Cross-medium visible light communication system and method based on deep learning

Abstract

The invention provides a cross-medium visible light communication system and method based on deep learning, on one hand, an angle estimator based on a convolutional neural network and a receiving end movement strategy are provided, a receiving end can make a movement decision according to an angle estimation result, and link alignment of a transceiver of the cross-medium visible light communication system is adaptively realized; on the other hand, the invention provides a signal detection algorithm based on wind speed estimation and channel classification. After the system completes link alignment, the algorithm can realize estimation of wind speed and adaptively detect received signals transmitted through channels under different wind speed conditions; in the algorithm, the invention further provides a loss function based on a corrected mean square error criterion, the adverse effect of wave-induced fading of a communication channel of the cross-medium visible light system on signal transmission can be effectively relieved, and therefore the bit error rate of the system is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and cross-media visible light communication, and more specifically, to a cross-media visible light communication system and method based on deep learning. Background Art

[0002] In recent years, with the increasing demand for marine exploration and rescue, cross-medium communication has received extensive research attention. Since the visible light communication (VLC) system can transmit signals across the air-water interface and has many advantages such as high data rate and high confidentiality, it has become one of the important technologies for cross-medium communication. However, the current cross-medium VLC system has certain limitations and still faces technical challenges such as light absorption, scattering, and dynamic wavefront refraction, which need further in-depth research and solutions.

[0003] In order to improve the signal transmission performance of the existing cross-media VLC system, scholars mainly optimize the following two aspects: On the one hand, from the perspective of link alignment (LA) optimization design, wide beams can be used to expand the coverage of LEDs, or non-line of sight (NLOS) communication can be used to improve link reliability, but the disadvantage of this optimization solution is that the signal transmission distance is relatively limited. In addition, multiple-input multiple-output (MIMO) technology can be used to relax the alignment requirements, or beam tracking technology can be used to maintain alignment in dynamic environments, but the cost is to increase equipment cost and complexity.

[0004] On the other hand, it is considered to use advanced transmission technology to improve the reliability and throughput of cross-media VLC. For example, the transmission rate can be improved by combining Orthogonal Frequency Division Multiplexing (OFDM) technology, or the bit error rate performance can be improved by using demodulation technology based on Adaptive Threshold (ATD), and the system robustness can be enhanced by using Adaptive Differential Pulse Position Modulation (ADPPM) technology. However, for this aspect of optimization, the existing technology has not fully considered how to adapt to the dynamic wave surface, or it is difficult to take into account the requirements of high transmission rate. Summary of the invention

[0005] In order to overcome the defects of the above-mentioned prior art in the cross-medium VLC system, such as limited signal transmission distance, high equipment cost and complexity, difficulty in adapting to dynamic wave surfaces, and difficulty in taking into account both high communication rate and bit error rate, the present invention provides a cross-medium visible light communication system and method based on deep learning. Based on deep learning (DL) technology, link alignment of the cross-medium VLC system transceiver can be adaptively achieved; after the system is aligned, high-precision adaptive signal detection under different conditions can be achieved, and at the same time, the adverse effects of wave-induced fading of the cross-medium VLC channel on signal transmission can be effectively alleviated, thereby reducing the bit error rate (BER) of the system.

[0006] In order to solve the above technical problems, the technical solution of the present invention is as follows: A cross-medium visible light communication system based on deep learning, comprising: a transmitting end, a transmission channel and a receiving end; the transmitting end and the receiving end are respectively placed in different signal transmission media; The receiving end includes: a signal receiving module, a data preprocessing module, a link alignment module and a signal detection module; The transmitting end is used to generate and transmit an optical signal, and the optical signal reaches the signal receiving module of the receiving end after being transmitted through the transmission channel; The data preprocessing module at the receiving end is used to preprocess the received optical signal to obtain a pilot signal and a data signal; The link alignment module of the receiving end is provided with an angle estimator based on a deep learning algorithm, and the link alignment module is used to calculate the equivalent angle between the receiving end and the transmitting end according to the pilot signal, and adjust the position of the receiving end according to the equivalent angle until the link alignment is achieved; The signal detection module at the receiving end is provided with a wind speed estimator based on a deep learning algorithm, a channel classifier and several signal detectors. The signal detection module is used to estimate the wind speed according to the pilot signal after the link is aligned and calculate the weight value corresponding to each signal detector, detect the data signal in real time, and realize cross-medium visible light communication.

[0007] Preferably, the transmitting end generates an optical signal through optical OFDM (Optical OFDM, OOFDM) modulation, and the optical OFDM modulation includes: any one of: asymmetrically clipped optical-orthogonal frequency division multiplexing (Asymmetrically Clipped Optical Orthogonal Frequency Division Multiplexing, ACO-OFDM) and direct current biased optical-orthogonal frequency division multiplexing (Direct Current Biased Optical Orthogonal Frequency Division Multiplexing, DCO-OFDM).

[0008] Preferably, the equivalent angle is the angle between the line connecting the transmitting end and the receiving end and the vertical direction; The vertical distance between the receiving end and the transmitting end remains constant; the signal receiving plane of the receiving end includes a plurality of concentric circles with different radii, each of which represents an angle layer, and the equivalent angle of the receiving end and the transmitting end at any position in each angle layer is the same.

[0009] Preferably, the transmitting end is arranged on an autonomous underwater vehicle (AUV), and the autonomous underwater vehicle (AUV) is arranged underwater; The receiving end is arranged on an unmanned aerial vehicle (UAV), and the unmanned aerial vehicle (UAV) is arranged in an airspace at a specific height above the water surface.

[0010] The present invention also provides a cross-medium visible light communication method based on deep learning, based on the above-mentioned cross-medium visible light communication system based on deep learning, comprising the following steps: S1: The transmitter will transmit a bit vector After mapping and optical modulation, the time domain transmit signal vector is generated and sending it to the receiving end through the transmission channel; S2: The signal receiving module receives the time domain transmission signal vector And perform photoelectric conversion to obtain the time domain received signal ; S3: The data preprocessing module processes the received signal in the time domain Perform preprocessing to obtain the received signal , the received signal Including pilot signal and data signals ; S4: In the link alignment module, the angle estimator is based on the pilot signal Calculate the equivalent angle between the receiver and the transmitter , and according to the equivalent angle Adjust the position of the receiving end until link alignment is achieved, and execute step S5; S5: In the signal detection module, the wind speed estimator detects the wind speed according to the pilot signal. Estimated wind speed And input channel classifier, the channel classifier is based on wind speed Calculate the weight value corresponding to each signal detector; The signal detector detects the data signal in real time , and calculate the soft bit vector according to the corresponding weight value , after recovering the data bits through hard decision, cross-medium visible light communication is realized.

[0011] Preferably, in step S4, according to the equivalent angle Adjusting the position of the receiving end until link alignment is achieved includes: Real-time according to the equivalent angle Determine the angle layer of the receiver , according to the angle layer Determine the next moving step of the receiving end; Real-time based on the current equivalent angle Determine the next moving direction of the receiving end, move the receiving end along the moving direction by one step each time, and continuously move the position of the receiving end until the current equivalent angle Less than or equal to the preset alignment angle threshold , fix the receiving end position and achieve link alignment.

[0012] Preferably, the angle layer The larger the radius, the larger the next moving step of the receiving end; The real-time equivalent angle Determining the next moving direction of the receiving end, moving the receiving end along the moving direction by a step length each time, and continuously moving the position of the receiving end includes: S41: the receiving end moves a step length along a preset direction; S42: Determine the current equivalent angle Is it less than or equal to the equivalent angle before moving? If yes, the receiving end keeps the preset direction unchanged, continues to move one step along the preset direction, and re-executes step S42; otherwise, executes step S43; S43: the receiving end returns to the last position and moves a step length in the vertical direction on either side of the preset direction; After moving in the vertical direction on either side of the preset direction, determine the current equivalent angle Is it less than or equal to the equivalent angle before moving? If so, the vertical direction is used as the new preset direction and step S41 is re-executed; otherwise, the direction opposite to the vertical direction is used as the new preset direction and step S41 is re-executed.

[0013] Preferably, in step S4, the angle estimator is specifically a trained first CNN network; the input of the first CNN network is a pilot signal , the output is the equivalent angle ; The loss function used to supervise the training of the first CNN network is the cross entropy loss function; In step S5, the wind speed estimator is specifically a trained second CNN network; the input of the second CNN network is a pilot signal , the output is wind speed Estimated value of ; The loss function used to supervise the training of the second CNN network is the cross entropy loss function; Each of the signal detectors is a trained third CNN network; the input of the third CNN network is a data signal , the output is a soft bit vector ; The loss function used to supervise the training of the third CNN network is the mean square error loss function; The channel classifier is specifically a trained DNN network, and the input of the DNN network is the wind speed Estimated value of , the output is the weight value corresponding to each signal detector; the loss function used to supervise the DNN network training is a loss function based on the CMSE criterion.

[0014] Preferably, the loss function formula based on the CMSE criterion is specifically:

[0015] in, is the correction factor; is the expected value function; is the weight value corresponding to the i-th signal detector; is the number of signal detectors.

[0016] Preferably, in step S3, the data preprocessing module processes the time domain received signal Perform analog-to-digital conversion, serial-to-parallel conversion, remove the cyclic prefix and perform fast Fourier transform in sequence to complete the preprocessing and obtain the received signal .

[0017] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: The present invention provides a cross-medium visible light communication system and method based on deep learning. On the one hand, an angle estimator and a receiving end movement strategy based on a convolutional neural network (CNN) are proposed. The receiving end can make a movement decision according to the angle estimation result, and adaptively realize the link alignment of the transceiver of the cross-medium visible light communication system. On the other hand, the present invention proposes a signal detection (SD) algorithm based on wind speed estimation (WE) and channel classification (CC); after the system completes link alignment, the algorithm can estimate the wind speed and adaptively detect the received signal transmitted through the channel under different wind speed conditions; in this algorithm, the present invention also proposes a loss function based on the Correction of Mean Square Error (CMSE) criterion, which can effectively alleviate the adverse effects of wave-induced fading in the communication channel of the cross-medium visible light system on signal transmission, thereby reducing the bit error rate of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a structural diagram of a cross-medium visible light communication system based on deep learning provided in Example 1.

[0019] Figure 2 This is a schematic diagram of the W2A-VLC system communication scenario provided in Example 1.

[0020] Figure 3 This is a schematic diagram of the signal receiving plane of the receiving end provided in Example 1.

[0021] Figure 4 This is a flow chart of a cross-medium visible light communication method based on deep learning provided in Example 2.

[0022] Figure 5 This is a flowchart of the receiving end mobile strategy processing provided in Example 2.

[0023] Figure 6 This is a performance diagram of the accuracy of the CNN angle estimator provided in Example 3.

[0024] Figure 7This is an example diagram of the UAV movement trajectory provided in Example 3.

[0025] Figure 8 The bit error rate performance diagram of the CNN signal detector provided in Example 3 Fig. 9 This is the accuracy performance diagram of the CNN wind speed estimator provided in Example 3 Fig.10 For the example 3 provided in BER performance of the WE-CC-SD algorithm at m / s.

[0026] Fig.11 For the example 3 provided in BER performance of the WE-CC-SD algorithm at m / s. DETAILED DESCRIPTION

[0027] The drawings are for illustrative purposes only and should not be construed as limiting the present patent; In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0028] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0029] Example 1 like Figure 1 As shown, this embodiment provides a cross-medium visible light communication system based on deep learning, including: a transmitting end, a transmission channel and a receiving end; the transmitting end and the receiving end are respectively placed in different signal transmission media; The receiving end includes: a signal receiving module, a data preprocessing module, a link alignment module and a signal detection module; The transmitting end is used to generate and transmit an optical signal, and the optical signal reaches the signal receiving module of the receiving end after being transmitted through the transmission channel; The data preprocessing module at the receiving end is used to preprocess the received optical signal to obtain a pilot signal and a data signal; The link alignment module of the receiving end is provided with an angle estimator based on a deep learning algorithm, and the link alignment module is used to calculate the equivalent angle between the receiving end and the transmitting end according to the pilot signal, and adjust the position of the receiving end according to the equivalent angle until the link alignment is achieved; The signal detection module of the receiving end is provided with a wind speed estimator based on a deep learning algorithm, a channel classifier and a plurality of signal detectors. The signal detection module is used to estimate the wind speed according to the pilot signal and calculate the weight value corresponding to each signal detector after the link is aligned, detect the data signal in real time, and realize cross-medium visible light communication; The transmitting end generates an optical signal by optical OFDM modulation, wherein the optical OFDM modulation includes any one of asymmetric limited optical-orthogonal frequency division multiplexing and direct current biased optical-orthogonal frequency division multiplexing; The equivalent angle is the angle between the line connecting the transmitting end and the receiving end and the vertical direction; The vertical distance between the receiving end and the transmitting end is kept constant; the signal receiving plane of the receiving end includes a plurality of concentric circles of different radii, each of the concentric circles represents an angle layer, and the equivalent angle of the receiving end and the transmitting end at any position in each angle layer is the same; The transmitting end is arranged on an autonomous underwater vehicle AUV, and the autonomous underwater vehicle AUV is arranged underwater; The receiving end is arranged on an unmanned aerial vehicle (UAV), and the unmanned aerial vehicle (UAV) is arranged in an airspace at a specific height above the water surface.

[0030] In the specific implementation process, this embodiment proposes a W2A-VLC system based on deep learning for the underwater-to-air (W2A) cross-medium visible light communication scenario, and solves the link alignment problem and adaptive signal detection problem therein; in this embodiment, the designed system can be divided into three parts: AUV transmitter, W2A-VLC channel and UAV receiver; The communication scenario diagram of this embodiment is as follows Figure 2 As shown, the UAV flies in the airspace at a certain height above the water surface and communicates with the AUV transmitter through the W2A-VLC channel; Figure 2 middle, is the distance of the UAV from the horizontal plane, is the distance of the AUV from the horizontal plane, equivalent to the angle is the angle between the line connecting UAV and AUV and the vertical direction; Figure 3 As shown in the figure, it is the flight plane of the UAV receiver. The flight plane includes several concentric circles of different radii. Each concentric circle represents an angle layer. Each angle layer represents A set of equal positions, defined as an angle layer ,in is the number of angle layers; exist Figure 1 middle, represents the pilot signal vector received during the LA process, Indicates the pilot signal received during data transmission. Indicates data signal, represents the estimated equivalent angle, Indicates wind speed The estimated value of express Input signal detector i The output soft bit vector of express Corresponding weight values; the communication process of this system is as follows: The AUV transmitter will transmit a bit vector After mapping and optical OFDM modulation, the time domain transmission signal vector is obtained ,and After being transmitted through the W2A-VLC channel, it reaches the receiving end and undergoes subsequent processing; this system can adopt a variety of common OOFDM schemes, such as asymmetric limited optical-orthogonal frequency division multiplexing, DC biased optical-orthogonal frequency division multiplexing, etc. At the receiving end, the signal receiving module receives the signal and performs photoelectric conversion to obtain the time domain received signal ; The data preprocessing module receives the time domain signal Perform analog-to-digital conversion, serial-to-parallel conversion, remove the cyclic prefix and complete the fast Fourier transform to obtain the received signal , which includes the pilot signal and data signals ; The link alignment module is based on Determine whether the UAV is aligned and make a movement decision until the alignment conditions are met, the UAV hovers at the alignment position, and activates the signal detection module; The signal detection module is based on Perform signal detection to obtain the estimated soft bit vector of the transmitted bit vector , and then recover the data bits through hard decision to complete the detection of W2A-VLC signal; The system performs link alignment based on angle estimation and adaptive channel selection and signal detection based on wind speed estimation. Compared with the existing technology, the system can realize adaptive and continuous alignment of the UAV receiver to the AUV transmitter. It can also realize adaptive wave mitigation and low bit error rate signal detection under dynamic channels.

[0031] Example 2 like Figure 4 As shown, this embodiment provides a cross-medium visible light communication method based on deep learning, based on a cross-medium visible light communication system based on deep learning described in Embodiment 1, comprising the following steps: S1: The transmitter will transmit a bit vector After mapping and optical modulation, the time domain transmit signal vector is generated and sending it to the receiving end through the transmission channel; S2: The signal receiving module receives the time domain transmission signal vector And perform photoelectric conversion to obtain the time domain received signal ; S3: The data preprocessing module processes the received signal in the time domain Perform preprocessing to obtain the received signal , the received signal Including pilot signal and data signals ; S4: In the link alignment module, the angle estimator is based on the pilot signal Calculate the equivalent angle between the receiver and the transmitter , and according to the equivalent angle Adjust the position of the receiving end until link alignment is achieved, and execute step S5; S5: In the signal detection module, the wind speed estimator detects the wind speed according to the pilot signal. Estimated wind speed And input channel classifier, the channel classifier is based on wind speed Calculate the weight value corresponding to each signal detector; The signal detector detects the data signal in real time , and calculate the soft bit vector according to the corresponding weight value ,Recovering data bits through hard decision, and realizing cross-medium visible light communication; In step S4, according to the equivalent angle Adjusting the position of the receiving end until link alignment is achieved includes: Real-time according to the equivalent angle Determine the angle layer of the receiver , according to the angle layer Determine the next moving step of the receiving end; Real-time based on the current equivalent angle Determine the next moving direction of the receiving end, move the receiving end along the moving direction by one step each time, and continuously move the position of the receiving end until the current equivalent angle Less than or equal to the preset alignment angle threshold , fixing the receiving end position to achieve link alignment; The angle layer The larger the radius, the larger the next moving step of the receiving end; The real-time equivalent angle Determining the next moving direction of the receiving end, moving the receiving end along the moving direction by a step length each time, and continuously moving the position of the receiving end includes: S41: the receiving end moves a step length along a preset direction; S42: Determine the current equivalent angle Is it less than or equal to the equivalent angle before moving? If yes, the receiving end keeps the preset direction unchanged, continues to move one step along the preset direction, and re-executes step S42; otherwise, executes step S43; S43: the receiving end returns to the last position and moves a step length in the vertical direction on either side of the preset direction; After moving in the vertical direction on either side of the preset direction, determine the current equivalent angle Is it less than or equal to the equivalent angle before moving? If yes, the vertical direction is used as a new preset direction and step S41 is executed again; otherwise, the direction opposite to the vertical direction is used as a new preset direction and step S41 is executed again; In step S4, the angle estimator is specifically a trained first CNN network; the input of the first CNN network is a pilot signal , the output is the equivalent angle ; The loss function used to supervise the training of the first CNN network is the cross entropy loss function; In step S5, the wind speed estimator is specifically a trained second CNN network; the input of the second CNN network is a pilot signal , the output is wind speed Estimated value of ; The loss function used to supervise the training of the second CNN network is the cross entropy loss function; Each of the signal detectors is a trained third CNN network; the input of the third CNN network is a data signal , the output is a soft bit vector ; The loss function used to supervise the training of the third CNN network is the mean square error loss function; The channel classifier is specifically a trained DNN network, and the input of the DNN network is the wind speed Estimated value of , the output is the weight value corresponding to each signal detector; the loss function used to supervise the DNN network training is a loss function based on the CMSE criterion; The loss function formula based on the CMSE criterion is specifically:

[0032] in, is the correction factor; is the expected value function; is the weight value corresponding to the i-th signal detector; is the number of signal detectors; In step S3, the data preprocessing module processes the time domain received signal Perform analog-to-digital conversion, serial-to-parallel conversion, remove the cyclic prefix and perform fast Fourier transform in sequence to complete the preprocessing and obtain the received signal .

[0033] In the specific implementation process, the transmitter first transmits the bit vector After mapping and optical modulation, the time domain transmit signal vector is generated and sent to the receiving end through a transmission channel; Then the signal receiving module receives the time domain transmitted signal vector And perform photoelectric conversion to obtain the time domain received signal ; The data preprocessing module receives the time domain signal Perform analog-to-digital conversion, serial-to-parallel conversion, remove the cyclic prefix and perform fast Fourier transform in sequence to complete the preprocessing and obtain the received signal , receiving signal Including pilot signal and data signals ; Then the link alignment module is used to perform link alignment. Figure 2 In the process, the AUV transmitter transmits signals in the vertical direction, and the UAV receiver is at a constant height of On the plane, according to the result of angle estimation, get the current angle layer , and then adopt the corresponding moving strategy, move along the trajectory of the "zigzag spiral" with an appropriate step length each time, and finally reach the alignment point; Specifically, the angle estimator is based on the pilot signal Calculate the equivalent angle between the receiver and the transmitter , the angle estimator is the trained first CNN network, and the input is the pilot signal , the output is the equivalent angle ; Then, in real time, according to the equivalent angle Determine the angle layer of the receiver , according to the angle layer Determine the next moving step of the receiver; Angle layer The larger the radius, the larger the next moving step of the receiving end (that is, the outer angle layer corresponds to a longer step, and the inner angle layer corresponds to a shorter step); Real-time based on the current equivalent angle Determine the next moving direction of the receiver, move the receiver one step along the moving direction each time, and keep moving the receiver until the current equivalent angle is reached. Less than or equal to the preset alignment angle threshold , fix the receiving end position and achieve link alignment; like Figure 5 As shown, according to the current equivalent angle Flowchart for determining the mobility strategy. The UAV adjusts the position of the receiver according to the mobility strategy. The specific steps are as follows: S41: The receiving end first moves a step along a preset direction (random direction); S42: Determine the current equivalent angle Is it less than or equal to the equivalent angle before moving? If yes, the receiving end keeps the preset direction unchanged, continues to move one step along the preset direction, and re-executes step S42; otherwise, executes step S43; S43: the receiving end returns to the last position and moves a step length in the vertical direction on either side of the preset direction; After moving in the vertical direction on either side of the preset direction, determine the current equivalent angle Is it less than or equal to the equivalent angle before moving? If yes, the vertical direction is used as a new preset direction and step S41 is executed again; otherwise, the direction opposite to the vertical direction is used as a new preset direction and step S41 is executed again; It is worth mentioning that after link alignment, the wind speed is estimated It can be used as auxiliary information to facilitate the UAV to better adjust its control parameters and maintain the hovering state at the alignment point to facilitate the link to be persistently aligned; for example, the UAV can adjust the speed of the corresponding rotor motor according to the estimated wind speed to compensate for the displacement and attitude changes caused by the wind; in addition, the hovering control can also be achieved according to the special technology of the UAV manufacturer to maintain the hovering state at the alignment point; if the UAV is offset due to environmental factors, the above link alignment process can be executed again to re-acquire the alignment point; since the UAV hovering control technology belongs to other professional fields such as automatic control and machinery, it does not belong to the protection content of the present invention, and will not be described in detail here; After link alignment is completed, the system can perform data communication, and the signal detection module is activated at this time; exist The time domain received signal at time for:

[0034] in, Transmit signal in time domain; is the impulse response; * is the time domain convolution; for noise; By wind speed The Pearson-Mills (PM) wave power spectrum is determined as:

[0035] in: is a constant factor, is the acceleration due to gravity, is the angular frequency; according to the above formula, different The values ​​form different wave surfaces, which affect the channel through refraction on the sea surface. Causes wave-induced fading, which in turn leads to a decrease in system detection performance; In order to achieve wave mitigation, this method proposes a signal detection module consisting of a wind speed estimator composed of a CNN network, several signal detectors, and a channel classifier composed of a deep neural network (DNN). Among them, the CNN signal detector can be trained to detect the received data signal. Get the estimated soft bit vector ; In order to improve the system's adaptability to the corresponding channels under untrained wind speed conditions, this method targets Training at different wind speeds A set of CNN signal detectors, which will be used for signal detection under various unknown wind speeds; If the wind speed is unknown, enter Set CNN signal detector to get output , the CNN wind speed estimator uses the received pilot signal Estimate the wind speed corresponding to the channel at this time , and the estimated value Input DNN channel classifier to get A single combined weight vector for each CNN detector ,but It can be obtained by the following formula:

[0036] In addition, according to the formula of wave power spectrum, in order to ensure that the decision threshold of signal detection remains unchanged, it should be satisfied ; Therefore, this method designs a loss function based on the CMSE criterion for the DNN channel classifier to help the system achieve better bit error rate performance. The formula is as follows:

[0037] in, is the correction factor; is the expected value function; is the weight value corresponding to the i-th signal detector; is the number of signal detectors; After the link alignment is completed, the signal detection module is used to perform signal detection, wherein the wind speed estimator is based on the pilot signal Estimated wind speed And input the channel classifier, the channel classifier is based on the wind speed Calculate the weight value corresponding to each signal detector; Signal detector detects data signal in real time , and calculate the soft bit vector according to the corresponding weight value ,Recovering data bits through hard decision, and realizing cross-medium visible light communication; This method proposes an angle estimator based on a convolutional neural network and a receiving end movement strategy. The receiving end can make movement decisions based on the angle estimation results and adaptively realize the link alignment of the transceiver of the cross-medium visible light communication system. In addition, this method proposes a signal detection algorithm based on wind speed estimation and channel classification. After the system completes the link alignment, the algorithm can estimate the wind speed and adaptively detect the received signal transmitted through the channel under different wind speed conditions. In this algorithm, this method also proposes a loss function based on the modified mean square error criterion, which can effectively alleviate the adverse effects of wave-induced fading in the communication channel of the cross-medium visible light system on signal transmission, thereby reducing the bit error rate of the system.

[0038] Example 3 This embodiment provides a simulation experiment to verify the effectiveness of a cross-medium visible light communication method based on deep learning proposed in Embodiment 2.

[0039] In the specific implementation process, in this embodiment, DCO-OFDM is selected as the modulation scheme, the DC offset is set to 13dB, the number of subcarriers is 64, and QPSK modulation is adopted; each signal frame contains 10 data OFDM symbols and 1 pilot OFDM symbol; the W2A-VLC channel is obtained based on the Monte Carlo simulation model, and the distance between the UAV and the horizontal plane is m, distance between AUV and horizontal plane m, PD (Photo Diode, photodetector) diameter is m, LED light source wavelength is nm, LED half power angle is , the number of simulated photons is , PD half field of view is ; The simulation uses the Jerlov I water body in the Jerlov water body classification, and its absorption and scattering coefficients , for ; In the simulation of the link alignment module, the network structure and main parameters of the CNN angle estimator are shown in Table 1, where: is the number of convolutional layers, is the number of convolution kernels, is the convolution kernel dimension, is the convolution step size, is the number of fully connected layers, is the number of neurons in the fully connected layer; Table 1: Main parameters of CNN angle estimator

[0040] This embodiment sets Angle layer, that is ,in: ; During training, take wind speed m / s, randomly select 4 positions on each of the above angle layers and obtain Received signal , obtained from all locations constitute the training set; in particular, since The mean square error of the channel sample values ​​in the inner distribution is less than , so choose As the alignment angle threshold; since the receiver plane area range corresponding to the same angle varies , The value of is different, so it should be based on , Choose the appropriate value of Value; In addition, the training strategy adopted in this embodiment is to respectively The following trains are performed for 100 epochs each; During the test, the wind speed was set at 4 m / s condition, under Inside, and between, and between, and Four sub-angle layers are randomly selected between each sub-angle layer, and each sub-angle layer is collected separately at random positions The data is used as a test set; Finally, set the input vector dimension of the CNN network to , the training set size is input vectors, and the test set size is input vector; the accuracy performance of the CNN angle estimator is as follows Figure 6 As shown; like Figure 6 As shown, the angle estimator proposed by the present invention is effective under high SNR conditions (such as ) can achieve 100% accuracy; at the same time, The network trained at m / s can also adapt to various wind speed conditions that have not been trained, which can prove the robustness of the solution proposed in Example 2; Further, Figure 7 An example of the trajectory of the UAV moving in the receiver plane is provided, where the angles of the angle layer from inside to outside are , numbers 1 to 10 represent the order of the UAV movement trajectory, the thick arrow represents the movement direction of the UAV, and the thin arrow represents each historical movement trajectory; Figure 7 It can be seen that the UAV can reach the alignment point after several moves, which proves the effectiveness of the solution proposed in Example 2; After the alignment is completed, the W2A-VLC system can perform data communication; in the simulation of the WE-CC-SD algorithm, the network structures and main parameters are shown in Table 2; Table 2: Main parameters of WE-CC-SD

[0041] The simulation of the WE-CC-SD algorithm is divided into the following three parts: 1) Part 1: Using data signals The CNN signal detector is trained and tested. The training strategy is the same as above. The specific parameters are shown in Table 2. The output of the training network is the estimated soft bit vector ; This embodiment selects m / s (still water) and The data generated by the channel at m / s is used to train and test the CNN signal detector; the bit error rate performance of the two CNN signal detectors is shown in Figure 8 As shown; Depend on Figure 8 It can be seen that the performance of the separately trained CNN signal detector is close to the ideal scheme assuming perfect channel state information (CSI) conditions, and is better than the least squares (LS) estimation and linear minimum mean square error (LMMSE) estimation schemes in the presence of wave-induced fading; 2) The second part: the received pilot signal Used for training and testing of CNN wind speed estimator; generated in m / s wind speed As the input data for training, the real wind speed The one-hot encoding of the true value is used to train the network output corresponding to the estimated wind speed , the loss function is the cross entropy of the two; the accuracy performance of the CNN wind speed estimator is shown in Fig. 9 ,It can be seen that the CNN wind speed estimator has a higher accuracy under ,high SNR conditions; 3) Part III: Training the DNN channel classifier and testing the overall performance of the WE-CC-SD algorithm; m / s wind speed Input the signal detector trained in the first part respectively to obtain 11 groups corresponding to 11 wind speeds. Then 11 groups and corresponding wind speed The one-hot encoding is used as the input of the channel classifier; then, the network is trained using the CMSE loss function proposed in Example 2 to obtain the corresponding combined weights ; Fig.10 and Fig.11 Tested separately m / s and BER performance of the WE-CC-SD algorithm at m / s; Depend on Fig.10 and Fig.11 It can be seen that the bit error rate performance of the WE-CC-SD algorithm is always better than the traditional solution, which verifies that the algorithm can make the system Get better bit error rate performance under the conditions.

[0042] The same or similar reference numerals correspond to the same or similar components; The terms used in the drawings to describe positional relationships are only used for illustrative purposes and should not be construed as limiting this patent; Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A cross-medium visible light communication system based on deep learning, characterized in that: include: Transmitter, transmission channel and receiver; The transmitting end and the receiving end are respectively placed in different signal transmission media; The receiving end includes: a signal receiving module, a data preprocessing module, a link alignment module and a signal detection module; The transmitting end is used to generate and transmit an optical signal, and the optical signal reaches the signal receiving module of the receiving end after being transmitted through the transmission channel; The data preprocessing module at the receiving end is used to preprocess the received optical signal to obtain a pilot signal and a data signal; The link alignment module of the receiving end is provided with an angle estimator based on a deep learning algorithm, and the link alignment module is used to calculate the equivalent angle between the receiving end and the transmitting end according to the pilot signal, and adjust the position of the receiving end according to the equivalent angle until the link alignment is achieved; The signal detection module at the receiving end is provided with a wind speed estimator based on a deep learning algorithm, a channel classifier and several signal detectors. The signal detection module is used to estimate the wind speed according to the pilot signal after the link is aligned and calculate the weight value corresponding to each signal detector, detect the data signal in real time, and realize cross-medium visible light communication.

2. According to claim 1, a cross-medium visible light communication system based on deep learning is characterized in that: The transmitting end generates an optical signal through optical OFDM modulation, and the optical OFDM modulation includes: any one of asymmetric limited optical-orthogonal frequency division multiplexing and direct current biased optical-orthogonal frequency division multiplexing.

3. The cross-medium visible light communication system based on deep learning according to claim 1, characterized in that: The equivalent angle is the angle between the line connecting the transmitting end and the receiving end and the vertical direction; The vertical distance between the receiving end and the transmitting end remains constant; the signal receiving plane of the receiving end includes a plurality of concentric circles with different radii, each of which represents an angle layer, and the equivalent angle of the receiving end and the transmitting end at any position in each angle layer is the same.

4. A cross-medium visible light communication system based on deep learning according to any one of claims 1 to 3, characterized in that: The transmitting end is arranged on an autonomous underwater vehicle AUV, and the autonomous underwater vehicle AUV is arranged underwater; The receiving end is arranged on an unmanned aerial vehicle (UAV), and the unmanned aerial vehicle (UAV) is arranged in an airspace at a specific height above the water surface.

5. A cross-medium visible light communication method based on deep learning, based on a cross-medium visible light communication system based on deep learning as described in any one of claims 1 to 4, characterized in that: The following steps are involved: S1: The transmitter will transmit a bit vector After mapping and optical modulation, the time domain transmit signal vector is generated and sending it to the receiving end through the transmission channel; S2: The signal receiving module receives the time domain transmission signal vector And perform photoelectric conversion to obtain the time domain received signal ; S3: The data preprocessing module processes the received signal in the time domain Perform preprocessing to obtain the received signal , the received signal Including pilot signal and data signals ; S4: In the link alignment module, the angle estimator is based on the pilot signal Calculate the equivalent angle between the receiver and the transmitter , and according to the equivalent angle Adjust the position of the receiving end until link alignment is achieved, and execute step S5; S5: In the signal detection module, the wind speed estimator detects the wind speed according to the pilot signal. Estimated wind speed And input channel classifier, the channel classifier is based on wind speed Calculate the weight value corresponding to each signal detector; The signal detector detects the data signal in real time , and calculate the soft bit vector according to the corresponding weight value , after recovering the data bits through hard decision, cross-medium visible light communication is realized.

6. A cross-medium visible light communication method based on deep learning according to claim 5, characterized in that: In step S4, according to the equivalent angle Adjusting the position of the receiving end until link alignment is achieved includes: Real-time according to the equivalent angle Determine the angle layer of the receiver , according to the angle layer Determine the next moving step of the receiving end; Real-time based on the current equivalent angle Determine the next moving direction of the receiving end, move the receiving end along the moving direction by one step each time, and continuously move the position of the receiving end until the current equivalent angle Less than or equal to the preset alignment angle threshold , fix the receiving end position and achieve link alignment.

7. The method for cross-medium visible light communication based on deep learning according to claim 6, characterized in that: The angle layer The larger the radius, the larger the next moving step of the receiving end; The real-time equivalent angle Determining the next moving direction of the receiving end, moving the receiving end along the moving direction by a step length each time, and continuously moving the position of the receiving end includes: S41: the receiving end moves a step length along a preset direction; S42: Determine the current equivalent angle Is it less than or equal to the equivalent angle before moving? If yes, the receiving end keeps the preset direction unchanged, continues to move one step along the preset direction, and re-executes step S42; otherwise, executes step S43; S43: the receiving end returns to the last position and moves a step length in the vertical direction on either side of the preset direction; After moving in the vertical direction on either side of the preset direction, determine the current equivalent angle Is it less than or equal to the equivalent angle before moving? If so, the vertical direction is used as the new preset direction and step S41 is re-executed; otherwise, the direction opposite to the vertical direction is used as the new preset direction and step S41 is re-executed.

8. The method for cross-medium visible light communication based on deep learning according to claim 5, characterized in that: In step S4, the angle estimator is specifically a trained first CNN network; the input of the first CNN network is a pilot signal , the output is the equivalent angle ; The loss function used to supervise the training of the first CNN network is the cross entropy loss function; In step S5, the wind speed estimator is specifically a trained second CNN network; the input of the second CNN network is a pilot signal , the output is wind speed Estimated value of ; The loss function used to supervise the training of the second CNN network is the cross entropy loss function; Each of the signal detectors is a trained third CNN network; the input of the third CNN network is a data signal , the output is a soft bit vector ; The loss function used to supervise the training of the third CNN network is the mean square error loss function; The channel classifier is specifically a trained DNN network, and the input of the DNN network is the wind speed Estimated value of , the output is the weight value corresponding to each signal detector; the loss function used to supervise the DNN network training is a loss function based on the CMSE criterion.

9. The method for cross-medium visible light communication based on deep learning according to claim 8, characterized in that: The loss function formula based on the CMSE criterion is specifically: in, is the correction factor; is the expected value function; is the weight value corresponding to the i-th signal detector; is the number of signal detectors.

10. A cross-medium visible light communication method based on deep learning according to any one of claims 5 to 9, characterized in that: In step S3, the data preprocessing module processes the time domain received signal Perform analog-to-digital conversion, serial-to-parallel conversion, remove the cyclic prefix and perform fast Fourier transform in sequence to complete the preprocessing and obtain the received signal .

Images