Complex signal adaptive compensation method adopting extraction and fusion strategy

Abstract

The invention relates to the technical field of artificial intelligence and the field of free space optical communication in the optical communication technology, in particular to a complex signal adaptive compensation method adopting an extraction and fusion strategy, which comprises the following steps: constructing a deep learning equalizer which comprises a signal receiving module, a damage classification module, a signal equalization module and a signal output module; wherein a plurality of equalization networks with different complexities are integrated in the signal equalization module, each equalization network corresponds to one damage type, and each equalization network comprises an extraction sub-network and a fusion sub-network; adaptive compensation of complex signals is realized through a deep learning equalizer; according to the method, the optimal equalization network matched with the current system state can be automatically selected, dynamic optimization of equalization processing is realized, and the real-time performance and the resource utilization efficiency of the system are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology and the field of free space optical communication in optical communication technology, and in particular to a complex signal adaptive compensation method using an extraction and fusion strategy. Background Art

[0002] Visible Light Communication (VLC) is an emerging high-speed wireless communication method that uses visible light as an information carrier, modulates optical signals through current drive of light-emitting diodes (LEDs), and converts optical signals into electrical signals at the receiving end to complete data transmission. It does not require wired media such as optical fibers. Its low cost, high efficiency, and high security make it an important part of future communication technology and is widely used in indoor positioning, vehicle networks, and underwater communications. However, VLC also faces challenges such as noise interference and nonlinear effects. When the signal passes through a system affected by linear and nonlinear damage, distortion and distortion may occur. In severe cases, it may even cause the receiving system to be unable to accurately identify the signal, thereby significantly reducing the communication quality.

[0003] Signal equalization technology is used to compensate for signal damage during transmission, and the widespread application of complex signals, especially in efficient data transmission technology, has put forward higher requirements for equalization technology. For example, the QAM-CAP modulation method transmits multiple bits through complex symbols, which improves transmission efficiency, but processing complex signals requires both retaining the relationship between the real part and the imaginary part and avoiding over-design of the neural network. Currently, neural networks are widely used in visible light communication signal equalization, in which the size of the receptive field is a key factor affecting the signal equalization effect. Although increasing the number of network layers can expand the receptive field and thus improve the equalization performance, this method will increase the complexity of the network, may cause gradient problems, and reduce the practicality in resource-constrained scenarios.

[0004] In summary, current equalization technology has the following problems: (1) The complexity of complex signal processing is relatively high, especially how to effectively preserve the relationship between the real part and the imaginary part, which is not adequately supported by existing methods; (2) Although increasing the complexity of the network structure can help improve performance, it will increase the network computing cost, which may cause gradient problems and reduce the feasibility of practical applications; (3) The ability to compensate for linear and nonlinear damage is still insufficient, making it difficult to effectively deal with diverse signal damage in complex communication environments. Summary of the invention

[0005] To solve the above problems, the present invention provides a complex signal adaptive compensation method using an extraction and fusion strategy, including constructing a deep learning equalizer, which includes a signal receiving module, a damage classification module, a signal equalization module and a signal output module; wherein the signal equalization module integrates a plurality of equalization networks of different complexity, each equalization network corresponds to a damage type, and each equalization network includes an extraction subnetwork and a fusion subnetwork; the process of realizing complex signal adaptive compensation by a deep learning equalizer includes

[0006] S1. Using a signal receiving module to obtain a receiving end complex signal, the receiving end complex signal includes a normal transmission complex signal and a receiving pilot signal;

[0007] S2. Input the received pilot signal into the damage classification module to obtain the damage type;

[0008] S3. Input the normal transmission complex signal into the signal equalization module, use the equalization network corresponding to the damage type to perform compensation processing, and obtain comprehensive damage information;

[0009] S4. Input the normal transmission complex signal and the comprehensive damage information into the signal output module, and output the compensation signal.

[0010] Furthermore, step S1 specifically includes:

[0011] The transmitter embeds a preset pilot signal at the front end of the normal transmission signal to form a transmission signal, and then uses LED to perform intensity modulation on the transmission signal to obtain an optical signal;

[0012] The receiving end receives the optical signal transmitted through the channel and converts it into an electrical signal;

[0013] The signal receiving module distinguishes the electrical signal to obtain a received normal transmission signal and a received pilot signal; determines whether the received normal transmission signal adopts a complex modulation method. If so, the received normal transmission signal is directly used as a normal transmission complex signal. If not, the received normal transmission signal is mapped to the complex domain to obtain a normal transmission complex signal.

[0014] Furthermore, the damage classification module uses a multi-layer perceptron, which receives the pilot signal as input and the damage type as output, expressed as

[0015] C=argmax(σ(W2·φ(W1·presig+b1)+b2))

[0016] Among them, C represents the output of the loss classification module, presig represents the reception of the pilot signal, W1 and W2 represent weight matrices, b1 and b2 represent biases, φ() represents a nonlinear activation function, σ() represents a softmax activation function, and argmax() represents the selection of the category corresponding to the maximum probability.

[0017] Furthermore, each equalizing network in the signal equalizing module is constructed based on a convolutional neural network, and the hyperparameters of each equalizing network are set differently, including the number of hidden layers, the convolution void ratio, and the convolution kernel dimension.

[0018] Further, step S3 uses an equalization network corresponding to the damage type to perform compensation processing, including:

[0019] S31. For the real part signal Rx in the normal transmission complex signal RE and the imaginary signal Rx IM , the extraction sub-network is used to extract the real part loss information Feature RE and imaginary damage information Feature IM , expressed as

[0020] Feature RE =φ1(W extract ·Rx RE +b)

[0021] Feature IM =φ1(W extract ·Rx IM +b)

[0022] Among them, W extract represents the weight matrix, b represents the convolution bias, and φ1() represents the nonlinear activation function;

[0023] S32. Feature the real part loss information RE and imaginary damage information Feature IM Splicing to get features

[0024] S33.Characteristics Input the fusion sub-network to obtain the comprehensive damage information Feature, which is expressed as

[0025]

[0026] Among them, W fuse 1. W fuse 2 represents the weight matrix, b3 and b4 represent the bias, and φ2() represents the nonlinear activation function.

[0027] Furthermore, in step S4, in the signal output module, the normal transmission complex signal is subtracted from the comprehensive damage information to obtain a compensation signal. Finally, the compensation signal is subjected to corresponding demodulation processing to obtain an output data stream.

[0028] Beneficial effects of the present invention:

[0029] The method proposed in the present invention has the ability to perceive the strength of system damage effects and can automatically select the best balancing network that matches the current system state, thereby achieving dynamic optimization of balancing processing and improving the real-time performance and resource utilization efficiency of the system.

[0030] The signal equalization module adopts an extraction and fusion strategy, which not only simplifies the processing flow of complex signals, but also fully preserves the integrity and phase relationship of the signal. The high receptive field convolution kernel significantly improves the quality of feature extraction and fusion without increasing the complexity of the network, enabling the network to adapt to complex damage scenarios more efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a structural diagram of the optical communication system of the present invention;

[0032] Figure 2 This is a schematic diagram of the structure of the deep learning equalizer of the present invention;

[0033] Figure 3 This is a schematic diagram of a signal equalization module of the present invention;

[0034] Figure 4 Schematic diagram of the high receptive field convolution kernel of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] The present invention provides a complex signal adaptive compensation method using an extraction and fusion strategy, wherein the visible light communication system structure used is as follows: Figure 1 As shown, it mainly includes 5 parts: data input, signal modulation, LED signal transmission, PIN signal reception, signal equalization, signal demodulation and data output.

[0037] Preferably, in the data input part, the present invention divides the input data stream into two parts: a pilot signal and a normal transmission signal. Specifically, at the transmitting end, the pilot signal is embedded into the communication data stream through a special encoding method and transmitted together with the normal transmission signal. The pilot signal and the normal transmission signal are consistent in form, but the pilot signal is a fixed-length system known data and is sent at the beginning of a round of communication. In particular, a redundant check field is added to the pilot signal to ensure that even if there is a certain interference, the pilot signal and the normal transmission signal can be accurately distinguished through the verification mechanism. Subsequently, the pilot signal will be identified at the receiving end based on the redundant check field to distinguish it from the normal transmission signal. The transmission characteristics of the entire channel can be roughly judged based on the degree of distortion of the pilot signal, which is also the basis for damage classification and adaptive signal equalization.

[0038] Preferably, at the signal modulation part of the transmitting end, the transmitting signal (a signal consisting of a normal transmission signal plus a pilot signal) is preprocessed and coded and modulated, and then the LED is driven for intensity modulation to convert the transmitting signal from an electrical signal to an optical signal.

[0039] Preferably, in the LED signal transmission part, the communication system will be affected by noise interference and system distortion, which is mainly divided into linear damage and nonlinear damage. Linear damage mainly comes from inter-symbol crosstalk between adjacent symbols and multipath effect in light propagation; nonlinear damage is caused by nonlinear characteristics of system components and square law detection of receivers.

[0040] Therefore, the signal received by the receiving end after the normal transmission signal is transmitted through the channel can be expressed as:

[0041] Rx=Tx+Distortion+noise,

[0042] Among them, Rx represents receiving normal transmission signals, Tx represents normal transmission signals, Distortion represents system distortion, and noise represents noise.

[0043] Preferably, in the signal equalization part, a deep learning equalizer is constructed, which includes a signal receiving module, a damage classification module, a signal equalization module and a signal output module. The present invention utilizes the powerful fitting ability of neural networks to achieve signal equalization, but the type and intensity of signal damage will be different in different transmission environments and system equipment. Although a single complex network theoretically has the ability to handle diversified damage, it has obvious deficiencies in computing resources, real-time and generalization capabilities. Therefore, the signal equalization module of the present invention integrates multiple equalization networks of different complexities to improve the equalization efficiency and adaptability of the system. Each equalization network corresponds to a type of damage, and each equalization network includes an extraction subnetwork and a fusion subnetwork. Each subnetwork is constructed based on a convolutional neural network, and includes an input layer, a hidden layer and an output layer.

[0044] like Figure 2 As shown, the process of implementing adaptive compensation of complex signals through deep learning equalizer includes

[0045] S1. Using a signal receiving module to obtain a receiving end complex signal, the receiving end complex signal includes a normal transmission complex signal and a receiving pilot signal.

[0046] Specifically, the signal receiving module performs corresponding processing according to the modulation form of the input signal to provide a consistent input form for subsequent modules. If the normal transmission signal received adopts a complex modulation method (such as CAP modulation), it is directly passed to the subsequent equalization module in the form of real and imaginary parts for processing. If the normal transmission signal received adopts a real modulation method (such as PAM modulation), the signal receiving module first performs complex domain mapping. For example, for the sampling signal sequence a, b, c, d, complex symbols can be generated as follows:

[0047] s1=a+jc,s2=b+jd,

[0048] Where s1 and s2 represent the generated complex symbols, and j is the imaginary unit. By mapping adjacent symbols to the real and imaginary parts of complex numbers, the correlation between symbols can be fully preserved.

[0049] S2. Input the received pilot signal into the damage classification module to obtain the damage type.

[0050] Specifically, the damage classification module is based on the multi-layer perceptron (MLP) structure, takes the received pilot signal as input, and uses the distortion information in the received pilot signal to quickly evaluate the damage intensity in the communication environment through a lightweight classification task, and outputs the corresponding classification results. The classification results correspond one-to-one to the preset equalization network, so that the matching equalization network is selected for signal equalization processing. The process can be expressed as:

[0051] C=argmax(σ(W2·φ(W1·presig+b1)+b2))

[0052] Among them, C represents the output of the loss classification module, presig represents the reception of the pilot signal, W1 and W2 represent weight matrices, b1 and b2 represent biases, φ() represents a nonlinear activation function, σ() represents a softmax activation function, and argmax() represents the selection of the category corresponding to the maximum probability.

[0053] S3. Input the normally transmitted complex signal into the signal equalization module, use the equalization network corresponding to the damage type to perform compensation processing, and obtain comprehensive damage information.

[0054] Specifically, Figure 3 As shown in the figure, the signal equalization module integrates multiple equalization networks of different complexity, and an equalization network is determined by the damage classification result to perform the current round of damage compensation task. The difference in network complexity is achieved by setting different hyperparameters, including the number of hidden layers, convolutional dilation rate, and convolution kernel dimension.

[0055] Step S3 uses an equalization network corresponding to the damage type to perform compensation processing, including:

[0056] S31. For the real part signal Rx in the normal transmission complex signal RE and the imaginary signal Rx IM , the extraction sub-network is used to extract the real part loss information Feature RE and imaginary damage information Feature IM , expressed as

[0057] Feature RE =φ1(W extract ·Rx RE +b)

[0058] Feature IM =φ1(W extract ·Rx IM +b)

[0059] Among them, W extract represents the weight matrix, b represents the convolution bias, and φ1() represents the nonlinear activation function;

[0060] S32. Feature the real part loss information RE and imaginary damage information Feature IM Splicing to get features

[0061] S33.Characteristics Input the fusion sub-network to obtain the comprehensive damage information Feature, which is expressed as

[0062]

[0063] Among them, W fuse 1. W fuse 2 represents the weight matrix, b3 and b4 represent the bias, and φ2() represents the nonlinear activation function.

[0064] Specifically, each convolution layer in the fusion sub-network uses a high receptive field convolution kernel. Figure 4 As shown in the figure, unlike conventional convolutional layers, high receptive field convolution kernels set the hole value and padding value of the convolutional layer so that the convolution sampling area does not overlap when the number of samples is limited. The benefit of this method is that it significantly improves the efficiency of feature extraction, thereby capturing a wider range of feature information in a simpler network structure and enhancing the network's overall understanding of the input data. The calculation method of the hole value and padding value can be expressed as

[0065] dilationrate=kernelsize

[0066] padding=(kernelsize-1)×dilation÷2

[0067] Among them, kernelsize represents the convolution kernel size, dilationrate represents the hole value, and padding represents the padding value. By setting the padding value through the above formula, the length of the feature map after convolution can be ensured to be consistent with the input length.

[0068] S4. Input the normal transmission complex signal and the comprehensive damage information into the signal output module, and output the compensation signal.

[0069] Specifically, in step S4, in the signal output module, a compensation signal is obtained by subtracting the comprehensive damage information from the normal transmission complex signal.

[0070] Specifically, in the training process of the deep learning equalizer, the complex mean square error function CMSE is used to measure the training loss, which can be expressed as:

[0071] Loss=CMSE(Tx,Tx′)=MSE(Tx RE ′,Tx RE )+MSE(Tx IM ′,Tx IM )

[0072] Among them, Tx' represents the compensation signal. During the training process, the gradient descent method is used to update the weights to minimize the loss value of the network and achieve the optimal signal compensation effect.

[0073] The recovered signal is processed by corresponding signal demodulation and finally the output data stream is obtained.

[0074] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "rotation" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0075] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A complex signal adaptive compensation method using an extraction and fusion strategy, characterized in that: A deep learning equalizer is constructed, which includes a signal receiving module, an impairment classification module, a signal equalization module and a signal output module; wherein the signal equalization module integrates multiple equalization networks of different complexity, each equalization network corresponds to a type of impairment, and each equalization network includes an extraction subnetwork and a fusion subnetwork; the process of realizing adaptive compensation of complex signals through a deep learning equalizer includes: S1. Using a signal receiving module to obtain a complex signal at the receiving end, the complex signal at the receiving end includes a normal transmission complex signal and a receiving pilot signal; S2. Input the received pilot signal into the damage classification module to obtain the damage type; S3. Input the normal transmission complex signal into the signal equalization module, use the equalization network corresponding to the damage type to perform compensation processing, and obtain comprehensive damage information; S4. Input the normal transmission complex signal and the comprehensive damage information into the signal output module, and output the compensation signal.

2. The method for adaptively compensating complex signals using an extraction and fusion strategy according to claim 1, characterized in that: Step S1 specifically includes: The transmitter adds a preset pilot signal to the front end of the normal transmission signal to form a transmission signal, encodes and modulates the transmission signal, and then uses LED to perform intensity modulation to obtain an optical signal; The receiving end receives the optical signal transmitted through the channel and converts it into an electrical signal; The signal receiving module distinguishes the electrical signal to obtain a received normal transmission signal and a received pilot signal; determines whether the received normal transmission signal adopts a complex modulation method. If so, the received normal transmission signal is directly used as a normal transmission complex signal. If not, the received normal transmission signal is mapped to the complex domain to obtain a normal transmission complex signal.

3. The method for adaptively compensating complex signals using an extraction and fusion strategy according to claim 1, characterized in that: The damage classification module uses a multi-layer perceptron, with the received pilot signal as input and the damage type as output, expressed as C = argmax(σ(W2·φ(W1·presig+b1)+b2)) Among them, C represents the output of the loss classification module, presig represents the reception of the pilot signal, W1 and W2 represent weight matrices, b1 and b2 represent biases, φ( ) represents the nonlinear activation function, σ( ) represents the softmax activation function, and argmax( ) represents the selection of the category corresponding to the maximum probability.

4. The method for adaptively compensating complex signals using an extraction and fusion strategy according to claim 1, characterized in that: Each equalizing network in the signal equalizing module is constructed based on a convolutional neural network, and the hyperparameters of each equalizing network are set differently, including the number of hidden layers, the convolution void ratio, and the convolution kernel dimension.

5. The method for adaptively compensating complex signals using an extraction and fusion strategy according to claim 1, characterized in that: Step S3 uses an equalization network corresponding to the damage type to perform compensation processing, including: S31. For the real part signal Rx in the normal transmission complex signal RE and the imaginary signal Rx IM , the extraction sub-network is used to extract the real part loss information Feature RE and imaginary damage information Feature IM , expressed as Feature RE =φ1(W extract ·Rx RE +b) Feature IM =φ1(W extract ·Rx IM +b) Among them, W extract represents the weight matrix, b represents the convolution bias, and φ1( ) represents the nonlinear activation function; S32. Feature the real part loss information RE and imaginary damage information Feature IM Splicing to get features S33.Characteristics Input the fusion sub-network to obtain the comprehensive damage information Feature, which is expressed as Among them, W fuse 1. W fuse 2 represents the weight matrix, b3 and b4 represent the bias, and φ2( ) represents the nonlinear activation function.

6. The method for adaptively compensating complex signals using an extraction and fusion strategy according to claim 1, characterized in that: The fusion sub-network uses a high receptive field convolution kernel, where dilation rate = kernel size padding=(kernel size-1)×dilation÷2 Kernel size indicates the size of the convolution kernel, dilation rate indicates the hole value, and padding indicates the edge padding value.

7. The method for adaptively compensating complex signals using an extraction and fusion strategy according to claim 1, characterized in that: In step S4, in the signal output module, a compensation signal is obtained by subtracting the comprehensive damage information from the normal transmission complex signal.

Images