Intelligent scene recognition method and device suitable for 6g wireless communication

Abstract

The application discloses a kind of intelligent scene identification method and device suitable for 6G wireless communication, belong to scene identification field, method includes: the physical environment parameter of the scene to be identified is collected;The physical environment parameter of the scene to be identified is input to pre-trained classifier and classified, wherein, the classifier is obtained by training the classification model using the physical environment parameter of the scene.This method can collect the environment characteristics easily obtained in different 6G wireless communication scenes, through data preprocessing, classifier training, hyperparameter optimization and other steps, so that it has good scene identification performance, and can be applied to 6G multi-scene channel characteristic research, system design, network optimization and other key technologies.

Description

Technical Field

[0001] This invention relates to the field of scene recognition technology, and in particular to an intelligent scene recognition method and device suitable for 6G wireless communication. Background Technology

[0002] To meet users' needs for easy access to high-quality, stable network connections in various environments and to provide equitable network connectivity, 6th Generation Wireless Communication Technology (6G) aims to achieve the technological vision of "full coverage" and "full application." Compared to 5th Generation Wireless Communication Technology (5G), 6G will no longer be limited to terrestrial communication scenarios but will also supplement satellite communication, drone communication, and maritime communication, realizing an integrated "air-space-terrestrial-sea" communication network and supporting diverse application scenarios. Different communication scenarios exhibit different propagation methods for wireless signals, resulting in varying channel characteristics that affect signal reception quality and communication system performance. Therefore, in practical applications, only by quickly and accurately identifying 6G wireless communication scenarios can 6G wireless channel research be conducted more fully and comprehensively, thereby ensuring the efficiency, stability, and reliability of 6G communication systems. Traditional scenario identification methods are often based on subjective experience or simple table lookup matching, thus failing to meet the accuracy requirements of the 6G era. For example, scenario identification based on subjective experience is difficult to guarantee accuracy. Furthermore, matching scenarios through table lookups requires acquiring a large amount of physical environment features, making the data acquisition process extremely resource-intensive. Moreover, the physical boundaries of different scenarios inevitably overlap to some extent, making it difficult for traditional methods to make intelligent decisions. In view of these problems, there is an urgent need for new 6G wireless communication scenario identification methods.

[0003] In recent years, AI technology has been increasingly applied to the field of wireless channel research. Compared to traditional methods that often require significant subjective experience and resources, AI's powerful adaptability and big data processing capabilities enable it to intelligently adjust and optimize wireless communication scene recognition strategies in real time. It extracts multi-dimensional features from scene data, providing a more comprehensive understanding of the characteristics of wireless communication scenarios, ultimately achieving more flexible and intelligent scene recognition. AI's powerful ability to capture nonlinear relationships and its generalization capabilities allow it to more intelligently characterize the complexity of communication scenarios to cope with the numerous communication scenarios in 6G, providing more comprehensive and accurate scene recognition. In the data processing stage, for high-dimensional data, commonly used AI algorithms such as Auto Encoders (AEs) can be used to learn efficient representations of the data, which is crucial for constructing efficient scene recognition methods.

[0004] However, existing research on wireless communication scene recognition often uses channel measurement data as classifier input, which requires prior channel measurement activities, reducing the efficiency and real-time performance of scene recognition. Furthermore, many scene recognition methods fail to eliminate the influence of the magnitude relationship between dataset sample values ​​on training, leading to instability in the classifier learning process and slow convergence speed. Moreover, when the number of features in the dataset is limited, existing research rarely considers data augmentation methods, causing the classifier to be prone to overfitting and exhibiting poor generalization ability. In addition, current research is often limited to typical 5G scenarios with high discriminative power or to coarsely identify line-of-sight (LOS) or non-line-of-sight (NLOS) scenarios, with limited research on intelligent scene recognition for 6G communication scenarios. Summary of the Invention

[0005] This invention provides an intelligent scene recognition method and transposition method suitable for 6G wireless communication, used to capture 6G wireless communication scene features to identify different types of 6G communication scenes. The method utilizes unsupervised learning of scene physical environment parameters to extract their efficient representations and construct composite features to enhance the dataset. Then, based on the enhanced dataset and scene category, a classifier is trained in a supervised manner. Finally, a hyperparameter optimization algorithm is used to achieve high recognition accuracy.

[0006] A first aspect of this invention provides an intelligent scene recognition method suitable for 6G wireless communication, comprising the following steps:

[0007] Collect physical environment parameters of the scene to be identified;

[0008] The physical environment parameters of the scene to be identified are input into a pre-trained classifier for classification, wherein the classifier is obtained by training a classification model using the physical environment parameters of the scene.

[0009] In one embodiment of the present invention, before inputting the physical environment parameters of the scene to be identified into a pre-trained classifier for classification, the method further includes:

[0010] Collect scene physical environment parameters and label the corresponding scene categories. Preprocess the scene physical environment parameters and corresponding scene category data to obtain the scene dataset.

[0011] An efficient representation of the scene physical environment parameters in the scene dataset is extracted using an autoencoder, and the efficient representation of the scene physical environment parameters is combined with the original scene physical environment parameters to construct a composite feature of the scene physical environment parameters;

[0012] The classification model built using a multilayer perceptron takes the composite features of the scene's physical environment parameters and the corresponding scene category as input, and trains the classifier under the initial hyperparameter settings to build the classification model.

[0013] Set the range of hyperparameters to be optimized, use the optimization algorithm to automatically traverse all hyperparameter combinations, obtain the hyperparameter combination that makes the classifier perform optimally as the hyperparameters of the classification model, and obtain the trained classifier.

[0014] In one embodiment of the present invention, preprocessing of scene physical environment parameters and corresponding scene category data includes:

[0015] The scene's physical environment parameters and corresponding scene category data are normalized and one-hot encoded.

[0016] In one embodiment of the present invention, preprocessing of scene physical environment parameters and corresponding scene category data includes:

[0017] The scene physical environment parameters are processed using normalization. The minimum value is subtracted from the sample value of each type of parameter, and then divided by the range of sample values ​​to obtain the normalized scene physical environment parameters.

[0018] One-hot encoding is used to process scene categories, mapping each scene's true category to a unique binary vector, with one element being 1 and the rest being 0. The category corresponding to the element with the value of 1 is the true category of the scene.

[0019] A scene dataset is constructed using a comma-separated value format, containing normalized physical environment parameters for different scenarios and scene categories with one-hot encoding.

[0020] In one embodiment of the present invention, an autoencoder is used to extract an efficient representation of the scene physical environment parameters in the scene dataset, and the efficient representation of the scene physical environment parameters is combined with the original scene physical environment parameters to construct a composite feature of the scene physical environment parameters, including:

[0021] Define the scene sample data in the scene dataset as X. j This includes the preprocessed normalized scene physical environment parameters F. j And the pre-processed one-hot encoded scene category L j ,

[0022] X j =[F j ,L j ], j=1,2,3,…,n

[0023] Where n represents the total number of scene samples;

[0024] Construct an autoencoder and initialize its network parameters. The size of the autoencoder's input layer is related to F. n Dimensions are consistent, encoder output layer is smaller than F n Dimension, the output layer of the decoder and F n Consistent dimensions;

[0025] Compress F using the encoder in the automatic encoder. n To a lower dimension, C n Use a decoder to restore it to R j Calculate R j and F j The difference between them is described using mean squared error, and the network parameters of the autoencoder are iteratively adjusted to minimize this difference.

[0026] C j With F j Combine to construct composite feature S j The resulting formula is:

[0027] S j =[C j ,F j ], j=1,2,3,…,n.

[0028] In one embodiment of the present invention, a classification model constructed using a multilayer perceptron takes the composite features of the scene's physical environment parameters and the corresponding scene category as input, and trains the constructed classification model under initial hyperparameter settings, including:

[0029] Construct a classification model and initialize its parameters. The input layer size of the classification model is related to the composite feature S. j The model has consistent dimensions, consisting of multiple fully connected layers following the input layer. ReLU activation is used to introduce non-linearity. Each fully connected layer is preceded by a batch normalization layer to normalize the data for each batch. Each fully connected layer is followed by a random deactivation layer, which temporarily ignores certain neurons during iteration with a preset probability. The output layer size of the classification model is consistent with the total number of categories in the scene. Softmax is used as the activation function, ensuring that the predicted output P of the classification model... j Category L of real-world scenarios j They have the same format;

[0030] The classification model is trained with the following hyperparameters: batch size, number of training epochs, optimizer, learning rate, and loss function set to P0. j With L j The cross-entropy between categories.

[0031] In one embodiment of the present invention, a range of hyperparameters to be optimized is set, and an optimization algorithm is used to automatically traverse all hyperparameter combinations to obtain the hyperparameter combination that makes the classifier perform optimally as the hyperparameters of the classification model, thereby obtaining a trained classifier, including:

[0032] Based on the training results of the classification model during training, set the range of hyperparameters to be optimized;

[0033] Based on the set range of hyperparameters to be optimized, candidate hyperparameter combinations are constructed. A classifier is trained using each set of hyperparameters. The average accuracy of the classifier on the validation set is evaluated through K-fold cross-validation to assess the performance of the classifier under the set of hyperparameters. The best-performing set of all candidate hyperparameter combinations is selected as the optimal hyperparameter combination for the classifier. The classification model is then retrained under the optimal hyperparameter combination.

[0034] A second aspect of the present invention provides an intelligent scene recognition device suitable for 6G wireless communication, comprising:

[0035] The data acquisition module is used to collect physical environment parameters of the scene to be identified.

[0036] The recognition module is used to input the physical environment parameters of the scene to be recognized into a pre-trained classifier for classification, wherein the classifier is obtained by training a classification model using the physical environment parameters of the scene.

[0037] In one embodiment of the present invention, the device further includes:

[0038] The module is used to collect scene physical environment parameters and label the corresponding scene categories. The scene physical environment parameters and corresponding scene category data are preprocessed to obtain the scene dataset.

[0039] The processing module is used to extract an efficient representation of the scene physical environment parameters in the scene dataset using an autoencoder, and to combine the efficient representation of the scene physical environment parameters with the original scene physical environment parameters to construct a composite feature of the scene physical environment parameters.

[0040] The training module is used to build a classification model using a multilayer perceptron. It takes the composite features of the scene's physical environment parameters and the corresponding scene category as input and trains the classifier under the initial hyperparameter settings to build the classification model.

[0041] The search module is used to set the range of hyperparameters to be optimized, and to automatically traverse all hyperparameter combinations using an optimization algorithm to obtain the hyperparameter combination that makes the classifier perform optimally as the hyperparameters of the classification model, thereby obtaining a trained classifier.

[0042] In one embodiment of the present invention, preprocessing of scene physical environment parameters and corresponding scene category data includes:

[0043] The scene's physical environment parameters and corresponding scene category data are normalized and one-hot encoded.

[0044] The intelligent scene recognition method and apparatus for 6G wireless communication, as described in this invention, can capture the characteristics of physical environment parameters of 6G wireless communication scenarios and identify scene categories based on these parameters. By training on a dataset processed with normalization, one-hot encoding, and an autoencoder, and using a hyperparameter optimization algorithm to obtain the optimal hyperparameter combination, the proposed classifier can improve scene recognition performance and achieve high recognition accuracy. Furthermore, this invention has high versatility and can be applied to different types of 6G wireless communication scenarios, outperforming the widely used K-Nearest Neighbor (K-NN) algorithm. This invention demonstrates excellent performance in intelligent scene recognition for 6G wireless communication and can be used for key technologies such as 6G multi-scenario channel characteristic research, system design, network optimization, and network planning.

[0045] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0046] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0047] Figure 1 A flowchart illustrating an intelligent scene recognition method suitable for 6G wireless communication according to an embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram illustrating the training process of an intelligent scene recognition method for 6G wireless communication provided according to an embodiment of the present invention.

[0049] Figure 3 An electronic map of a communication scenario provided according to an embodiment of the present invention;

[0050] Figure 4 This is a schematic diagram of the algorithm for the intelligent scene recognition method provided according to an embodiment of the present invention;

[0051] Figure 5 This is a comparison chart showing the recognition accuracy of a classifier trained using the optimal hyperparameter combination and a K-NN classifier on the same dataset, according to an embodiment of the present invention.

[0052] Figure 6This is a schematic diagram of the structure of an intelligent scene recognition device suitable for 6G wireless communication provided according to an embodiment of the present invention. Detailed Implementation

[0053] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0054] Figure 1 This is a flowchart illustrating an intelligent scene recognition method suitable for 6G wireless communication according to an embodiment of the present invention.

[0055] like Figure 1 As shown, the intelligent scene recognition method suitable for 6G wireless communication includes the following steps:

[0056] In step S1, the physical environment parameters of the scene to be identified are collected.

[0057] In embodiments of the present invention, the collected scene physical environment parameters may include information containing key geographical features such as terrain and buildings.

[0058] In step S2, the physical environment parameters of the scene to be identified are input into a pre-trained classifier for classification. The classifier is obtained by training a classification model using the physical environment parameters of the scene.

[0059] It is understood that the embodiments of the present invention pre-train a classifier, the input of which is the collected scene physical environment parameters, and the output is the category corresponding to the scene.

[0060] Combination Figure 2 As shown, the training process of the classifier in the embodiments of the present invention is described in detail.

[0061] In one embodiment of the present invention, before inputting the physical environment parameters of the scene to be identified into a pre-trained classifier for classification, the method further includes:

[0062] Collect scene physical environment parameters and label the corresponding scene categories. Preprocess the scene physical environment parameters and corresponding scene category data to obtain the scene dataset.

[0063] An efficient representation of scene physical environment parameters in a scene dataset is extracted using an autoencoder. The efficient representation of scene physical environment parameters is then combined with the original scene physical environment parameters to construct a composite feature of scene physical environment parameters.

[0064] The classification model built using a multilayer perceptron takes the composite features of the scene's physical environment parameters and the corresponding scene category as input, and trains the classifier under the initial hyperparameter settings to build the classification model.

[0065] Set the range of hyperparameters to be optimized, use the optimization algorithm to automatically traverse all hyperparameter combinations, obtain the hyperparameter combination that makes the classifier perform best as the hyperparameters of the classification model, and obtain the trained classifier.

[0066] First, the training data required for training is collected. In this embodiment of the invention, the scene physical environment parameters are collected based on an electronic map containing detailed information on key geographical features such as terrain, buildings, and roads, and the corresponding scene categories are recorded.

[0067] To eliminate the influence of numerical magnitude on training, normalization and one-hot encoding were used to process the scene physical environment parameter data and the corresponding scene category data to construct a scene dataset. Then, the dataset was divided into training, validation, and test sets for subsequent training, optimization, and evaluation of the classifier.

[0068] For the training set data, an autoencoder is used to extract efficient representations of the scene physical environment parameters. These representations are then combined with the original scene physical environment parameters to construct composite features, thereby enhancing the dataset and improving the generalization ability of the classifier.

[0069] For a classifier built using a multilayer perceptron, composite features and the corresponding scene category are used as inputs. The classifier is trained under the initial hyperparameter settings, and the recognition accuracy is verified using a validation set.

[0070] Set the range of hyperparameters to be optimized, and use an optimization algorithm to automatically traverse all hyperparameter combinations to obtain the hyperparameter combination that makes the classifier perform optimally. Then, retrain the classifier with the optimal hyperparameter settings, and evaluate the final performance of the classifier using a test set.

[0071] In embodiments of the present invention, the scene physical environment parameters and corresponding scene category data are normalized and subjected to one-hot encoding. Specifically:

[0072] The scene physical environment parameters are processed by normalization. The minimum value is subtracted from the sample value of each type of parameter, and then divided by the range of sample values ​​to obtain the normalized scene physical environment parameters.

[0073] One-hot encoding is used to process scene categories, mapping each scene's true category to a unique binary vector, with one element being 1 and the rest being 0. The category corresponding to the element with the value of 1 is the true category of the scene.

[0074] A scene dataset containing normalized physical environment parameters and one-hot encoded scene categories for different scenarios is constructed using comma-separated values ​​(CSV) format.

[0075] The constructed scenario dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The training set is used for subsequent classifier training; the validation set is used to test the classifier's performance and help understand the direction of hyperparameter optimization; and the test set is used to evaluate the classifier's final performance, ensuring that the classifier has good generalization ability on unseen data.

[0076] In one embodiment of the present invention, an autoencoder is used to extract an efficient representation of scene physical environment parameters from a scene dataset. This efficient representation is then combined with the original scene physical environment parameters to construct a composite feature of the scene physical environment parameters, including:

[0077] Define the scene sample data in the scene dataset as X j This includes the preprocessed normalized scene physical environment parameters F. j And the pre-processed one-hot encoded scene category L j ,

[0078] X j =[F j ,L j ], j=1,2,3,…,n

[0079] Where n represents the total number of scene samples;

[0080] Construct an autoencoder and initialize its network parameters. The size of the autoencoder's input layer is related to F. n Dimensions are consistent, encoder output layer is smaller than F n Dimension, the output layer of the decoder and F n Consistent dimensions;

[0081] Compress F using the encoder in the automatic encoder. n To a lower dimension, C n Use a decoder to restore it to R j Calculate R j and F j The difference between them is described using mean squared error (MSE), and the network parameters of the autoencoder are iteratively adjusted to minimize this difference;

[0082] C j With F j Combine to construct composite feature S jThe resulting formula is:

[0083] S j =[C j ,F j ], j=1,2,3,…,n.

[0084] In one embodiment of the present invention, a classification model constructed using a multilayer perceptron takes the composite features of the scene's physical environment parameters and the corresponding scene category as input, and trains the constructed classification model under initial hyperparameter settings, including:

[0085] Construct a classification model (MLP network) and initialize the model parameters. The input layer size of the classification model and the composite features S are... j The model has consistent dimensions, consisting of multiple fully connected layers following the input layer. ReLU activation is used to introduce non-linearity. Each fully connected layer is preceded by a batch normalization layer to normalize the data for each batch, making the model more convergent. Each fully connected layer is followed by a random deactivation layer, which temporarily ignores certain neurons with a preset probability during iteration, increasing the model's generalization ability. The output layer size of the classification model is consistent with the total number of categories in the scene. Softmax is used as the activation function, ensuring that the predicted output P of the classification model... j Category L of real-world scenarios j They have the same format;

[0086] The classification model is trained with the following hyperparameters: batch size, number of training epochs, optimizer, learning rate, and loss function set to P0. j With L j Categorical Cross Entropy (CCR) between the two.

[0087] Furthermore, using samples from the validation set, the ratio of correctly identified samples to the total number of samples is calculated as the recognition accuracy to test the performance of the classification model.

[0088] In one embodiment of the present invention, a range of hyperparameters to be optimized is set, and an optimization algorithm is used to automatically traverse all hyperparameter combinations to obtain the hyperparameter combination that makes the classifier perform optimally as the hyperparameters of the classification model, thereby obtaining a trained classifier, including:

[0089] Based on the training results of the classification model during training, set the range of hyperparameters to be optimized;

[0090] Based on the set range of hyperparameters to be optimized, candidate hyperparameter combinations are constructed. A classifier is trained using each set of hyperparameters. The average accuracy of the classifier on the validation set is evaluated through K-fold cross-validation to assess the performance of the classifier under the set of hyperparameters. The best-performing set of all candidate hyperparameter combinations is selected as the optimal hyperparameter combination for the classifier. The classification model is then retrained under the optimal hyperparameter combination.

[0091] Specifically, the performance of the root classification model during training determines the range of hyperparameters to be optimized. For example, if the classification model has low accuracy on the training set, it indicates underfitting, and the number of training epochs and the learning rate can be appropriately increased. If the classification model has high accuracy on the training set but low accuracy on the validation set, it indicates that the model may be overfitting, and the number of training epochs and the learning rate can be appropriately reduced, or a different optimizer can be used.

[0092] The range of hyperparameters to be optimized is defined, and candidate hyperparameter combinations are constructed. Then, a classifier is trained using each set of hyperparameters, and the average accuracy of the classifier on the validation set is evaluated using K-fold cross-validation (K-fold CV) to assess the classifier's performance under that set of hyperparameters. Finally, the best-performing set of hyperparameters is selected from all candidate combinations as the optimal hyperparameter combination for the classifier.

[0093] The classification model is retrained with the optimal hyperparameter combination settings. The recognition accuracy is calculated using samples from the test set to ultimately evaluate the classifier's performance and compare it with the contrast algorithm.

[0094] The present invention provides an intelligent scene recognition method applicable to 6G wireless communication through a specific embodiment.

[0095] Step S1: Based on an electronic map containing detailed information on key geographical features such as terrain, buildings, and roads, physical environment parameters were collected for four scenarios in a typical 6G city: main roads, secondary roads, high-rise residential streets, and low-rise residential streets. The electronic maps for these scenarios are available in [link to relevant documentation]. Figure 3 The scene's physical environment parameters include: average building spacing, average building height, building elevation change, area LOS percentage, and indoor percentage, which are respectively measured using... V b ,P LoS ,P i This indicates the average building spacing. Defined as:

[0096]

[0097] Among them, s sec s represents the total area of ​​the scene sectors. iLet c be the base area of ​​building i. i Let n be the perimeter of building i. b This represents the total number of buildings within the sector. The overall average building spacing formula reflects the average building height in the scene.

[0098] Average building height Defined as:

[0099]

[0100] Where h i This represents the height of building i. The formula for the overall average building height reflects the density of buildings in the scene.

[0101] Building undulation height V b This reflects the degree of change in building height within a scene, and is defined as:

[0102]

[0103] Regional LOS percentage P LoS Defined as:

[0104]

[0105] Where n LoS n represents the number of Loss grid cells within a scene sector. g This represents the total number of grid cells within a scene sector.

[0106] Indoor proportion P i This reflects the proportion of indoor area within a scene sector, and is defined as:

[0107]

[0108] At the same time, the corresponding scene categories are recorded.

[0109] Step S2: For the five types of scene physical environment parameter data and corresponding scene category data obtained in Step S1, normalization and one-hot encoding are used to process these data to eliminate the influence of the numerical magnitude relationship on training, and a scene dataset is constructed. Then, the dataset is divided into training set, validation set, and test set for subsequent training, optimization, and evaluation of the classifier. See the diagram for step S2. Figure 4 .

[0110] Specifically, in this embodiment, step S2 includes:

[0111] Step S201: Use normalization to process the scene's physical environment parameters, and convert each type of parameter X... i Sample x in i,j Numerical value minus minimum value x i,minand divide by the range x of the sample values ​​of that type. i,max -x i,min Normalized scene physical environment parameter samples are obtained.

[0112]

[0113] Where m is the physical environment parameter of the scene, which is 5 in this example.

[0114] Step S202: Use one-hot encoding to process scene categories. Map each scene's true category to a unique binary vector, where one element is 1 and the rest are 0. The category corresponding to the element that is 1 is the scene's true category. For example, the scene category of the main road scene is 0, so its one-hot encoded category is [1,0,0,0]; the scene category of the side road scene is 1, so its one-hot encoded category is [0,1,0,0].

[0115] Step S203: Construct a scene dataset using CSV format, which contains normalized physical environment parameters for 5 different scenarios and scene categories with one-hot encoding.

[0116] Step S204: For the dataset obtained in step S203, divide it into a training set, a validation set, and a test set in an 8:1:1 ratio. The training set is used for subsequent classifier training; the validation set is used to test the classifier's performance and help understand the direction of hyperparameter optimization; the test set is used to evaluate the final performance of the classifier and ensure that the classifier has good generalization ability on unseen data.

[0117] Step S3: For the training set data obtained in Step S2, after removing scene categories, use an autoencoder to extract efficient representations of the physical environment parameters of the five scene categories. Then, combine these representations with the original scene physical environment parameters to construct composite features, thereby enhancing the dataset and improving the generalization ability of the classifier. See the diagram for Step S3. Figure 4 .

[0118] Specifically, in this embodiment, step S3 includes:

[0119] Step S301: Define the scene sample data in the training set obtained in step S204 as X. j It includes: collecting the preprocessed normalized scene physical environment parameters F in step S201 in step S101. j And the one-hot encoded scene category L after step S202 processing j The resulting formula is:

[0120]

[0121] Xj =[F j ,L j ], j=1,2,3,…,n

[0122] Where n represents the total number of scene samples.

[0123] Step S302: Construct an autoencoder and initialize its network parameters. Set the encoder input layer size to 5, the encoder output layer size to 3, and the decoder output layer size to 5.

[0124] In step S303, the parameters of the encoder and decoder are respectively represented by θ. encoder and θ decoder express:

[0125] θ encoder =[w encoder ,b encoder ]

[0126] θ decoder =[w decoder ,b decoder ]

[0127] Where w encoder and b encoder These are the weights and bias parameters of the encoder network, respectively; w decoder and b decoder These are the weights and bias parameters of the decoder network, respectively.

[0128] Compress F using the encoder in the automatic encoder. j To a lower dimension, C j The resulting formula is:

[0129] C j =ReLU(w encoder F j +b encoder )

[0130] Where ReLU() is the rectified linear unit (RCU) as the activation function, its expression is:

[0131]

[0132] Then use the decoder to restore it to R. j Use MSE to calculate R j and F j Differences between RE :

[0133]

[0134] Then, the network parameters of the autoencoder are iteratively adjusted to minimize l. RE The parameter update expression for step k is:

[0135]

[0136]

[0137] Where α represents the learning rate, l RE Regarding θ encoder and θ decoder The gradients are respectively and

[0138] Step S304: Use C obtained in S301 j to compare it with F j Combine to construct composite feature S j The resulting formula is:

[0139] S j =[C j ,F j ], j=1,2,3,…,n

[0140] Step S4: For a classifier built using a multilayer perceptron, use composite features S j and the corresponding scene category L j As input, the classifier is trained with the initial hyperparameter settings, and the recognition accuracy is tested using the validation set.

[0141] Specifically, in this embodiment, step S4 includes:

[0142] Step S401: Construct an MLP network and initialize its network parameters. The MLP's input layer size is 8, followed by 5 fully connected layers with sizes of 32, 64, 128, 64, and 32, respectively. The ReLU nonlinear activation function is used to introduce nonlinearity. Each fully connected layer is preceded by a batch normalization layer to normalize the data for each batch, making the model more likely to converge. Each fully connected layer is followed by a random deactivation layer, where certain neurons in each layer are temporarily ignored during iterations with probabilities of 0.05, 0.1, 0.2, 0.1, and 0.05, respectively, to increase the model's generalization ability. The MLP's output layer size is 4, using the Softmax function as the activation function, ensuring that the predicted output P of the MLP classifier is... j Category L of real-world scenarios j They have the same format. The Softmax function expression is:

[0143]

[0144] Step S402: Train the MLP with the following hyperparameter settings: batch size is set to 32; number of training epochs is set to 100; the optimizer uses stochastic gradient descent (SGD); the learning rate is set to 0.001. The loss function is set to P... j With L j Categorical Cross Entropy CE Its expression is:

[0145]

[0146] Step S403: Using samples from the validation set, calculate the ratio of correctly identified samples to the total number of samples as the recognition accuracy to verify the performance of the MLP.

[0147] Step S5: Set the range of hyperparameters to be optimized, and use the optimization algorithm to automatically traverse all hyperparameter combinations to obtain the hyperparameter combination that makes the classifier perform optimally. Then, retrain the classifier with the optimal hyperparameter settings, and use the test set to evaluate the final performance of the classifier.

[0148] Specifically, in this embodiment, step S5 includes:

[0149] Step S501: Based on the classifier's performance in step S403, set the range of hyperparameters to be optimized. The batch size is set to 8, 16, and 32; the number of training epochs is set to 100, 200, and 400; the optimizers are SGD and Adam, respectively; and the learning rate is set to 0.01, 0.001, and 0.0001.

[0150] Step S502: For the range of hyperparameters to be optimized set in step S501, construct candidate hyperparameter combinations. Then, train a classifier using each set of hyperparameters, and evaluate the average accuracy of the classifier on the validation set using K-fold CV to assess the classifier's performance under that set of hyperparameters. Finally, select the best-performing set from all candidate hyperparameter combinations as the optimal hyperparameter combination for the classifier.

[0151] Step S503: Retrain the MLP with the optimal hyperparameter combination. The optimal hyperparameter combination is: batch size of 32; number of training epochs of 400; Adam optimizer; learning rate of 0.001.

[0152] Step S504: Using samples from the test set, calculate the ratio of the number of correctly identified samples to the total number of samples as the recognition accuracy, in order to finally evaluate the performance of the classifier and compare it with the comparison algorithm.

[0153] Specifically, in this embodiment, the proposed intelligent scene recognition method suitable for 6G wireless communication is used to identify four scenes in a typical 6G city: main roads, secondary roads, high-rise residential streets, and low-rise residential streets, and the results are compared with the K-NN algorithm. Since the setting of the K value in the K-NN algorithm has a significant impact on its performance, the recognition accuracy was tested with K values ​​of 1, 2, 3, 4, and 5. The proposed method uses the optimal hyperparameter setting. See the comparison chart of the recognition accuracy of the two methods. Figure 5 The proposed method achieves a recognition accuracy of 90.71%. In comparison, the highest recognition accuracy of the K-NN classifier trained and tested on the same dataset is 80.77%. This demonstrates that the proposed method has good scene recognition performance.

[0154] The intelligent scene recognition method for 6G wireless communication proposed in this invention collects physical environment parameters of 6G communication scenarios based on electronic maps. It is trained using a dataset processed by normalization, one-hot encoding, and an autoencoder, and an optimal hyperparameter optimization algorithm is used to obtain the optimal hyperparameter combination. Finally, the classifier under the optimal hyperparameter settings achieves high-precision scene recognition accuracy for various 6G communication scenarios, providing valuable reference for research on communication scene recognition and 6G wireless channels.

[0155] Next, referring to the accompanying drawings, an intelligent scene recognition device suitable for 6G wireless communication is described according to an embodiment of the present invention.

[0156] Figure 6 This is a schematic diagram of the structure of an intelligent scene recognition device suitable for 6G wireless communication provided according to an embodiment of the present invention.

[0157] like Figure 6 As shown, the intelligent scene recognition device 10 suitable for 6G wireless communication includes: a data acquisition module 100 and a recognition module 200.

[0158] The acquisition module 100 is used to acquire the physical environment parameters of the scene to be identified; the identification module 200 is used to input the physical environment parameters of the scene to be identified into a pre-trained classifier for classification, wherein the classifier is obtained by training a classification model using the physical environment parameters of the scene.

[0159] In an embodiment of the present invention, the intelligent scene recognition device 10 suitable for 6G wireless communication further includes:

[0160] The module is used to collect scene physical environment parameters and label the corresponding scene categories. The scene physical environment parameters and corresponding scene category data are preprocessed to obtain the scene dataset.

[0161] The processing module is used to extract efficient representations of scene physical environment parameters from the scene dataset using an autoencoder, and to combine the efficient representations of scene physical environment parameters with the original scene physical environment parameters to construct composite features of scene physical environment parameters.

[0162] The training module is used to build a classification model using a multilayer perceptron. It takes the composite features of the scene's physical environment parameters and the corresponding scene category as input and trains the classifier under the initial hyperparameter settings to build the classification model.

[0163] The search module is used to set the range of hyperparameters to be optimized, and uses the optimization algorithm to automatically traverse all hyperparameter combinations to obtain the hyperparameter combination that makes the classifier perform best as the hyperparameters of the classification model, thus obtaining the trained classifier.

[0164] In embodiments of the present invention, preprocessing of scene physical environment parameters and corresponding scene category data includes:

[0165] The scene's physical environment parameters and corresponding scene category data are normalized and one-hot encoded.

[0166] In embodiments of the present invention, preprocessing of scene physical environment parameters and corresponding scene category data includes:

[0167] The scene physical environment parameters are processed by normalization. The minimum value is subtracted from the sample value of each type of parameter, and then divided by the range of sample values ​​of that type to obtain the normalized scene physical environment parameters.

[0168] One-hot encoding is used to process scene categories, mapping each scene's true category to a unique binary vector, with one element being 1 and the rest being 0. The category corresponding to the element with the value of 1 is the true category of the scene.

[0169] A scene dataset is constructed using a comma-separated value format, containing normalized physical environment parameters for different scenarios and scene categories with one-hot encoding.

[0170] In an embodiment of the present invention, the processing module is further configured to:

[0171] Define the scene sample data in the scene dataset as X j This includes the preprocessed normalized scene physical environment parameters F. j And the pre-processed one-hot encoded scene category L j ,

[0172] X j =[F j ,L j ], j=1,2,3,…,n

[0173] Where n represents the total number of scene samples;

[0174] Construct an autoencoder and initialize its network parameters. The size of the autoencoder's input layer is related to F. n Dimensions are consistent, encoder output layer is smaller than F n Dimension, the output layer of the decoder and F n Consistent dimensions;

[0175] Compress F using the encoder in the automatic encoder. n To a lower dimension, C n Use a decoder to restore it to R j Calculate R j and F j The difference between them is described using mean squared error, and the network parameters of the autoencoder are iteratively adjusted to minimize this difference.

[0176] C j With F j Combine to construct composite feature S j The resulting formula is:

[0177] S j =[C j ,F j ], j=1,2,3,…,n.

[0178] In an embodiment of the present invention, the training module is further configured to:

[0179] Construct a classification model and initialize its parameters. The input layer size of the classification model is related to the composite feature S. j The model has consistent dimensions, consisting of multiple fully connected layers following the input layer. ReLU activation is used to introduce non-linearity. Each fully connected layer is preceded by a batch normalization layer to normalize the data for each batch. Each fully connected layer is followed by a random deactivation layer, which temporarily ignores certain neurons during iteration with a preset probability. The output layer size of the classification model is consistent with the total number of categories in the scene. Softmax is used as the activation function, ensuring that the predicted output P of the classification model... j Category L of real-world scenarios j They have the same format;

[0180] The classification model is trained with the following hyperparameters: batch size, number of training epochs, optimizer, learning rate, and loss function set to P0. j With L j The cross-entropy between categories.

[0181] In an embodiment of the present invention, the search module is further configured to: set the range of hyperparameters to be optimized based on the training results of the classification model during training;

[0182] Based on the set range of hyperparameters to be optimized, candidate hyperparameter combinations are constructed. A classifier is trained using each set of hyperparameters. The average accuracy of the classifier on the validation set is evaluated through K-fold cross-validation to assess the performance of the classifier under the set of hyperparameters. The best-performing set of all candidate hyperparameter combinations is selected as the optimal hyperparameter combination for the classifier. The classification model is then retrained under the optimal hyperparameter combination.

[0183] It should be noted that the foregoing explanation of the intelligent scene recognition method embodiment applicable to 6G wireless communication also applies to the intelligent scene recognition device applicable to 6G wireless communication in this embodiment, and will not be repeated here.

[0184] The intelligent scene recognition device for 6G wireless communication proposed in this embodiment of the invention collects physical environment parameters of 6G communication scenarios based on electronic maps, trains the device using a dataset processed by normalization, one-hot encoding, and autoencoder, and obtains the optimal hyperparameter combination using a hyperparameter optimization algorithm. Finally, the classifier under the optimal hyperparameter settings achieves high-precision scene recognition accuracy for various 6G communication scenarios, providing a reference for research on communication scene recognition and 6G wireless channels.

[0185] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0186] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0187] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of the invention pertain.

Claims

1. A smart scene recognition method suitable for 6G wireless communication, characterized in that, Includes the following steps: Collect physical environment parameters of the scene to be identified; The physical environment parameters of the scene to be identified are input into a pre-trained classifier for classification, wherein the classifier is obtained by training a classification model using the physical environment parameters of the scene; Before inputting the physical environment parameters of the scene to be identified into a pre-trained classifier for classification, the method further includes: Collect scene physical environment parameters and label the corresponding scene categories. Preprocess the scene physical environment parameters and corresponding scene category data to obtain the scene dataset. An efficient representation of the scene physical environment parameters in the scene dataset is extracted using an autoencoder, and the efficient representation of the scene physical environment parameters is combined with the original scene physical environment parameters to construct a composite feature of the scene physical environment parameters; A classification model constructed using a multilayer perceptron takes the composite features of the scene's physical environment parameters and the corresponding scene category as input, and trains and constructs the classification model under initial hyperparameter settings. Set the range of hyperparameters to be optimized, use the optimization algorithm to automatically traverse all hyperparameter combinations, obtain the hyperparameter combination that makes the classifier perform optimally as the hyperparameters of the classification model, and obtain the trained classifier. An efficient representation of the scene physical environment parameters in the scene dataset is extracted using an autoencoder. This efficient representation is then combined with the original scene physical environment parameters to construct a composite feature of the scene physical environment parameters, including: Define the scene sample data in the scene dataset as This includes preprocessed normalized scene physical environment parameters. and preprocessed one-hot encoded scene categories , in, Indicates the total number of scene samples; Construct an autoencoder and initialize its network parameters. The size of the autoencoder's input layer is... Dimensions are consistent, encoder output layer is smaller than Dimensions, the size of the decoder output layer and Consistent dimensions; Compression using the encoder in the automatic encoder. To a lower dimension Use the decoder to then Restore to ,calculate and The difference between them is described using mean squared error, and the difference is minimized by iteratively adjusting the network parameters of the autoencoder. Will and Combine to construct composite features The resulting formula is: 。 2. The method according to claim 1, characterized in that, Preprocessing of scene physical environment parameters and corresponding scene category data includes: The scene's physical environment parameters and corresponding scene category data are normalized and one-hot encoded.

3. The method according to claim 2, characterized in that, Preprocessing of scene physical environment parameters and corresponding scene category data includes: The scene physical environment parameters are processed using normalization. The minimum value is subtracted from the sample value of each type of parameter, and then divided by the range of sample values ​​of that type to obtain the normalized scene physical environment parameters. One-hot encoding is used to process scene categories, mapping each scene's true category to a unique binary vector, with one element being 1 and the rest being 0. The category corresponding to the element with the value of 1 is the true category of the scene. A scene dataset is constructed using a comma-separated value format, containing normalized physical environment parameters for different scenarios and scene categories with one-hot encoding.

4. The method according to claim 1, characterized in that, A classification model constructed using a multilayer perceptron takes the composite features of the scene's physical environment parameters and the corresponding scene category as input, and trains and constructs the classification model under initial hyperparameter settings, including: Build a classification model and initialize its parameters, including the input layer size and composite features. The model has consistent dimensions, consisting of multiple fully connected layers following the input layer. ReLU activation is used to introduce non-linearity. Each fully connected layer is preceded by a batch normalization layer to normalize the data for each batch. Each fully connected layer is followed by a random deactivation layer, which temporarily ignores certain neurons with a predetermined probability during iteration. The output layer size of the classification model is consistent with the total number of categories in the scene. Softmax is used as the activation function to optimize the predicted output of the classification model. Real-world scenario categories They have the same format; The classification model is trained with the following hyperparameters: batch size, number of training epochs, optimizer, and learning rate. The loss function is set to... and The cross-entropy between categories.

5. The method according to claim 1, characterized in that, Set the range of hyperparameters to be optimized, use an optimization algorithm to automatically traverse all hyperparameter combinations, and obtain the hyperparameter combination that makes the classifier perform optimally as the hyperparameters of the classification model, thus obtaining a trained classifier, including: Based on the training results of the classification model during training, set the range of hyperparameters to be optimized; Based on the set range of hyperparameters to be optimized, candidate hyperparameter combinations are constructed. A classifier is trained using each set of hyperparameters. The average accuracy of the classifier on the validation set is evaluated through K-fold cross-validation to assess the performance of the classifier under the set of hyperparameters. The best-performing set of all candidate hyperparameter combinations is selected as the optimal hyperparameter combination for the classifier. The classification model is then retrained under the optimal hyperparameter combination.

6. A smart scene recognition device suitable for 6G wireless communication, used in the smart scene recognition method for 6G wireless communication as described in any one of claims 1-5, characterized in that, include: The data acquisition module is used to collect physical environment parameters of the scene to be identified. The recognition module is used to input the physical environment parameters of the scene to be recognized into a pre-trained classifier for classification, wherein the classifier is obtained by training a classification model using the physical environment parameters of the scene; The device further includes: The module is used to collect scene physical environment parameters and label the corresponding scene categories. The scene physical environment parameters and corresponding scene category data are preprocessed to obtain the scene dataset. The processing module is used to extract an efficient representation of the scene physical environment parameters in the scene dataset using an autoencoder, and to combine the efficient representation of the scene physical environment parameters with the original scene physical environment parameters to construct a composite feature of the scene physical environment parameters. The training module is used to train and construct a classification model using a multilayer perceptron, taking the composite features of the scene's physical environment parameters and the corresponding scene category as input, under the initial hyperparameter settings. The search module is used to set the range of hyperparameters to be optimized, use the optimization algorithm to automatically traverse all hyperparameter combinations, obtain the hyperparameter combination that makes the classifier perform optimally as the hyperparameters of the classification model, and obtain the trained classifier. The processing module is further used for: An efficient representation of the scene physical environment parameters in the scene dataset is extracted using an autoencoder. This efficient representation is then combined with the original scene physical environment parameters to construct a composite feature of the scene physical environment parameters, including: Define the scene sample data in the scene dataset as This includes preprocessed normalized scene physical environment parameters. and preprocessed one-hot encoded scene categories , in, Indicates the total number of scene samples; Construct an autoencoder and initialize its network parameters. The size of the autoencoder's input layer is... Dimensions are consistent, encoder output layer is smaller than Dimensions, the size of the decoder output layer and Consistent dimensions; Compression using the encoder in the automatic encoder. To a lower dimension Use the decoder to then Restore to ,calculate and The difference between them is described using mean squared error, and the difference is minimized by iteratively adjusting the network parameters of the autoencoder. Will and Combine to construct composite features The resulting formula is: 。 7. The apparatus according to claim 6, characterized in that, Preprocessing of scene physical environment parameters and corresponding scene category data includes: The scene's physical environment parameters and corresponding scene category data are normalized and one-hot encoded.

Images