A lightweight design method for neural network for AI-enabled timing synchronization

Abstract

The present invention discloses a neural network lightweight design method for AI-enabled timing synchronization, the design method comprising: designing the initial layer receptive field parameters of the convolution module according to the length of the timing synchronization training sequence, designing a lightweight convergence layer according to the protection interval length of the timing synchronization training sequence and the initial layer receptive field parameters, designing a lightweight refining layer according to the protection interval length of the timing synchronization training sequence and the convergence layer receptive field parameters, achieving network lightweight while achieving accurate timing measurement, and finally, using a fully connected module to output a timing synchronization offset estimate. The present invention can improve the timing synchronization performance in wireless networks and mobile communication scenarios with low processing delay requirements, and compared with the existing timing synchronization method based on compressed sensing, it has lower complexity, higher timing synchronization accuracy and lower processing delay.

Description

Technical Field

[0001] The present invention relates to the technical field of AI timing synchronization, and in particular to a neural network lightweight design method and device for AI-enabled timing synchronization. Background Art

[0002] AI-enabled timing synchronization is a key technology and research hotspot for future wireless networks and mobile communication technologies. The quality of its synchronization performance has a direct impact on subsequent signal processing decisions. In modern wireless and mobile communication systems, traditional signal processing methods usually require the sacrifice of the computational complexity of the processing method to improve the accuracy of the method. Unfortunately, however, processing methods that sacrifice complexity are not suitable for some wireless and mobile communication scenarios that have strict requirements on processing delay. In addition, more stringent requirements are put forward for timing synchronization with lower processing delay in wireless communication scenarios. The timing synchronization method based on compressed sensing can perform interference cancellation through multiple iterations, thereby improving the timing synchronization performance of the system. However, compressed sensing requires a series of problems and challenges such as a complex iterative process, resulting in ultra-high processing delays, which urgently need to be improved. Summary of the invention

[0003] The purpose of the present invention is to provide a neural network lightweight design method for AI-enabled timing synchronization. Compared with the existing timing synchronization method based on compressed sensing, this method designs a convolutional neural network to achieve neural network lightweight, reduces the computational complexity and processing delay of timing synchronization, and achieves high-precision AI-enabled timing synchronization performance. It not only effectively improves the synchronization performance of AI-enabled timing synchronization, but also reduces the difficulty of deploying the timing synchronization device.

[0004] A lightweight design method for a neural network with timing synchronization for AI empowerment, the method comprising:

[0005] S1 constructs a convolutional neural network including sequentially connected convolutional layer modules and fully connected layer modules, wherein the convolutional layer module includes an initial layer, a convergence layer and a refinement layer that are sequentially spliced, wherein the initial layer includes a first parameter recognition unit, a first parameter initialization unit and a first convolutional layer, the convergence layer includes a second parameter initialization unit and a second convolutional layer, the refinement layer includes a third parameter initialization unit and a third convolutional layer, and each convolutional layer is connected in a cascade manner; the fully connected layer module includes a τ max +1 fully connected layer with output neuron nodes and an output layer with softmax activation function, where τ max It indicates the maximum possible propagation delay obtained through actual measurement;

[0006] S2 designs the receptive field parameters of the initial layer according to the length N and the tuning parameter k of the timing synchronization training sequence, and the number K of input features of the received and stacked stored signal y corresponding to the timing synchronization training sequence obtained by the first parameter identification unit;

[0007] S3 designs the receptive field parameter of the convergence layer according to the timing synchronization protection interval G set for the synchronization training sequence, the tuning parameter k, the number of input features K obtained by the first parameter identification unit of the initial layer, and the receptive field parameter of the initial layer;

[0008] S4, designing the receptive field parameter of the refining layer according to the timing synchronization protection interval G, the tuning parameter k, and the receptive field parameter of the convergence layer;

[0009] S5 determines the convolution layer module according to the receptive field parameter of the initial layer, the receptive field parameter of the convergence layer, and the receptive field parameter of the refinement layer, and the obtained convolution neural network is a lightweight convolution neural network that can realize AI-enabled timing synchronization;

[0010] The receptive field parameters include the convolution kernel size and the number of convolution kernels.

[0011] Among them, the specific structures of the first parameter identification unit, the first, second and third parameter initialization units can be selected from the existing convolutional neural network structure, such as the first parameter identification unit is a structure for determining the tuning parameter k according to the existing feature engineering (such as reference document "Wang J, Jiang C, Zhang H, et al. Thirty years of machine learning: The road to Pareto-optimal wireless networks [J]. IEEE Communications Surveys & Tutorials, 2020, 22 (3): 1472-1514."); the first, second and third parameter initialization units are implemented according to the neural network structure for initializing the convolutional layer weights, biases and activation functions in the prior art (such as reference document "Wei Xiucan. Analysis of Deep Learning: Principles and Visual Practice of Convolutional Neural Networks. Publishing House of Electronics Industry, 2018.).

[0012] According to some preferred embodiments of the present invention, the receptive field parameters of the initial layer are obtained by the following calculation model:

[0013] Q=(2+k)N

[0014]

[0015]

[0016] Where Q represents the convolution kernel size of the initial layer, that is, the length of the receptive field; k represents the tuning parameter, whose value range is an integer between {-1,0}; K represents the number of input features of the received stack signal y with a length of M; C α Indicates the number of convolution kernels in the initial layer. The length of the timing synchronization sequence is known to both the sender and the receiver and is set based on engineering experience.

[0017] According to some preferred embodiments of the present invention, the receptive field parameter of the convergence layer is obtained by the following calculation model:

[0018]

[0019] C β =max(C α ,(2+k)K)

[0020] Among them, P represents the convolution kernel size of the pooling layer; Indicates rounding up; C β Indicates the number of convolution kernels in the pooling layer.

[0021] According to some preferred embodiments of the present invention, the convolution kernel parameters of the refining layer are obtained by the following calculation model:

[0022]

[0023]

[0024] Among them, L represents the convolution kernel size of the refinement layer, C γ Represents the number of convolution kernels in the refinement layer

[0025] Based on the above method, the present invention can further obtain a method for the lightweight convolutional neural network to perform AI empowerment timing synchronization, which includes:

[0026] Input the signal sequence x received online and stacked for processing to the initial layer, identify the number of features of the received stack signal y of the timing synchronization training sequence based on the first parameter identification unit of the initial layer of the convolution module, set the tuning parameter k and the synchronization training sequence protection interval G, and identify and obtain the initial extraction feature number K, wherein the input feature number identification includes: signal sequence real part input feature number identification, imaginary part input feature number identification, power input feature number identification, amplitude input feature number extraction and phase input feature number extraction;

[0027] According to the obtained initial extraction feature number K and the timing synchronization training sequence length N, the size Q and number C of the convolution kernel of the initial layer are generated by the first parameter initialization unit of the initial layer of the convolution module. α , and initialize the weights, biases, and activation functions for the first convolutional layer;

[0028] Extracting features of the signal sequence x received online and stacked by completing the initialization of the first convolutional layer, obtaining the timing metric features of the signal sequence, and forming a complete timing metric initial feature from the timing metric features output by each channel;

[0029] According to the set guard interval G and tuning parameter k, the number of initial recognition features K and the number of initial layer convolution kernels C α , the size P and number C of the convolution kernel of the convolution layer are generated by the second parameter initialization unit of the convolution layer β , and initialize the weights, biases, and activation functions for the second convolutional layer;

[0030] Further feature extraction is performed on the initial features of the timing measurement by completing the initialization of the second convolutional layer to obtain a significant feature of the pooling layer, and the significant features of the pooling layer output by each channel form a significant feature of the timing measurement;

[0031] According to the set guard interval G and tuning parameter k, the number of convolution kernels C in the pooling layer is obtained β , the size L of the convolution kernel and the number C of the convolution kernel of the third convolution layer of the convolution layer are generated by the third parameter initialization unit of the refining layer γ , and initialize the weights, biases, and activation functions for the third convolutional layer;

[0032] Extracting the timing metric salient features through the initialized third convolutional layer to obtain refined timing metric features, and forming timing metric refined features from the refined timing metric features output by each channel;

[0033] Input the timing metric refinement feature into the fully connected layer module to obtain the output timing synchronization offset estimate

[0034] According to some preferred embodiments of the present invention, the timing synchronization offset estimate It is obtained from the following formula:

[0035]

[0036] Among them, p j It is represented as the output of the j-th output node of the fully connected layer module after the softmax activation function.

[0037] The present invention has the following beneficial effects:

[0038] The present invention forms a neural network lightweight design method for AI-enabled timing synchronization based on the length of the timing synchronization training sequence and its protection interval, thereby reducing the computational complexity and processing delay of the timing synchronization device. The method specifically includes first designing the receptive field parameters of the initial layer for capturing the initial characteristics of the timing measurement based on the length of the timing synchronization training sequence; then designing a lightweight convergence layer and a refinement layer based on the protection interval length of the training sequence, optimizing the timing measurement characteristics while realizing the lightweight design of the network; finally, utilizing a fully connected layer module using a softmax activation function to perform timing classification and output a timing synchronization offset estimate.

[0039] For wireless network and mobile communication scenarios with low processing delay requirements, the present invention can construct a lightweight AI-enabled timing synchronization device according to the length of the timing synchronization sequence, forming a low-complexity timing synchronization method and an easy-to-deploy device, thereby improving the synchronization performance in actual scenarios while reducing processing delays, and bringing many implementation plans for timing synchronization with low-latency processing requirements in actual scenarios, which is of great significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the process of the present invention.

[0041] Figure 2 This is the design diagram of the initial layer of the convolution module of the present invention.

[0042] Figure 3 This is the first parameter identification unit process of the initial layer of the convolution module of the present invention.

[0043] Figure 4 This is the design diagram of the convolution module aggregation layer of the present invention.

[0044] Figure 5 This is a design diagram of the convolution module refinement layer of the present invention.

[0045] Figure 6 Schematic diagram of an AI-enabled timing synchronization device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The present invention is described in detail below in conjunction with the embodiments and drawings, but it should be understood that the embodiments and drawings are only used to exemplify the present invention and do not constitute any limitation on the protection scope of the present invention. All reasonable changes and combinations within the scope of the inventive concept of the present invention fall within the protection scope of the present invention.

[0047] According to the technical solution of the present invention, a specific implementation method is as follows: Figure 1 As shown, the following steps are included:

[0048] S1 constructs a convolutional neural network including sequentially connected convolutional layer modules and fully connected layer modules, wherein the convolutional layer module includes an initial layer, a convergence layer and a refinement layer that are sequentially spliced, wherein the initial layer includes a parameter recognition unit, a first parameter initialization unit and a first convolutional layer, the convergence layer includes a second parameter initialization unit and a second convolutional layer, the refinement layer includes a third parameter initialization unit and a third convolutional layer, and each convolutional layer is connected in a cascade manner; the fully connected layer module includes a τ max +1 fully connected layer with output neuron nodes and an output layer with sofmax activation function, where τ max It indicates the maximum possible propagation delay obtained through actual measurement;

[0049] S2 designs the receptive field parameters of the initial layer according to the length N and the tuning parameter k of the timing synchronization training sequence, and the number K of input features of the received and stacked stored signal y corresponding to the timing synchronization training sequence obtained by the first parameter identification unit;

[0050] S3 designs the receptive field parameter of the convergence layer according to the timing synchronization protection interval G set for the synchronization training sequence, the tuning parameter k, the number of input features K obtained by the first parameter identification unit of the initial layer, and the receptive field parameter of the initial layer;

[0051] S4, designing the receptive field parameter of the refining layer according to the timing synchronization protection interval G, the tuning parameter k, and the receptive field parameter of the convergence layer;

[0052] S5 determines the convolution layer module according to the receptive field parameter of the initial layer, the receptive field parameter of the convergence layer, and the receptive field parameter of the refinement layer, and the obtained convolution neural network is a lightweight convolution neural network that can realize AI-enabled timing synchronization;

[0053] The receptive field parameters of each convolution layer include the convolution kernel size and the number of convolution kernels;

[0054] The fully connected module uses a softmax activation function to complete the timing synchronization offset value estimation;

[0055] The design method of the initial layer receptive field of the convolution module described in step S2 can be further described in the attached Figure 2 As shown, specifically including:

[0056] According to the length N of the timing synchronization training sequence, the initial layer convolution kernel size Q is generated in the following way, namely:

[0057] Q=(2+k)N

[0058] The training sequence length N is based on the agreement between the transmitter and receiver, k is a tuning parameter, and its value range is {-1,0}. The specific value of k is set according to engineering experience;

[0059] The received sampling signal is stacked and stored to form a received stack signal y with a length of M. The received stack signal y is used to identify the number of input features. The number of input features K is calculated by the following formula:

[0060]

[0061] The input feature number recognition method includes: real part input feature number recognition, imaginary part input feature number recognition, power input feature number recognition, amplitude input feature number extraction and phase input feature number extraction completed before the signal sequence is sent to the convolution module. For example, when the tuning parameter k=0, the number of input features obtained is K=2;

[0062] According to the number of input features K identified, the number of convolution kernels in the initial layer C α It is obtained from the following formula:

[0063]

[0064] The number of convolution kernels is also called the number of channels of the convolutional neural network;

[0065] The design method of the convolution kernel of the convolution module of step S3 can be further described in the attached figure. Figure 4 As shown, specifically including:

[0066] According to the length G of the guard interval set for the timing synchronization training sequence, the convolution kernel size P of the convergence layer is calculated by the following formula:

[0067]

[0068] According to the number of convolution kernels C in the initial layer α , tuning parameter k and number of input features K, number of convolution kernels in the pooling layer C β It is obtained from the following formula:

[0069] C β =max(C α ,(2+k)K)

[0070] Among them, the length of the timing synchronization protection interval can be set according to the field equipment measurement;

[0071] The relationship between the convolution kernel sizes of the initial layer and the pooling layer satisfies And the number of convolution kernels in the initial layer and the pooling layer satisfies the relationship C α ≤C β, which reduces the computational complexity of convolution layer convolution and thus makes the network lightweight;

[0072] The design method of the convolution kernel of the refinement layer of the convolution module described in step S4 can be further described in the attached Figure 5 As shown, specifically including:

[0073] According to the length G of the guard interval set for the timing synchronization training sequence, the size L of the convolution kernel of the refinement layer is calculated by the following formula:

[0074]

[0075] According to the number of convolution kernels C in the pooling layer β , determine the number of convolution kernels C in the refinement layer γ for:

[0076]

[0077] The number of convolution kernels in the pooling layer and the refining layer satisfies the relationship Reduce the computational complexity of convolutional layers and further reduce the weight of the network.

[0078] The process of AI-enabled timing synchronization through the above lightweight neural network can be further described as shown in the attached figure. Figure 6 As shown, specifically including:

[0079] Input the signal sequence x received online and stacked for processing to the initial layer, identify the number of features of the received stack signal y of the timing synchronization training sequence through the parameter identification unit of the initial layer of the convolution module, set the tuning parameter k and the synchronization training sequence protection interval G, and identify and obtain the initial extraction feature number K, wherein the input feature number identification includes: signal sequence real part input feature number identification, imaginary part input feature number identification, power input feature number identification, amplitude input feature number extraction and phase input feature number extraction;

[0080] According to the obtained initial extraction feature number K and the timing synchronization training sequence length N, the size Q and number C of the convolution kernel of the initial layer are generated by the first parameter initialization unit of the initial layer of the convolution module. α , and initialize the weights, biases, and activation functions for the first convolutional layer;

[0081] Extracting features of the signal sequence x received online and stacked by completing the initialization of the first convolutional layer, obtaining the timing metric features of the signal sequence, and forming a complete timing metric initial feature from the timing metric features output by each channel;

[0082] According to the set guard interval G and tuning parameter k, the number of initial recognition features K and the number of initial layer convolution kernels C α , the size P and number C of the convolution kernel of the convolution layer are generated by the second parameter initialization unit of the convolution layer β , and initialize the weights, biases, and activation functions for the second convolutional layer;

[0083] Further feature extraction is performed on the initial features of the timing measurement by completing the initialization of the second convolutional layer to obtain a significant feature of the pooling layer, and the significant features of the pooling layer output by each channel form a significant feature of the timing measurement;

[0084] According to the set guard interval G and tuning parameter k, the number of convolution kernels C in the pooling layer is obtained β , the size L of the convolution kernel and the number C of the convolution kernel of the third convolution layer of the convolution layer are generated by the third parameter initialization unit of the refining layer γ , and initialize the weights, biases, and activation functions for the third convolutional layer;

[0085] Extracting the timing metric salient features through the initialized third convolutional layer to obtain refined timing metric features, and forming timing metric refined features from the refined timing metric features output by each channel;

[0086] The timing metric refined features are input into the fully connected layer module, and the timing synchronization offset estimation value is obtained through the fully connected module using the softmax activation function have:

[0087]

[0088] Among them, p j It is represented as the output of the jth output node of the fully connected layer module after the timing metric refined features output by the convolution module's aggregation layer are activated using the softmax function.

[0089] In the above specific implementation manner, the present invention further provides the following embodiments:

[0090] Example 1

[0091] As attached Figure 2 The initial layer design of the convolution module described in the above example assumes that the length of the timing synchronization training sequence is N=256, k=0, and the initial layer convolution kernel size is Q=512 according to Q=(2+k)N;

[0092] The received sampled signal is stacked and stored to form a received stack signal y with a length of M = 512. The initial feature extraction is performed using y to obtain the number of features.

[0093] Assume that the initial feature extraction uses real feature extraction and imaginary feature extraction;

[0094] Using the number K = 2, according to the formula The number of convolution kernels in the initial layer is C α =4.

[0095] At this point, the convolution layer receptive field parameters of the initial layer of the convolution module obtained in the embodiment 1 are: convolution kernel size Q=512, convolution kernel number C α =4;

[0096] Example 2

[0097] As attached Figure 3 The convolution module convergence layer design described in the above assumes that the timing synchronization protection interval length is G = 32, k = 0, according to The convolution kernel size of the pooling layer is obtained to be P = 16;

[0098] According to the tuning parameter k=0 in Example 1, the number of convolution kernels in the initial layer C α =4 and the initial number of features K = 2, using C β =max(C α ,(2+k)K)The number of convolution kernels in the pooling layer is C β =4;

[0099] At this point, the convolution layer receptive field parameters of the convolution module convergence layer obtained in the embodiment 2 are: convolution kernel size P = 16, number of convolution kernels C β =4.

[0100] Example 3

[0101] As attached Figure 5 The AI-enabled timing synchronization device described in the embodiment outputs a tuning parameter k and the number of input features K according to the offline received stack signal y using the parameter recognition unit of the initial layer of the convolution module, such as k=0 and K=2;

[0102] According to the timing synchronization training sequence length N and the guard interval length G in y, such as N=128, G=32, combined with the tuning parameter k and the number of input features K, the initial layer receptive field parameters are obtained according to the following calculation model:

[0103] Q=(2+k)N

[0104]

[0105] The convolution kernel size is Q=256, and the number of convolution kernels is C. α =4, and use the attached Figure 5The parameter initialization unit in the initial layer of the convolution module gives the convolution layer weights, biases, and activation functions;

[0106] Similarly, according to the timing synchronization sequence protection interval length G and the initial layer convolution kernel size C α , input the number of features K and the tuning parameter k, such as G = 32, C α =4, K=2 and k=0, the receptive field parameters of the pooling layer are calculated as follows:

[0107]

[0108] C β =max(C α ,(2+k)K)

[0109] The convolution kernel size is P=16, and the number of convolution kernels is C β =4, and use the attached Figure 5 The parameter initialization unit in the convolutional module pooling layer gives the convolutional layer weights, biases, and activation functions;

[0110] Similarly, according to the timing synchronization sequence protection interval length G and the convolution kernel size C of the convergence layer β , and tuning parameters k, such as G = 32, C β =4 and k=0, the receptive field parameters of the refinement layer are calculated as follows:

[0111]

[0112]

[0113] The convolution kernel size is L=16, and the number of convolution kernels is C. γ =2, and use the attached Figure 5 The parameter initialization unit in the refinement layer of the convolution module gives the convolution layer weights, biases, and activation functions.

[0114] Example 4

[0115] In one operating environment, namely: the processor is 11th Gen Intel(R) Core(TM) i5-11300H@3.1GHz3.11GHz, the memory is 16GB, and the 64-bit Windows11 operating system;

[0116] According to the length M of the received stack signal y, the length N of the timing synchronization training sequence, the timing synchronization training sequence protection interval G, the number of input features K, and the convolution receptive field parameters Q, P, L, C given in the above embodiment 3, α ,C β ,C γ ,like,

[0117] M=288, N=128, G=32, K=2, Q=256, P=16, L=16, C α =4,C β =4,C γ = 2 and assuming a maximum propagation delay τ max =127, the specific architecture of the convolution module and the fully connected module of the preferred embodiment is given as shown in the following table:

[0118] name Output Dimensions Convolution kernel size Number of convolution kernels Activation Function enter KM=576 none none none The first convolutional layer <![CDATA[(N+G)×C α =160×4]]> Q＝256 <![CDATA[C α =4]]> ReLU The second convolutional layer <![CDATA[(N+G2)×C β =144×4]]> P＝16 <![CDATA[C β =4]]> ReLU The third convolutional layer <![CDATA[N×C γ =128×2]]> L＝16 <![CDATA[C γ =4]]> ReLU Fully connected modules <![CDATA[τ max +1=128]]> none none softmax

[0119] According to the network architecture and the received stack signal y of length M, the equivalent complex multiplication complexity of the designed lightweight network is calculated in the following way:

[0120]

[0121] Wherein, according to the preferred implementation, the network processing delay is 0.35 seconds;

[0122] According to the M, N, and assuming that the number of wireless communication propagation paths is P=28, in a comparative literature method based on compressed sensing, the complexity of the complex multiplication used is:

[0123]

[0124] Among them, the processing delay of the comparative document is 81.41 seconds.

[0125] According to the computational complexity and processing delay of the above-designed neural network and the comparative literature, the present invention achieves network lightweighting.

[0126] Example 5

[0127] As attached Figure 5 The AI-enabled timing synchronization device, according to the online received stack signal, obtains the timing metric initial feature, the timing metric significant feature and the timing metric refined feature through the initialized convolution module initial layer, the convergence layer and the refinement layer respectively; according to the timing metric refined feature output by the convolution module, outputs the timing synchronization offset estimation value through the fully connected module;

[0128] The fully connected module uses the softmax activation function to activate the timing metric refinement feature; assuming that the actual measurement obtains the maximum possible propagation delay τ max is τ max =128, the timing synchronization offset estimate is obtained in the following manner, namely:

[0129]

[0130] Among them, p j Represents the output of the j-th output node of the fully connected layer module after the timing metric refined features of the convolution module convergence layer output are activated by the softmax function.

[0131] The above embodiments are only preferred implementations of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A lightweight design method for AI-enabled timing synchronization neural networks, It is characterized in that For wireless network and mobile communication scenarios with low processing latency requirements, including: S1 constructs a convolutional neural network including sequentially connected convolutional layer modules and fully connected layer modules, wherein the convolutional layer module includes an initial layer, a convergence layer and a refinement layer that are sequentially spliced, wherein the initial layer includes a first parameter recognition unit, a first parameter initialization unit and a first convolutional layer, the convergence layer includes a second parameter initialization unit and a second convolutional layer, the refinement layer includes a third parameter initialization unit and a third convolutional layer, and each convolutional layer is connected in a cascade manner; the fully connected layer module includes a τ max +1 fully connected layer with output neuron nodes and an output layer with softmax activation function, where τ max It indicates the maximum possible propagation delay obtained through actual measurement; S2, designing the receptive field parameter of the initial layer according to the length N and the tuning parameter k of the timing synchronization training sequence, and the number K of input features of the received and stacked stored signal y corresponding to the timing synchronization training sequence obtained by the first parameter identification unit; S3 designs the receptive field parameter of the convergence layer according to the timing synchronization protection interval G set for the synchronization training sequence, the tuning parameter k, the number of input features K obtained by the first parameter identification unit of the initial layer, and the receptive field parameter of the initial layer; S4, designing the receptive field parameter of the refining layer according to the timing synchronization protection interval G, the tuning parameter k, and the receptive field parameter of the convergence layer; S5 determines the convolution layer module according to the receptive field parameter of the initial layer, the receptive field parameter of the convergence layer, and the receptive field parameter of the refinement layer, and the obtained convolution neural network is a lightweight convolution neural network that can realize AI-enabled timing synchronization; The receptive field parameters include the convolution kernel size and the number of convolution kernels.

2. The method according to claim 1, It is characterized in that The receptive field parameters of the initial layer are obtained by the following calculation model: Q=(2+k)N Wherein, Q represents the convolution kernel size of the first convolution layer; k represents a tuning parameter, whose value range is an integer between {-1,0}; K represents the number of initial extracted features of the received stack signal y with a length of M; C α Represents the number of convolution kernels in the initial layer. The length N is known to both the sender and the receiver and is set based on engineering experience.

3. The design method according to claim 2, It is characterized in that The receptive field parameters of the pooling layer are obtained through the following calculation model: C β =max(C α ,(2+k)K) Wherein, P represents the convolution kernel size of the second convolution layer of the pooling layer; Indicates rounding up; C β Indicates the number of convolution kernels of the second convolution layer.

4. The design method according to claim 3, It is characterized in that The receptive field parameters of the refinement layer are obtained by the following calculation model: Where L represents the convolution kernel size of the third convolutional layer of the refinement layer, C γ Represents the number of convolution kernels of the third convolution layer.

5. A method for AI-enabled timing synchronization using the lightweight convolutional neural network obtained by the design method described in any one of claims 1 to 4, wherein include: Input the baseband signal sequence x collected by the wireless communication receiver for online reception and stack processing to the initial layer, identify the feature number of the received stack signal y of the timing synchronization training sequence through the parameter identification unit of the initial layer of the convolution module, set the tuning parameter k and the synchronization training sequence protection interval G, and identify and obtain the initial extraction feature number K, wherein the input feature number identification includes: signal sequence real part input feature number identification, imaginary part input feature number identification, power input feature number identification, amplitude input feature number extraction and phase input feature number extraction; According to the obtained initial extraction feature number K and the timing synchronization training sequence length N, the size Q and number C of the convolution kernel of the initial layer are generated by the first parameter initialization unit of the initial layer of the convolution module. α , and initialize the weights, biases, and activation functions for the first convolutional layer; Extracting features of the signal sequence x received online and stacked by completing the initialization of the first convolutional layer, obtaining the timing metric features of the signal sequence, and forming a complete timing metric initial feature from the timing metric features output by each channel; According to the set guard interval G and tuning parameter k, the number of initial recognition features K and the number of initial layer convolution kernels C α , the size P and number C of the convolution kernel of the convolution layer are generated by the second parameter initialization unit of the convolution layer β , and initialize the weights, biases, and activation functions for the second convolutional layer; Further feature extraction is performed on the initial features of the timing measurement by completing the initialization of the second convolutional layer to obtain a significant feature of the pooling layer, and the significant features of the pooling layer output by each channel form a significant feature of the timing measurement; According to the set guard interval G and tuning parameter k, the number of convolution kernels C in the pooling layer is obtained β , the size L of the convolution kernel and the number C of the convolution kernel of the third convolution layer of the convolution layer are generated by the third parameter initialization unit of the refining layer γ , and initialize the weights, biases, and activation functions for the third convolutional layer; Extracting the timing metric salient features through the initialized third convolutional layer to obtain refined timing metric features, and forming timing metric refined features from the refined timing metric features output by each channel; Input the timing metric refinement feature into the fully connected layer module to obtain the output timing synchronization offset estimate 6. The method for AI-enabled timing synchronization according to claim 5, It is characterized in that The timing synchronization offset estimate It is obtained by the following formula: Among them, p j It is represented as the output of the jth output node of the fully connected layer module after the timing metric refined features of the output of the pooling layer of the convolution module are activated by the softmax function.

Images