represents the input at time t, ω ic Represents the iterative matrix of size i×c of the core unit, I represents the number of rows of the iterative matrix of the core unit, F f Represents the forget gate activation function, c represents the number of columns of the iterative matrix of the core unit, F g Represents the input gate activation function, Represents the output of the core unit, Iy(m) represents the I-channel signal, Qy(m) represents the Q-channel signal, x(m) represents the pilot signal, Represents the channel impulse response estimated based on the pilot signal.

[0084] For the output gate, the pilot signal passed through the output layer at the current time and the input of the moment conversion matrix parameters And all the core units within the same memory block at the current time As the input, the output result of the most output gate is:

[0085]

[0086]

[0087] Among them, Represents the input at time t, ω iω Represents the iterative matrix of size i×ω of the output gate, I represents the number of rows of the iterative matrix of the output gate, Represents the output of the memory unit at time t, ω cω Represents the iterative matrix of size c×ω of the core unit of the output gate, C represents the number of rows of the iterative matrix of the core unit of the output gate, F h Represents the output gate activation function, F f Represents the forget gate activation function, Represents the output of the output gate, c′ is the output of the core unit, Iy(m) represents the I-channel signal, Qy(m) represents the Q-channel signal, x(m) represents the pilot signal, Represents the channel impulse response estimated based on the pilot signal.

[0088] For the final overall memory unit, through the output of the output gate at the current time And the output of the core unit at the current time The output result of the final memory unit is obtained. Among them, F f (Gate activation function) selects the sigmoid function; F g (Input activation function) and F h (Output activation function) both select the tanh function, and the specific representation is as follows:

[0089]

[0090]

[0091] In the above formula, x represents the input of the corresponding gate.

[0092] Furthermore, after combining multiple LSTM units into an LSTM layer, the output result of the LSTM layer can be obtained: ω iω represents an iterative matrix with the size of the output gate being i×ω; represents an iterative matrix with the size of the forget gate being i×φ; ω il represents an iterative matrix with the size of the input gate being i×l; ω ic represents; c′ represents the output of the core unit; out represents the output of the bidirectional long short-term memory network layer, c represents the number of rows of the iterative matrix of the input gate core unit.

[0093] The output of the bidirectional long short-term memory network layer is further input to the fully connected layer, and the fully connected layer realizes feature fusion, that is, the fully connected layer aggregates all feature information and further sends the output value to the next layer; the specific representation form of the fully connected layer is: [ω full ; outf] = FullConnectedLayer[c′; out]; where, ω full represents the iterative matrix of the fully connected layer; outf represents the output of the fully connected layer; c′ represents the output of the core unit; out represents the output of the bidirectional long short-term memory network layer. The output of the fully connected layer is specifically input to the Softmax layer, and the Softmax layer will obtain the probability value of each data and send the output value to the classifier to label the data used for feature classification; its specific representation form is: [ω softmax ; outs] = SoftmaxLayer[I*outf + b]; where, ω softmax represents the iterative matrix of the softmax layer, outs represents the output of the softmax layer, I represents the number of rows of the iterative matrix of the fully connected layer, outf represents the output of the fully connected layer, and b represents the model parameter. Furthermore, the output layer generates label information under different satellite channel states; that is, [Feature net = OutLayer[outs].

[0094] Since a bidirectional LSTM network structure is adopted for the entire deep learning network, forward propagation and backward propagation can be performed simultaneously for output calculation and iterative update of weight gradients. For the forward propagation process of LSTM, the input gate, forget gate, cell gate, and output gate can be analyzed one by one starting from the input data, and finally the output value can be obtained; for the backward propagation process of LSTM, the gradient of each weight in the entire LSTM structure should be solved through the chain derivative. For the final output of the network, the Softmax function is used to calculate the probability that the result belongs to a certain category in the total number of overall target categories. At this time, the sum of the probabilities that the result belongs to the total number of overall target categories is 1. Then, the matrix results of each weight and bias are updated using the gradient, and multiple iterations are performed to finally complete the network training.

[0095] Exemplarily, Figure 3 is a curve graph showing the change of the calculated value of the loss function during the update iteration process of the channel scenario classification model according to an embodiment of the present invention; in this embodiment, 100 update iterations are performed, that is, the entire data set has iterated the data 100 times, and the network is successfully trained without overfitting or underfitting. It can be seen from the figure that the calculated value of the loss function of the deep learning network has decreased from the initial starting point value greater than 1.1 to 0.0002. And Figure 4 is a curve graph showing the change of the calculated value of the accuracy function during the update iteration process of the channel scenario classification model according to an embodiment of the present invention. From Figure 4 it can be seen that the calculated value of the accuracy function of the deep learning network has increased from 0.33 to 0.99. From Figure 3 and Figure 4 it can be known that the network model in this embodiment has been well trained; thus, the channel scenario can be accurately classified based on the trained channel scenario classification model.

[0096] Furthermore, the trained model can be tested based on the data in the test set. In this embodiment, 100 groups of data are selected from the test set to test the channel scenario classification model, and the 100 groups of data include random data obtained from three different scenario channels. The test results show that the final classification accuracy of this channel scenario classification model remains between 0.99 and 1. And as the communication scenarios in the city, suburb, and rural areas change, the influence caused by small-scale fading gradually decreases.

[0097] In this embodiment, the channel scenario classification model obtained based on training can successfully distinguish three channel types, namely, ITU-LOS-A (rural), ITU-LOS-B (suburban), and ITU-LOS-C (urban) environments. The comparison results of the test accuracies obtained under different data sample sizes are shown in the following table. In the online communication phase, before inputting the data sample into the neural network for dimension conversion, 200 frames of data can meet the required sample size, which means that the input data exactly corresponds to 2s.

[0098] Accuracy Comparison of Different Data Sample Sizes

[0099] Number of frames 1 50 100 150 200 Accuracy rate (%) 33.5294 55.8824 67.0588 98.2353 100

[0100] For this application, the throughput analysis and comparison of different strategies (traditional ACM strategy, fixed MCS strategy, and MCS strategy after channel scenario classification using the channel scenario classification method of this application) were further simulated during the entire process from satellite communication access to departure. In the link simulation, the considered channel models are urban channel scenarios and rural channel scenarios, namely, the ITU-LOS-A (rural) and ITU-LOS-C (urban) channel environments mentioned above. In the link, the satellite orbit altitude is set to 1175 km, the signal waveform is selected as DFT-S-OFDM, the subcarrier spacing is 120 kHz, the number of NRBs is 264, the bandwidth is 400 MHz, the coding method is selected as LDPC coding, and the modulation methods are selected as QPSK and 8PSK. The calculation formula for throughput is as follows: Where, W is the bandwidth of the signal, T is the total time length of communication transmission, Eff(i) is the spectral efficiency of the MCS strategy selected at the current moment, and BER(i) is the calculated bit error rate value. When the environment of the receiving terminal changes, it can be clearly observed that the strategy after channel scenario classification based on the channel scenario classification method of this application has an overall throughput increase compared to other solutions.

[0101] Exemplarily, Figure 5 Fig. is the first comparison schematic diagram of the system throughput performance after channel scenario classification and the system throughput performance without channel scenario classification. This figure shows the change in the overall system throughput corresponding to different transmission strategies when the communication scenario changes from urban to rural, that is, when the channel state improves. From Figure 5As can be seen, when compared with the traditional ACM strategy, within 0 - 593.924 s, due to more accurate estimation of the channel state, the ACM strategy solution adopting the channel scenario classification technology shows better performance, with the throughput increased by 3352 Mbit; when the communication scenario switches after 593.924 s, the overall throughput is increased by 85656 Mbit. However, the traditional ACM solution selects a lower MCS solution with outdated CSI, resulting in waste of resources and thus a decrease in the overall system throughput. Compared with the fixed MCS and the traditional ACM strategy, the system throughput performance of the channel scenario classification of the present invention has been greatly improved compared with that of the system without channel scenario classification, and the overall throughput is increased by 331848 Mbit.

[0102] Figure 6 It is the second comparison schematic diagram of the system throughput performance of the system adopting channel scenario classification and the system without channel scenario classification. This figure shows the environmental change of the communication scenario from rural to urban, that is, when the channel state deteriorates, the change of the overall throughput of different transmission strategy systems. As can be seen from Figure 6 As can be seen, when compared with the traditional ACM strategy, within 0 - 260.779 s, due to more accurate estimation of the channel state, the ACM strategy solution adopting the channel scenario classification technology shows better performance, with the throughput increased by 2466 Mbit; when the communication scenario switches after 260.779 s, the overall throughput is increased by 128683 Mbit. It can be seen from the figure that the traditional ACM solution selects a higher MCS solution with outdated CSI, resulting in an increase in the bit error rate and thus a decrease in the overall system throughput; the system throughput performance of the channel scenario classification of the present invention has been greatly improved compared with that of the system without channel scenario classification, and the overall throughput is increased by 335313 Mbit.

[0103] Since it is necessary to train, verify and simulate the channel scenario classification model based on the bidirectional long short - term memory network, before training the network model based on the bidirectional long short - term memory network, the model parameters are set first. The specific model parameters are shown in the following table:

[0104] Parameter name Parameter value Deep learning network BI-LSTM channel classifier Number of hidden layer nodes 100*2 Initial learning rate 0.005 Loss function Cross-entropy function Optimizer Adam optimizer Target channel type ITU-LOS-A / B / C

[0105] After setting the model parameters, configure the deep learning network model according to the parameter configuration in the above table and perform model training. To obtain the dataset samples, the channel estimation results and transceiver data during the transmission process can be collected and stored as the training dataset samples. Calculate the round-trip channel state information delay through the track position information, perform delay compensation processing on the stored data, and according to the estimation results and delay compensation operations, map the channel quality and communication time one by one; cut each data sample of the above dataset according to the time slot size and perform dimensional conversion; correspond the data in different channel scenarios to their respective channel types, and use the corresponding channel scenario type as the label of the network classification output result to obtain the final dataset samples. It should also be understood that the number of samples in the training set and test set set in the embodiments of the present application is only an example, and can be specifically set according to the actual application scenario. In addition, the scalability of the channel scenario analogy in the above-mentioned channel scenario classification method based on deep learning is relatively strong. In some embodiments, only three channels, namely ITU-LOS-A / B / C, are selected for the channel types of the low-earth orbit satellite communication scenario; in subsequent engineering application research, other channel types and other influencing factors such as weather conditions can be introduced for training to enrich the satellite channel feature database, so it has high engineering application value.

[0106] Correspondingly, the present invention also provides a channel scenario classification system based on deep learning. The system includes a processor and a memory. Computer instructions are stored in the memory, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method described in any of the above embodiments.

[0107] In addition, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps of the method described in any of the above embodiments.

[0108] Through the above embodiments, it can be found that for the channel scenario classification method and system based on deep learning, the channel scenario classification model based on the bidirectional long short-term memory network can achieve accurate classification of channel scenarios, that is, it can effectively distinguish the communication scenarios where the receiving terminal is located, so as to quickly select the correct MCS mapping criterion, thereby reducing resource consumption and improving the throughput performance of the system.

[0109] Those of ordinary skill in the art should understand that the various exemplary components, systems, and methods described in connection with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Specifically, whether to implement in hardware or software depends on the specific application and design constraints of the technical solution. A professional technician can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a functional card, and so on. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium or transmitted via a data signal carried in a carrier wave over a transmission medium or a communication link. "Machine-readable medium" can include any medium that can store or transmit information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical discs, hard disks, fiber optic media, radio frequency (RF) links, and so on. The code segment can be downloaded via a computer network such as the Internet, an intranet, and so on.

[0110] It should also be noted that the exemplary embodiments mentioned in the present invention describe some methods or systems based on a series of steps or devices. However, the present invention is not limited to the order of the above steps, that is, the steps can be executed in the order mentioned in the embodiments, can be different from the order in the embodiments, or several steps can be executed simultaneously.

[0111] In the present invention, the features described and / or illustrated for one embodiment can be used in the same or a similar manner in one or more other embodiments, and / or combined with the features of other embodiments or replace the features of other embodiments.

[0112] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, various changes and modifications can be made to the embodiments of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A channel scenario classification method based on deep learning, characterized in that, The method includes: Obtaining channel data of the beam where the terminal is located, where the channel data includes the I-channel signal, Q-channel signal, pilot signal corresponding to the channel, and the channel impulse response estimated based on the obtained pilot signal; Constructing a channel scenario classification model, where the channel scenario classification model includes an input layer, a bidirectional long short-term memory network layer, a fully connected layer, a softmax layer, and an output layer; Inputting the obtained channel data into the bidirectional long short-term memory network layer to obtain feature data of the channel; Inputting the feature data into the fully connected layer to obtain a target feature vector; Inputting the target feature vector into the softmax layer to obtain the probability distribution of various channel scenarios, and obtaining a channel scenario classification result based on the probability distribution; The activation function adopted by the forgetting gate in the bidirectional long short-term memory network layer is the sigmoid function, and the activation functions adopted by the input gate and the output gate are both the tanh function; The output result of the input gate is: Among them, represents the input at time t, ω il is the iteration matrix of size i×l for the input gate, I represents the number of rows of the input gate iteration matrix, represents the output of the memory cell at time t-1, ω cl represents the iteration matrix of size c×l for the input gate core unit, C represents the number of rows of the input gate core unit iteration matrix, F g represents the input gate activation function, F f represents the forget gate activation function, represents the output of the input gate, Iy(m) represents the I-channel signal, Qy(m) represents the Q-channel signal, and x(m) represents the pilot signal, represents the channel impulse response estimated based on the pilot signal; The output result of the forgetting gate is: Among them, represents the input at time t, ω iφ represents the iteration matrix of the forget gate with a size of i×φ, and I represents the number of rows of the forget gate iteration matrix, represents the output of the memory cell at time t-1, ω cφ represents the iteration matrix of the forget gate core unit with a size of c×φ, and C represents the number of rows of the forget gate core unit iteration matrix, F f represents the activation function of the forget gate, represents the output of the forget gate; The output result of the output gate is: Among them, represents the input at time t, ω iω represents the iteration matrix of the output gate with a size of i×ω, I represents the number of rows of the output gate iteration matrix, represents the output of the memory cell at time t, ω cω represents the iteration matrix of the output gate core unit with a size of c×ω, C represents the number of rows of the output gate core unit iteration matrix, F h represents the output gate activation function, F f represents the forget gate activation function, represents the output of the output gate, c′ is the output of the core unit in the bidirectional long short-term memory network layer, 2. The channel scenario classification method based on deep learning according to claim 1, wherein Constructing a channel scenario classification model includes: Constructing a loss function; Pre-training the network model to obtain a trained channel scenario classification model.

3. The channel scenario classification method based on deep learning according to claim 2, characterized in that, Pre-training the network model to obtain a trained channel scenario classification model includes: Obtaining a sample data set, where the sample data in the sample data set includes channel data and corresponding channel scenario labels; Generating a training set and a test set from the sample data set according to a preset ratio; Inputting the training set and the test set into the network model for pre-training to obtain a trained channel scenario classification model.

4. The channel scenario classification method based on deep learning according to claim 2, characterized in that The loss function is: Where N is the number of channel scenario categories, x represents the input channel data, and label represents the output channel scenario label.

5. The deep learning-based channel scenario classification method according to any one of claims 1 to 4, characterized in that The scenario categories of the channel include: rural scenario, urban scenario, and suburban scenario.

6. The method for classifying channel scenarios based on deep learning according to claim 1, wherein The bidirectional long short-term memory network layer includes two hidden layers.

7. A channel scenario classification system based on deep learning, the system comprising a processor and a memory, characterized in that, The memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 6.

Images