A method and system for intelligent modulation format identification of photon-assisted millimeter wave OTFS

Abstract

This invention discloses a photon-assisted intelligent modulation format recognition method and system for millimeter-wave OTFS. At the transmitting end, modulation symbols corresponding to different modulation formats are mapped to a two-dimensional data matrix in the time-delay-Doppler domain of the OTFS frame to generate a time-domain OTFS signal. The time-domain OTFS signal is converted into a millimeter-wave signal using photon assistance and transmitted into free space. At the receiving end, the time-delay dimension subcarrier signals corresponding to preset indices in the OTFS time-delay-Doppler domain signal are extracted and separated into real and imaginary data. These are then concatenated in the channel dimension to form real / imaginary data (I / Q data), and constellation diagram data is generated based on the I / Q data. The I / Q data and constellation diagram data are used to construct an input dataset. A pre-constructed intelligent modulation format recognition neural network model is trained based on the training and validation datasets. A dual-channel input dataset is constructed based on the signal to be modulated and classified, and input into the trained intelligent modulation format recognition neural network model to obtain the modulation classification result. This invention achieves high-precision intelligent recognition of multiple modulation formats in millimeter-wave OTFS systems and is suitable for 6G high-speed mobile communication and optical-wireless converged communication.

Description

Technical Field

[0001] This invention relates to the field of communication technology, specifically to a method and system for intelligent modulation format identification of photon-assisted millimeter-wave OTFS, applicable to 6G high-speed mobile communication and optical-wireless converged communication scenarios. Background Technology

[0002] Orthogonal Time-Frequency Modulation (OTFS), a novel modulation technique adapted to high-speed mobile and high-frequency millimeter-wave scenarios, effectively addresses the time-varying nature of channels in high Doppler shift environments by multiplexing signals in the delay-Doppler (DD) domain, and has become one of the core candidate modulation schemes for 6G systems. With the increasing complexity of communication scenarios and the diversification of service requirements, the modulation formats of OTFS systems are showing a diversified development trend, and traditional signal processing mechanisms relying on fixed modulation formats are no longer sufficient to meet the needs of dynamic adaptive communication.

[0003] Modulation Format Recognition (MFR), a key technology in intelligent communication and cognitive radio, can automatically identify the modulation type of received signals under unknown channel parameters and complex interference environments, providing core support for dynamic spectrum allocation, anti-interference communication, and flexible networking. In recent years, deep learning has made significant progress in the field of modulation format recognition due to its powerful feature extraction and pattern classification capabilities. Compared with traditional methods based on likelihood ratio or feature engineering, it can achieve end-to-end blind recognition, demonstrating significant advantages in recognition accuracy and robustness.

[0004] As 6G communication evolves to higher frequency bands, such as millimeter waves and terahertz bands, photonic-assisted millimeter wave technology has become a key enabling technology for achieving high-capacity, long-distance OTFS signal transmission due to its advantages such as low loss, large bandwidth, and resistance to electromagnetic interference. However, existing research on OTFS modulation format identification mainly focuses on wireless mobile communication scenarios. For photonic-assisted millimeter wave OTFS systems, the nonlinear effects introduced by the photonic-assisted link, the multipath fading of the millimeter wave channel, and the coupling of the OTFS's unique time-delay-Doppler domain signal structure make it difficult to directly apply traditional modulation format identification methods based on the time or frequency domain. Summary of the Invention

[0005] To address the aforementioned issues, this invention discloses an intelligent modulation format identification method and system for photon-assisted millimeter-wave OTFS. By constructing a parallel dual-branch feature extraction network that integrates delay-Doppler domain I / Q data and constellation diagram data, high-precision intelligent identification of multiple modulation formats in the photon-assisted millimeter-wave OTFS system is achieved.

[0006] A method for intelligent modulation format identification of photon-assisted millimeter-wave OTFS, characterized in that the method includes the following steps:

[0007] Step 1: At the transmitting end, the modulation symbols corresponding to different modulation formats are mapped to the time-delay-Doppler domain two-dimensional data matrix of the OTFS frame, and the inverse symplectic finite Fourier transform and Heisenberg transform are performed in sequence to generate the time-domain OTFS signal.

[0008] Step 2: The time-domain OTFS signal is converted into a millimeter-wave signal using a photon-assisted method and transmitted into free space;

[0009] Step 3: At the receiving end, perform Wigner transform and symplectic finite Fourier transform on the received signal in sequence to obtain the OTFS time-delay-Doppler domain signal; extract the time-delay dimension subcarrier signal corresponding to the preset index in the OTFS time-delay-Doppler domain signal, separate the extracted subcarrier data into real part data and imaginary part data, and concatenate them in the channel dimension to form real part / imaginary part data (I / Q data), and generate constellation diagram data based on the I / Q data; construct the I / Q data and constellation diagram data together to form an input dataset containing two types of data, and divide it into training dataset, validation dataset and test dataset according to the proportion;

[0010] Step 4: Train the pre-built intelligent modulation format recognition neural network model based on the training dataset and validation dataset;

[0011] Step 5: Obtain the signal to be modulated and classified, construct a dual-channel input dataset based on the signal to be modulated and classified, and input it into the trained intelligent modulation format recognition neural network model to obtain the modulation classification result.

[0012] Further, in step 1, mapping the modulation symbols corresponding to different modulation formats to the two-dimensional data matrix of the time-delay-Doppler domain of the OTFS frame specifically includes: setting the modulation format index as t, where t corresponds to different modulation formats in sequence; within the same OTFS frame, different modulation formats are mapped to different subcarrier ranges in the time-delay-Doppler domain, and the subcarrier ranges corresponding to different modulation formats are continuous and do not overlap, satisfying:

[0013]

[0014] Where N is the number of subcarrier indices mapped for each modulation format. , Let k be the start and end subcarrier indices of the t-th modulation format, respectively; the modulation format category to which any subcarrier index k belongs satisfies:

[0015]

[0016] in, This is a rounding down operation; under the allocation method, the mapped subcarrier range of all modulation formats exactly covers all time delay dimension subcarriers in the time delay-Doppler domain, and the subcarrier intervals corresponding to any two types of modulation formats do not overlap, so as to realize independent identification and non-aliasing transmission of data of different modulation formats in the time delay-Doppler domain subcarrier dimension.

[0017] Further, in step 3, the I / Q data and constellation diagram data are jointly constructed into an input dataset containing two types of data. Specifically, this includes: performing virtual-real separation and channel-dimension concatenation on a row of 1×M data corresponding to each subcarrier received and transformed into the delay-Doppler domain. For each delay-dimension subcarrier, an I / Q data matrix of dimension 2×M is obtained, represented as:

[0018]

[0019] in, This represents the I / Q data matrix of the k-th subcarrier. This represents the complex vector data of the k-th subcarrier. This indicates the operation of taking the real part. This indicates the operation of taking the imaginary part, where M is the Doppler dimension.

[0020] The I / Q data matrix is ​​subjected to amplitude limiting, which is expressed as:

[0021]

[0022] in, This is the data matrix after amplitude limiting. and These are the maximum and minimum values ​​within the preset amplitude limit range, respectively;

[0023] The clipped I / Q signal is linearly mapped to pixel space, as follows:

[0024]

[0025] Where x and y are the pixel coordinates after mapping the real and imaginary parts of the signal, respectively. Let x,y ∈ [0,y] be the coordinate axis dimensions of the constellation diagram, and x,y ∈ [0,y]. A constellation diagram is drawn with a black background and white data pixels, represented as follows:

[0026]

[0027] in, This represents the pixel value matrix of the constellation map corresponding to the k-th subcarrier at coordinates (x, y); the constellation map is normalized to obtain the final constellation map data input to the network.

[0028]

[0029] in, The normalized constellation map pixel matrix has a dimension of 1×P×P; the generated I / Q data and constellation map data are combined to construct an input dataset containing two types of data.

[0030] Furthermore, the intelligent modulation format recognition neural network model pre-constructed in step 4 specifically includes: the intelligent modulation format recognition neural network model comprises a first branch, a second branch, and a feature fusion and classification module, wherein:

[0031] The first branch employs a structure combining a residual network and a long short-term memory network to process the constructed I / Q data. The input to the first branch is the I / Q data matrix. The I / Q data matrix passes sequentially through a network structure composed of multiple stacked residual modules, each residual module including a convolutional layer, a Selu activation function layer, a batch normalization layer, and a dropout layer. The data output from the residual network is then input into a two-layer long short-term memory network (LSTM). After flattening the data output from the long short-term memory network, it sequentially passes through a fully connected layer, a batch normalization layer, and a dropout layer to output the first feature vector.

[0032] The second branch employs a two-dimensional convolutional neural network structure to process the constellation diagram data constructed above. The input to the second branch is the constellation diagram data matrix. The constellation diagram data matrix is ​​sequentially processed through a network structure consisting of three stacked convolutional modules. Each convolutional module includes a convolutional layer, a batch normalization layer, a ReLU activation function layer, and a pooling layer. After flattening the output data, it is sequentially processed through a fully connected layer, a batch normalization layer, and a dropout layer to output the second feature vector.

[0033] Feature fusion and classification module: The first feature vector and the second feature vector are concatenated along the channel dimension to obtain a fused feature vector; the fused feature vector is then passed through a combination structure containing two fully connected layers, a ReLU activation function layer and a batch normalization layer, followed by a classification header of a Softmax layer to output the modulation format classification result. The dimension of the modulation classification result is the same as the number of modulation categories.

[0034] Furthermore, in the first branch: the residual network is used to suppress the impact of gradient explosion on network training, and extracts the spatial features of the time-delay-Doppler domain I / Q data and reduces the feature map size through convolutional layers; the long short-term memory network is used to extract the temporal features of the time-delay-Doppler domain I / Q data.

[0035] Furthermore, the training process of the intelligent modulation format recognition neural network model includes: inputting the training dataset into the pre-built intelligent modulation format recognition neural network model to obtain the modulation format classification result output by the model; comparing the modulation format classification result with the label dataset of known modulation type data, calculating the loss function value, which is used to characterize the prediction bias; using an optimization algorithm to iteratively update the parameters of the neural network model until the number of iterations reaches a preset threshold and training is terminated; after the model training is completed, the modulation format classification performance is verified on the test dataset, and the modulation classification accuracy of the model is output.

[0036] A smart modulation format recognition system for photon-assisted millimeter-wave OTFS, characterized in that it comprises:

[0037] The signal generation module maps modulation symbols corresponding to different modulation formats to a two-dimensional data matrix in the time-delay-Doppler domain of the OTFS frame at the transmitting end, and performs inverse symplectic finite Fourier transform and Heisenberg transform sequentially to generate a time-domain OTFS signal. The transmitting module converts the time-domain OTFS signal into a millimeter-wave signal using a photon-assisted method and transmits it into free space for transmission. The signal receiving module performs Wigner transform and symplectic finite Fourier transform sequentially on the received signal at the receiving end to obtain the OTFS time-delay-Doppler domain signal. The data processing module extracts the time-delay dimension subcarrier signal corresponding to a preset index from the OTFS time-delay-Doppler domain signal, and processes the extracted subcarrier... Wave data is separated into real and imaginary data, concatenated along the channel dimension to form real / imaginary data (I / Q data), and constellation diagram data is generated based on the I / Q data. The I / Q data and constellation diagram data are combined to construct an input dataset containing two types of data, which are then divided proportionally into training, validation, and test datasets. The model training module is used to train a pre-built intelligent modulation format recognition neural network model based on the training and validation datasets. The modulation recognition module is used to acquire the signal to be modulated and classified, construct a dual-channel input dataset based on the signal to be modulated and classified, and input it into the trained intelligent modulation format recognition neural network model to obtain the modulation classification result. Attached Figure Description

[0038] To more clearly illustrate the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of the intelligent modulation format recognition system for photon-assisted millimeter-wave OTFS of the present invention;

[0040] Figure 2 This is a schematic diagram illustrating the mapping of different modulation formats of the present invention to OTFS frames;

[0041] Figure 3 This is a schematic diagram illustrating the feature extraction of I / Q data and constellation diagram data and the construction of the dataset according to the present invention;

[0042] Figure 4 This is a schematic diagram of the intelligent modulation format recognition neural network model of the present invention;

[0043] Figure 5 This is a diagram showing the results of the dual-branch confusion matrix of the present invention;

[0044] Figure 6 This is a diagram showing the I / Q data confusion matrix results of the present invention;

[0045] Figure 7 This is a diagram showing the result of the constellation diagram data branch confusion matrix of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0047] The technical solutions of the present invention will be described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0048] Figure 1This is a schematic diagram of the intelligent modulation format recognition system for photon-assisted millimeter-wave OTFS according to the present invention. In the signal generation module, at the transmitting end, the modulation symbols corresponding to different modulation formats are mapped to the time-delay-Doppler domain two-dimensional data matrix of the OTFS frame, and inverse symplectic finite Fourier transform and Heisenberg transform are performed sequentially to generate a time-domain OTFS signal. Through the transmitting module, the time-domain OTFS signal is converted into a millimeter-wave signal using a photon-assisted method and transmitted into free space for transmission. In the signal receiving module, at the receiving end, the received signal undergoes Wigner transform and symplectic finite Fourier transform sequentially to obtain the OTFS time-delay-Doppler domain signal. The data processing module extracts the time-delay dimension subcarrier signals corresponding to preset indices from the OTFS time-delay-Doppler domain signal, separating the extracted subcarrier data into real and imaginary data, concatenating them along the channel dimension to form real / imaginary data (I / Q data), and generating constellation diagram data based on the I / Q data. The I / Q data and constellation diagram data are combined to construct an input dataset containing two data types, proportionally divided into training, validation, and test datasets. The model training module trains a pre-built intelligent modulation format recognition neural network model based on the training and validation datasets. In the modulation recognition module, a dual-channel input dataset is constructed based on the signal to be modulated and classified, and input into the trained intelligent modulation format recognition neural network model to obtain the modulation classification result.

[0049] Figure 2 This diagram illustrates the mapping of different modulation formats to an OTFS frame according to the present invention. Information symbols are arranged on a two-dimensional time-delay-Doppler domain grid of size T×M, where T represents the number of subcarrier indices along the time-delay axis and M represents the number of symbol indices along the Doppler axis. Within an OTFS frame, the OOK modulation format maps to a subcarrier range of [1, N], the 4QAM modulation format maps to a subcarrier range of [N+1, 2N], the 16QAM modulation format maps to a subcarrier range of [2N+1, 3N], the 64QAM modulation format maps to a subcarrier range of [3N+1, 4N], the PAM4 modulation format maps to a subcarrier range of [4N+1, 5N], and the PAM8 modulation format maps to a subcarrier range of [5N+1, 6N], where N is the number of subcarrier indices mapped for each modulation format. The mapping subcarrier range of all modulation formats exactly covers all time delay dimension subcarriers [1,T] in the time delay-Doppler domain, and the subcarrier ranges corresponding to any two modulation formats do not overlap, so as to realize independent identification and non-aliasing transmission of data of different modulation formats in the time delay-Doppler domain subcarrier dimension.

[0050] Figure 3This diagram illustrates the feature extraction and dataset construction of I / Q data and constellation diagram data according to the present invention. For each row of 1×M data corresponding to the received data transformed to the delay-Doppler domain, virtual-real separation and channel-dimension concatenation are performed. For each delay-dimension subcarrier, an I / Q data matrix of dimension 2×M is obtained, represented as follows:

[0051]

[0052] in, This represents the I / Q data matrix of the k-th subcarrier. This represents the complex vector data of the k-th subcarrier. This indicates the operation of taking the real part. This indicates the operation of taking the imaginary part, where M is the Doppler dimension.

[0053] The I / Q data matrix is ​​subjected to amplitude limiting, which is expressed as:

[0054]

[0055] in, The data matrix after amplitude limiting has a dimension of 2×M; and These are the maximum and minimum values ​​within the preset amplitude limit range, respectively;

[0056] The clipped I / Q signal is linearly mapped to pixel space, as follows:

[0057]

[0058] Where x and y are the pixel coordinates after mapping the real and imaginary parts of the signal, respectively. Let x,y ∈ [0,y] be the coordinate axis dimensions of the constellation diagram, and x,y ∈ [0,y]. A constellation diagram is drawn with a black background and white data pixels, represented as follows:

[0059]

[0060] in, This represents the pixel value matrix of the constellation map corresponding to the k-th subcarrier at coordinates (x, y); the constellation map is normalized to obtain the final constellation map data input to the network.

[0061]

[0062] in, The normalized constellation map pixel matrix has a dimension of 1×P×P. The generated I / Q data in 6 different modulation formats and the constellation map data are combined to construct a multimodal input dataset containing two types of data. The dataset is divided into training dataset, validation dataset and test dataset according to the proportion.

[0063] Figure 4 This is a schematic diagram of the intelligent modulation format recognition neural network model of the present invention. The model includes a first branch, a second branch, and a feature fusion and classification module. The first branch adopts a structure combining a residual network and a long short-term memory network. The input is an I / Q data matrix, which passes sequentially through a network structure composed of multiple stacked residual modules. Each residual module includes a convolutional layer, a Selu activation function layer, a batch normalization layer, and a dropout layer. The data output from the residual network is then input to a two-layer long short-term memory network (LSTM). After flattening the data output from the LSTM network, it passes sequentially through a fully connected layer, a batch normalization layer, and a dropout layer to output the first feature vector. The second branch adopts a two-dimensional convolutional neural network structure. The input is a constellation diagram data matrix, which passes sequentially through... The network structure consists of three stacked convolutional modules, each including a convolutional layer, a batch normalization layer, a ReLU activation function layer, and a pooling layer. After flattening the output data, it is sequentially passed through a fully connected layer, a batch normalization layer, and a dropout layer to output a second feature vector. The feature fusion and classification module concatenates the first and second feature vectors along the channel dimension to obtain a fused feature vector. The fused feature vector is then sequentially passed through a combination structure containing two fully connected layers, a ReLU activation function layer, and a batch normalization layer, followed by a classification head with a Softmax layer to output classification results for six modulation formats.

[0064] Figure 5 This is a diagram showing the dual-branch confusion matrix result of the present invention. After the model training is completed, the test dataset is input into the trained model, and the modulation classification result is output, resulting in a modulation classification confusion matrix. It can be seen that the confusion matrix exhibits a clear diagonal distribution characteristic, with 16QAM, 64QAM, and PAM8 modulation formats achieving 100% classification accuracy; approximately 3% of OOK data, 2% of 4QAM data, and 2% of PAM4 data are identified as PAM8 data, resulting in an overall recognition accuracy of 98.86%. Experimental results demonstrate that the dual-branch fusion model used in this invention can achieve high-precision recognition of six modulation formats.

[0065] Figure 6This is a diagram showing the I / Q data confusion matrix results of this invention. The result represents the modulation classification confusion matrix of the I / Q data network model on the test dataset when only I / Q data is used as input and constellation diagram data is not fused. It can be seen that when using the I / Q data network alone, the model achieves 100% classification accuracy for OOK, 4QAM, and 64QAM modulation formats; approximately 4% of 16QAM data is misclassified as 64QAM data, 1% of PAM4 data is misclassified as 16QAM data, and approximately 14% of PAM8 signals are misclassified as PAM4 signals, with an overall recognition accuracy of 96.90%. The performance of modulation format classification relying solely on I / Q data is slightly lower than that of the dual-branch model.

[0066] Figure 7 This is the result of the constellation graph data branch confusion matrix of this invention. This result shows the modulation classification confusion matrix of the constellation graph data branch network model on the test set when only constellation graph data is used as input, without fusing I / Q data. It can be seen that when using the constellation graph data branch network alone, the model achieves 100% classification accuracy for the four modulation formats: OOK, 4QAM, 64QAM, and PAM8. Approximately 1% of 16QAM data is misclassified as 64QAM data, but nearly 30% of PAM4 data is misclassified as PAM8 data, resulting in an overall recognition accuracy of 94.61%. Experimental results demonstrate that the dual-branch model, which integrates both, can fully leverage complementary advantages to achieve optimal classification performance.

[0067] Implementation steps:

[0068] In this embodiment, Figure 1 This is a schematic diagram of the intelligent modulation format recognition system for photon-assisted millimeter-wave OTFS of the present invention. At the transmitting end, the signal generation module maps the modulation symbols corresponding to different modulation formats to the time-delay-Doppler domain two-dimensional data matrix of the OTFS frame. Figure 2This diagram illustrates the mapping of different modulation formats to an OTFS frame according to the present invention. Information symbols are arranged on a two-dimensional time-delay-Doppler domain grid of size T×M, where T represents the number of subcarrier indices along the time-delay axis and M represents the number of symbol indices along the Doppler axis. Within an OTFS frame, the OOK modulation format maps to a subcarrier range of [1, N], the 4QAM modulation format maps to a subcarrier range of [N+1, 2N], the 16QAM modulation format maps to a subcarrier range of [2N+1, 3N], the 64QAM modulation format maps to a subcarrier range of [3N+1, 4N], the PAM4 modulation format maps to a subcarrier range of [4N+1, 5N], and the PAM8 modulation format maps to a subcarrier range of [5N+1, 6N], where N is the number of subcarrier indices mapped for each modulation format. The mapped subcarrier ranges of all modulation formats exactly cover all time-delay dimension subcarriers [1,T] in the time-delay-Doppler domain, and the subcarrier ranges corresponding to any two modulation formats do not overlap, so as to achieve independent identification and aliasing-free transmission of data of different modulation formats in the time-delay-Doppler domain subcarrier dimension. Then, the inverse symplectic finite Fourier transform and Heisenberg transform are performed sequentially on the two-dimensional data matrix in the time-delay-Doppler domain to generate the time-domain OTFS signal.

[0069] The transmitting module uses a photon-assisted method to convert the time-domain OTFS signal into a millimeter-wave signal and transmits it into free space for transmission. In the signal receiving module, at the receiving end, the received signal undergoes a Wigner transform and a symplectic finite Fourier transform sequentially to obtain the OTFS time-delay-Doppler domain signal. The data processing module extracts the time-delay dimension subcarrier signals corresponding to preset indices from the OTFS time-delay-Doppler domain signal, separates the extracted subcarrier data into real and imaginary data, concatenates them along the channel dimension to form real / imaginary data (I / Q data), and generates constellation diagram data based on the I / Q data. The I / Q data and constellation diagram data are then combined to construct an input dataset containing both data types. Figure 3 This diagram illustrates the feature extraction of I / Q data and constellation diagram data and the construction of the dataset according to the present invention. For each row of 1×M data corresponding to the received subcarriers after transformation to the time-delay-Doppler domain, virtual-real separation and channel-dimension concatenation are performed. For each time-delay dimension subcarrier, a 2×M I / Q data matrix is ​​obtained. The I / Q data matrix is ​​then clipped, and the clipped I / Q signals are linearly mapped to pixel space. A constellation diagram is drawn with black as the background and white as the data pixels. The constellation diagram is normalized to a dimension of 1×P×P. The generated I / Q data and constellation diagram data of six different modulation formats are combined to construct a multimodal input dataset containing two types of data. The constructed dataset contains a total of 6120 samples. This dataset is then proportionally divided into a training dataset, a validation dataset, and a test dataset.

[0070] The model training module is used to train a pre-built intelligent modulation format recognition neural network model based on the training dataset and the validation dataset. Figure 4 This is a schematic diagram of the intelligent modulation format recognition neural network model of the present invention. The model includes a first branch, a second branch, and a feature fusion and classification module. The first branch adopts a structure combining a residual network and a long short-term memory network. The input is an I / Q data matrix, which passes sequentially through a network structure composed of multiple stacked residual modules. Each residual module includes a convolutional layer, a Selu activation function layer, a batch normalization layer, and a dropout layer. The data output from the residual network is then input to a two-layer long short-term memory network (LSTM). After flattening the data output from the LSTM network, it passes sequentially through a fully connected layer, a batch normalization layer, and a dropout layer to output the first feature vector. The second branch adopts a two-dimensional convolutional neural network structure. The input is a constellation diagram data matrix, which passes sequentially through... The network structure consists of three stacked convolutional modules, each including a convolutional layer, a batch normalization layer, a ReLU activation function layer, and a pooling layer. After flattening the output data, it is sequentially passed through a fully connected layer, a batch normalization layer, and a dropout layer to output a second feature vector. The feature fusion and classification module concatenates the first and second feature vectors along the channel dimension to obtain a fused feature vector. The fused feature vector is then sequentially passed through a combination structure containing two fully connected layers, a ReLU activation function layer, and a batch normalization layer, followed by a classification head with a Softmax layer to output classification results for six modulation formats.

[0071] Figure 5 This is a diagram showing the dual-branch confusion matrix result of the present invention. After the model training is completed, the test dataset is input into the trained model, and the modulation classification result is output, resulting in a modulation classification confusion matrix. It can be seen that the confusion matrix exhibits a clear diagonal distribution characteristic, with 16QAM, 64QAM, and PAM8 modulation formats achieving 100% classification accuracy; approximately 3% of OOK data, 2% of 4QAM data, and 2% of PAM4 data are identified as PAM8 data, resulting in an overall recognition accuracy of 98.86%. Experimental results demonstrate that the dual-branch fusion model used in this invention can achieve high-precision recognition of six modulation formats. Figure 6This is a diagram showing the I / Q data confusion matrix results of this invention. The result represents the modulation classification confusion matrix of the I / Q data network model on the test dataset when only I / Q data is used as input and constellation diagram data is not fused. It can be seen that when using the I / Q data network alone, the model achieves 100% classification accuracy for OOK, 4QAM, and 64QAM modulation formats; approximately 4% of 16QAM data is misclassified as 64QAM data, 1% of PAM4 data is misclassified as 16QAM data, and approximately 14% of PAM8 signals are misclassified as PAM4 signals, with an overall recognition accuracy of 96.90%. The performance of modulation format classification relying solely on I / Q data is slightly lower than that of the dual-branch model. Figure 7 This is the result of the constellation graph data branch confusion matrix of this invention. This result shows the modulation classification confusion matrix of the constellation graph data branch network model on the test set when only constellation graph data is used as input, without fusing I / Q data. It can be seen that when using the constellation graph data branch network alone, the model achieves 100% classification accuracy for the four modulation formats: OOK, 4QAM, 64QAM, and PAM4. Approximately 1% of 16QAM data is misclassified as 64QAM data, but nearly 30% of PAM4 data is misclassified as PAM8 data, resulting in an overall recognition accuracy of 94.61%. Experimental results demonstrate that the dual-branch model, which integrates both, can fully leverage complementary advantages to achieve optimal classification performance.

[0072] (vi) Key Technological Advantages

[0073] This invention discloses an intelligent modulation format identification method and system for photon-assisted millimeter-wave OTFS. Targeting the physical layer characteristics of photon-assisted millimeter-wave OTFS systems, a parallel dual-branch feature extraction network is proposed, fusing delay-Doppler domain I / Q data and constellation diagram data. This achieves high-precision intelligent identification of multiple modulation formats within the system. By mapping different modulation formats to different subcarrier intervals in the delay-Doppler domain, independent identification and aliasing-free transmission of multiple modulation formats are achieved, providing effective support for signal identification in scenarios where multiple modulation formats coexist. The intelligent modulation format identification neural network model constructed in this invention employs a first branch combining a residual network and a long short-term memory network to extract spatiotemporal features of the I / Q data, and a second branch using a convolutional neural network to extract image features of the constellation diagram. This dual-branch feature fusion improves the accuracy and robustness of modulation format identification.

Claims

1. A method and system for intelligent modulation format recognition of photon-assisted millimeter-wave OTFS, characterized in that, The method includes the following steps: Step 1: At the transmitting end, the modulation symbols corresponding to different modulation formats are mapped to the time-delay-Doppler domain two-dimensional data matrix of the OTFS frame, and the inverse symplectic finite Fourier transform and Heisenberg transform are performed in sequence to generate the time-domain OTFS signal. Step 2: The time-domain OTFS signal is converted into a millimeter-wave signal using a photon-assisted method and transmitted into free space; Step 3: At the receiving end, perform Wigner transform and symplectic finite Fourier transform on the received signal in sequence to obtain the OTFS time-delay-Doppler domain signal; extract the time-delay dimension subcarrier signal corresponding to the preset index in the OTFS time-delay-Doppler domain signal, separate the extracted subcarrier data into real part data and imaginary part data, and concatenate them in the channel dimension to form real part / imaginary part data (I / Q data), and generate constellation diagram data based on the I / Q data; construct the I / Q data and constellation diagram data together to form an input dataset containing two types of data, and divide it into training dataset, validation dataset and test dataset according to the proportion; Step 4: Train the pre-built intelligent modulation format recognition neural network model based on the training dataset and validation dataset; Step 5: Obtain the signal to be modulated and classified, construct a dual-channel input dataset based on the signal to be modulated and classified, and input it into the trained intelligent modulation format recognition neural network model to obtain the modulation classification result.

2. The intelligent modulation format recognition method for photon-assisted millimeter-wave OTFS according to claim 1, characterized in that, Step 1, which maps modulation symbols corresponding to different modulation formats to the time-delay-Doppler domain two-dimensional data matrix of the OTFS frame, specifically includes: setting the modulation format index as t, where t corresponds to different modulation formats sequentially; within the same OTFS frame, different modulation formats are mapped to different subcarrier ranges in the time-delay-Doppler domain, and the subcarrier ranges corresponding to different modulation formats are continuous and do not overlap, satisfying: Where N is the number of subcarrier indices mapped for each modulation format. , Let k be the start and end subcarrier indices of the t-th modulation format, respectively; the modulation format category to which any subcarrier index k belongs satisfies: in, This is a rounding down operation; under the allocation method, the mapped subcarrier range of all modulation formats exactly covers all time delay dimension subcarriers in the time delay-Doppler domain, and the subcarrier intervals corresponding to any two types of modulation formats do not overlap, so as to realize independent identification and non-aliasing transmission of data of different modulation formats in the time delay-Doppler domain subcarrier dimension.

3. The intelligent modulation format recognition method for photon-assisted millimeter-wave OTFS according to claim 1, characterized in that, Step 3 involves constructing an input dataset containing two types of data from the I / Q data and constellation diagram data. Specifically, this includes: performing virtual-real separation and channel-dimension concatenation on a row of 1×M data corresponding to each subcarrier received and transformed into the delay-Doppler domain. For each delay-dimension subcarrier, a 2×M I / Q data matrix is ​​obtained, represented as follows: in, This represents the I / Q data matrix of the k-th subcarrier. This represents the complex vector data of the k-th subcarrier. This indicates the operation of taking the real part. This indicates the operation of taking the imaginary part, where M is the Doppler dimension. The I / Q data matrix is ​​subjected to amplitude limiting, which is expressed as: in, This is the data matrix after amplitude limiting. and These are the maximum and minimum values ​​within the preset amplitude limit range, respectively; The clipped I / Q signal is linearly mapped to pixel space, as follows: Where x and y are the pixel coordinates after mapping the real and imaginary parts of the signal, respectively. Let x,y ∈ [0,y] be the coordinate axis dimensions of the constellation diagram, and x,y ∈ [0,y]. A constellation diagram is drawn with a black background and white data pixels, represented as follows: in, This represents the pixel value matrix of the constellation map corresponding to the k-th subcarrier at coordinates (x, y); the constellation map is normalized to obtain the final constellation map data input to the network. in, The normalized constellation map pixel matrix has a dimension of 1×P×P; the generated I / Q data and constellation map data are combined to construct an input dataset containing two types of data.

4. The intelligent modulation format recognition method for photon-assisted millimeter-wave OTFS according to claim 1, characterized in that, The pre-built intelligent modulation format recognition neural network model mentioned in step 4 specifically includes: the intelligent modulation format recognition neural network model comprises a first branch, a second branch, and a feature fusion and classification module, wherein: The first branch employs a structure combining a residual network and a long short-term memory network to process the I / Q data constructed in claim 3. The input to the first branch is the I / Q data matrix. The I / Q data matrix passes sequentially through a network structure composed of multiple stacked residual modules, each residual module including a convolutional layer, a Selu activation function layer, a batch normalization layer, and a dropout layer. The data output from the residual network is then input to a two-layer long short-term memory network (LSTM). After flattening the data output from the long short-term memory network, it sequentially passes through a fully connected layer, a batch normalization layer, and a dropout layer to output the first feature vector. The second branch employs a two-dimensional convolutional neural network structure to process the constellation diagram data constructed in claim 3. The input to the second branch is the constellation diagram data matrix. The constellation diagram data matrix is ​​sequentially processed through a network structure consisting of three stacked convolutional modules, each of which includes a convolutional layer, a batch normalization layer, a ReLU activation function layer, and a pooling layer. After flattening the output data, it is sequentially processed through a fully connected layer, a batch normalization layer, and a dropout layer to output the second feature vector. Feature fusion and classification module: The first feature vector and the second feature vector are concatenated along the channel dimension to obtain a fused feature vector; the fused feature vector is then passed through a combination structure containing two fully connected layers, a ReLU activation function layer and a batch normalization layer, followed by a classification header of a Softmax layer to output the modulation format classification result. The dimension of the modulation classification result is the same as the number of modulation categories.

5. The intelligent modulation format recognition method for photon-assisted millimeter-wave OTFS according to claim 4, characterized in that, In the first branch: the residual network is used to suppress the impact of gradient explosion on network training, and extracts the spatial features of the time-delay-Doppler domain I / Q data and reduces the feature map size through convolutional layers; the long short-term memory network is used to extract the temporal features of the time-delay-Doppler domain I / Q data.

6. The intelligent modulation format recognition method for photon-assisted millimeter-wave OTFS according to claim 4, characterized in that, The training process of the intelligent modulation format recognition neural network model includes: inputting the training dataset into the pre-built intelligent modulation format recognition neural network model to obtain the modulation format classification result output by the model; comparing the modulation format classification result with the label dataset of known modulation type data, calculating the loss function value, which is used to characterize the prediction bias; using an optimization algorithm to iteratively update the parameters of the neural network model until the number of iterations reaches a preset threshold and training is terminated; after the model training is completed, the modulation format classification performance is verified on the test dataset, and the modulation classification accuracy of the model is output.

7. A smart modulation format recognition system for photon-assisted millimeter-wave OTFS, characterized in that, The system is used to implement the intelligent modulation format identification method for photon-assisted millimeter-wave OTFS according to any one of claims 1-6, the system comprising: The signal generation module is used at the transmitting end to map the modulation symbols corresponding to different modulation formats to the time-delay-Doppler domain two-dimensional data matrix of the OTFS frame, and then perform inverse symplectic finite Fourier transform and Heisenberg transform in sequence to generate the time-domain OTFS signal. The transmitting module is used to convert time-domain OTFS signals into millimeter-wave signals using a photon-assisted method and transmit them into free space for transmission. The signal receiving module is used to perform Wigner transform and symmetric finite Fourier transform on the received signal at the receiving end to obtain the OTFS time delay-Doppler domain signal. The data processing module is used to extract the time-delay dimension subcarrier signal corresponding to the preset index in the OTFS time-delay-Doppler domain signal, separate the extracted subcarrier data into real part data and imaginary part data, concatenate them in the channel dimension to form real part / imaginary part data (I / Q data), and generate constellation diagram data based on the I / Q data; the I / Q data and constellation diagram data are used to construct an input dataset containing two types of data, and divide it into training dataset, validation dataset and test dataset according to the proportion; The model training module is used to train a pre-built intelligent modulation format recognition neural network model based on the training dataset and the validation dataset. The modulation recognition module is used to acquire the signal to be modulated and classified, construct a dual-channel input dataset based on the signal to be modulated and classified, and input it into the trained intelligent modulation format recognition neural network model to obtain the modulation classification result.

Images