6G network time delay prediction method based on multi-scale contrast learning
By adopting a multi-scale contrast learning method in the delay prediction of 6G networks, combined with multi-scale random cropping, timestamp masking and expansion convolution technology, the problem that existing methods are difficult to capture nonlinear and highly dynamic features is solved, and higher prediction accuracy and robustness are achieved.
Patent Information
- Application Number
- CN202510204618.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-24
- Publication Date
- 2025-05-27
AI Technical Summary
Existing 6G network delay prediction methods are difficult to effectively capture nonlinear and highly dynamic features, and relying on negative samples increases training complexity and may reduce robustness.
Using a multi-scale contrast learning method, combining multi-scale random cropping, multi-scale timestamp masking and multi-scale expansion convolution technology, multi-scale multi-scale features of 6G network delays are extracted, and prediction is made through improved time series hybrid prediction methods.
It improves the accuracy and robustness of the delay prediction of 6G network, reduces the training complexity, and enhances the generalization ability of the model.
Smart Images
Figure CN120050193A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of network delay prediction, and particularly relates to a 6G network delay prediction method based on multi-scale contrast learning. Background Art
[0002] In recent years, 6G network delay prediction has become a popular field in time series prediction, mainly divided into traditional methods and deep learning methods. Traditional methods usually assume that the data is linear and it is difficult to capture long-term dependencies, resulting in poor performance when dealing with non-linear and highly dynamic 6G network delay data. Deep learning methods capture the non-linear relationships and long-term dependencies of time series through neural networks, automatically extract features, and reduce the need for data preprocessing, showing significant advantages in dealing with non-linear and highly dynamic 6G delays. In particular, contrast learning performs well in 6G network delay prediction. Contrast learning extracts effective features by learning the similarities and differences between samples, significantly improving the quality and generalization ability of feature representation. However, most contrast learning methods rely on negative samples, increasing the training complexity and potentially reducing the robustness when dealing with noisy data. Summary of the Invention
[0003] The purpose of the present invention is to combine an improved general time series representation technique and an optimized activation function time series hybrid prediction method based on multi-scale contrast learning and apply it to the method of 6G network delay prediction. The improved time series representation method introduces multi-scale random cropping, multi-scale timestamp masking, and multi-scale dilated convolution techniques.
[0004] The present invention provides a 6G network delay prediction method based on multi-scale contrast learning, including the following steps:
[0005] Step 1: Obtain the original 6G network delay data set;
[0006] Step 2: Data preprocessing; introduce a multi-scale random cropping method and a multi-scale timestamp masking method to obtain overlapping multi-scale delay sequence sample pairs Y masked ;
[0007] Step 3: Feature extraction stage; pass the multi-scale 6G network delay sequence through the feature extraction stage based on multi-scale dilated convolution to obtain the local features and global features of the 6G network delay data;
[0008] Step 4: Delay prediction stage; pass the 6G network delay sequence with local features and global features extracted through the delay prediction stage based on the improved mixer to obtain the time pattern and features of the 6G network delay; use the obtained time pattern and features of the 6G network delay and the to-be-tested 6G network delay data set for multi-scale 6G network delay prediction to predict the network delay.
[0009] Further, in step 2, first, the multi-scale random cropping method is used to randomly cut the data for 6G network latency prediction to obtain sample pairs of overlapping sub-time series; then, the sample pairs of sub-time series are input into the fully connected layer; finally, the multi-scale timestamp masking method is used to mask some timestamps to obtain the enhanced multi-scale network latency sequence sample pair Y masked .
[0010] Further, in step 3, first, the residual connection method is respectively adopted for the multi-scale dilated convolutions with different dilation coefficients to obtain the time series X that extracts multi-scale features res ; then, the time series X that extracts multi-scale features with different dilation coefficients res are concatenated in the channel dimension to capture the features of 6G network latency data from multiple levels, and then the maximum pooling operation is applied to the concatenated output to reduce the size of the feature map; finally, the pooled output is flattened into a two-dimensional tensor and input into the fully connected layer.
[0011] Further, the residual connection method includes:
[0012] Step 3.1: Apply the multi-scale network latency sample pair Y masked after step 2 to the standard one-dimensional convolution to obtain X conv1d ;
[0013] X conv1d = Conv1D(Y masked )
[0014] Step 3.2: Apply X conv1d to the activation function Gaussian error linear unit to obtain X gelu1 , and then adopt the same padding convolution method to obtain the padded one-dimensional tensor X padding ;
[0015] X gelu = GELU(X conv1d )
[0016] Step 3.3: Repeat step 3.2, apply X padding to the activation function Gaussian error linear unit and obtain the corresponding output X gelu2 , and then adopt the same padding convolution method to obtain the padded one-dimensional tensor X padding2 ;
[0017] X gelu2 = GELU(X padding )
[0018] Step 3.4: Add the X conv1d obtained in step 3.1 to the X paddin2Perform residual connection to obtain the time series X that has extracted multi-scale features res ;
[0019] X res = X conv1d + X padding2 。
[0020] Furthermore, the padding convolution layer in the same padding convolution method is defined according to the convolution kernel size, padding size, and dilation rate, and specifically includes the following steps:
[0021] Step 3.2.1: Calculate the size of the receptive field;
[0022] receptive_field = (kernel_size - 1) * dilation + 1
[0023] In the formula, kernel_size is the convolution kernel size, and dilation is the dilation coefficient;
[0024] Step 3.2.2: Calculate the padding size according to the receptive field, and the padding size should be half of the receptive field;
[0025]
[0026] Step 3.2.3: Judge whether it is necessary to remove the redundant padding value according to the size of the receptive field. If the receptive field is even, one position needs to be removed;
[0027]
[0028] Furthermore, the specific steps of step 4 include:
[0029] Step 4.1: Use a time-mixed multi-layer perceptron to extract the time features of the multi-scale 6G network delay sample pairs;
[0030] Step 4.2: Use the feature multi-layer perceptron method to extract the time pattern for the sample pairs processed in step 4.1;
[0031] Step 4.3: Use the time projection method to apply the time patterns and features of the multi-scale 6G network delay obtained in steps 4.1 and 4.2 to the fully connected layer to project them to the predicted delay length required in the 6G network delay prediction task.
[0032] Furthermore, the specific steps of step 4.1 include:
[0033] Step 4.1.1: Take the sample pairs of the multi-scale 6G network delay features output in step 2 and complete the multi-scale 6G network delay normalization operation through a two-dimensional batch normalization layer;
[0034] X normalized = BatchNorm2D(X padded )
[0035] Step 4.1.2: Transpose the output of Step 4.1.1 and map the multi-scale 6G network latency to a high-dimensional latent vector through a fully connected layer;
[0036] X latent = FC(Transpose(X normalized ))
[0037] Step 4.1.3: Pass the output of Step 4.1.2 through an improved activation function, the parametric rectified linear unit, to complete the non-linear transformation of the multi-scale 6G network latency;
[0038] X activated = PReLU(X latent )
[0039] Step 4.1.4: Pass the output of Step 3.1.4 through a dropout layer to set the output of neurons to zero during the training process;
[0040] X dropped = Dropout(X activated )
[0041] Step 4.1.5: Perform a residual connection between the output of Step 4.1.4 and the multi-scale 6G network latency processed in the feature extraction stage based on multi-scale dilated convolution in Step 3 to complete the capture of the time pattern in the multi-scale 6G network latency;
[0042] X residual = X dropped + X padded .
[0043] Furthermore, Step 4.2 specifically includes:
[0044] Step 4.2.1: Transpose the multi-scale 6G network latency processed by the time mixing multi-layer perceptron method output in Step 4.1 and complete normalization through a two-dimensional batch normalization layer;
[0045] X normalized_2 = BatchNorm2D(Transpose(X residual ))
[0046] Step 4.2.2: Map the multi-scale 6G network latency processed in Step 4.2.1 to a high-dimensional latent vector through a fully connected layer;
[0047] X latent_2 = FC(X normalized_2 )
[0048] Step 4.2.3: Process the multi-scale 6G network latency processed in Step 4.2.2 through a parameterized rectified linear unit with an improved activation function to adjust its own non-linearity degree;
[0049] X activated_2 = PReLU(X latent_2 )
[0050] Step 4.2.4: Process the multi-scale 6G network latency processed in Step 4.2.3 through a dropout layer, a fully connected layer, and a dropout layer in sequence;
[0051] X dropped_2 = Dropout(X activated_2 )
[0052] X latent_3 = FC(X dropped_2 )
[0053] X dropped_3 = Dropout(X latent_3 )
[0054] Step 4.2.5: Perform a residual connection on the multi-scale 6G network latency processed by the intermixing multi-layer perceptron method in Step 4.1.5 and the multi-scale 6G network latency processed by the feature mixing multi-layer perceptron method in Step 4.2.4 to obtain the time pattern and features of the multi-scale 6G network latency;
[0055] X final = X residual + X dropped_3 。
[0056] The beneficial effects of the present invention are as follows:
[0057] 1) Introduce a multi-scale random cropping method and a multi-scale timestamp masking method in the data preprocessing stage to obtain overlapping multi-scale sub-latency sequences, reduce the complexity of the training process of the present invention, and enhance the robustness of the present invention; 2) Introduce a multi-scale dilated convolution method in the feature extraction stage based on multi-scale dilated convolution to obtain local and global features of 6G network latency data; 3) Introduce an improved time series hybrid prediction method in the improved general time series representation method, enhance the ability of the present invention to capture complex temporal features and time patterns, and improve the accuracy of 6G network latency prediction. Description of the Drawings
[0058] Figure 1 It is a flowchart of the 6G network latency prediction method based on multi-scale contrast learning of the present invention;
[0059] Figure 2This is the flowchart of the data preprocessing stage of the present invention;
[0060] Figure 3 This is the flowchart of the feature extraction stage based on multi-scale dilated convolution of the present invention;
[0061] Figure 4 This is the flowchart of the delay prediction stage based on the improved mixer. Detailed implementation manners
[0062] The present invention combines multi-scale contrast learning, an improved general time series representation technique, and an optimized activation function time series hybrid prediction method, and applies it to the method of 6G network delay prediction. The improved time series representation method introduces multi-scale random cropping, multi-scale timestamp masking, and multi-scale dilated convolution techniques. The overall process of the 6G network delay prediction of the present invention is as follows: First, read the original 6G network delay data set; then call the method in the data preprocessing stage to process the original 6G network delay data set; secondly, pass the multi-scale 6G network delay sample pairs through the feature extraction stage based on multi-scale dilated convolution, and capture the features of 6G network delay data from multiple levels through dilated convolutions with dilation coefficients of 1, 3, and 5 respectively, and increase the continuity of 6G network delay data through a one-dimensional convolutional layer to obtain global features from a global perspective; then pass the 6G network delay with local and global features through the delay prediction stage based on the improved mixer to obtain the time pattern and features of the 6G network delay; finally, use the extracted time pattern and features for the 6G network delay prediction of the test set in the multi-scale 6G network delay data set.
[0063] The present invention is mainly divided into three stages: (1) data preprocessing stage; (2) feature extraction stage based on multi-scale dilated convolution; (3) delay prediction stage based on the improved mixer.
[0064] In the above-mentioned (1) data preprocessing stage, first, the data for 6G network delay prediction is randomly cut through the multi-scale random cropping method to obtain subsequences of multi-scale 6G network delay as positive sample pairs for feature and time pattern extraction in the subsequent (2) and (3) stages; secondly, use the input projection layer to map the positive sample pairs to high-dimensional latent vectors through a fully connected layer to ensure dimension matching with the (2) and (3) stages; finally, use the multi-scale timestamp masking method to randomly block some timestamps to generate enhanced context views and improve the robustness of 6G network delay prediction.
[0065] In the above-mentioned (1) data preprocessing stage, a multi-scale random cropping method is combined with a multi-scale timestamp masking method for data preprocessing, which has the following characteristics: Compared with traditional contrastive learning methods that rely on a large number of negative samples and require complex sampling strategies to ensure the diversity and effectiveness of negative samples, the multi-scale random cropping method introduces overlapping multi-scale 6G sub-delay sequences as sample pairs, making it easier for the model to learn the internal structure of 6G network delays; compared with the random cropping method that only crops at the original scale, the multi-scale random cropping can crop at multiple scales, thereby capturing more relevant patterns of 6G network delays and improving the generalization ability of the model. The multi-scale timestamp masking method can generate different masking ratios according to the multi-scale 6G network delays compared with the timestamp masking method, thus achieving a better masking effect on the multi-scale 6G network delays. The specific process of the (1) data preprocessing stage of the present invention is as follows:
[0066] (1.1) Use the multi-scale random cropping method to generate positive sample pairs for subsequent multi-scale 6G network delay prediction;
[0067] (1.2) Apply the positive sample pairs of multi-scale 6G network delay prediction to the fully connected layer and obtain the corresponding output;
[0068] (1.3) Use the multi-scale timestamp masking method to mask part of the timestamps on the output of (1.2) to generate enhanced context views.
[0069] The specific process of the multi-scale random cropping method in the above-mentioned (1.1) is as follows:
[0070] (1.1.1) Obtain the original 6G network delay length length;
[0071] (1.1.2) Initialize an empty list to store the cropped subsequences;
[0072] (1.1.3) For each 6G network delay to be cropped, randomly select a ratio and calculate the length crop_length of the overlapping sub-delay sequence to be cropped;
[0073] (1.1.4) Randomly select the starting position start of the overlapping sub-delay sequence, where the size of start is a number between 0 and N 1 inclusive;
[0074] N 1 = length - crop_length + 1
[0075] (1.1.5) Calculate the ending position end of the overlapping sub-delay sequence;
[0076] end = start + crop_length
[0077] (1.1.6) Randomly select the left boundary of the 6G network latency in the positive sample pair as crop_left, where the size of crop_left is a number between 0 and start;
[0078] (1.1.7) Randomly select the right boundary of the 6G network latency in the positive sample pair as crop_right, where the size of crop_right is a number between end and length;
[0079] (1.1.8) Crop the positive sample pair with overlapping sub-latency sequences from the original 6G network latency;
[0080] (1.1.9) Add the cropped positive sample pairs to the list;
[0081] (1.1.10) Return all the cropped positive sample pairs for subsequent 6G network latency prediction.
[0082] In the above (1.3), the multi-scale timestamp masking method is used to partially mask the timestamps of the output of (1.2) to generate an enhanced 6G network latency context view.
[0083] The specific process of the multi-scale timestamp masking method in (1.3) is as follows:
[0084] (1.3.1) Obtain the multi-scale 6G network latency sample pairs processed by (1.2) and determine whether there are null values Na in the sample pairs. If so, set the Na values to 0 and obtain the corresponding output;
[0085] (1.3.2) Initialize the parameters related to the multi-scale timestamp masking method: the minimum masking ratio mask_ratio_min and the maximum masking ratio mask_ratio_max;
[0086] (1.3.3) Initialize an empty list to store the masked 6G network latency sample pairs;
[0087] (1.3.4) For each 6G network latency sample pair, randomly select a masking ratio mask_ratio between mask_ratio_min and mask_ratio_max;
[0088] (1.3.5) Calculate the number of masked timestamps num_masks, where l represents the length of the multi-scale 6G network latency sample processed by (1.2):
[0089] num_masks = l * mask_ratio
[0090] (1.3.6) Randomly select num_masks timestamp indices to generate a list of masked timestamp indices mask_indices;
[0091] (1.3.7) For each timestamp index i in mask_indices, set the i-th timestamp of the multi-scale 6G network delay sample pair processed in (1.2) to 0;
[0092] (1.3.8) Add the masked multi-scale 6G network delay sample pairs to the list, and apply all the sample pairs in the list to subsequent 6G network delay prediction.
[0093] (2) In the feature extraction stage based on multi-scale dilated convolution, obtain the multi-scale 6G network delay sample pairs with some timestamps masked processed by the multi-scale random cropping method and the multi-scale timestamp masking method in the (1) data preprocessing stage, and effectively capture the features of the multi-scale 6G network delay sample pairs from multiple levels by calling the corresponding methods in the (2) feature extraction stage based on multi-scale dilated convolution, and increase the data continuity through a one-dimensional convolutional layer to obtain global features from a global perspective, and combine these features for subsequent 6G network delay prediction. Among them, the residual connection method is that there is a residual connection inside each convolutional block, and then there are two applications of the same-padding convolution method for each convolutional block. The same-padding convolution is to keep the size of the one-dimensional convolution unchanged. The (2) feature extraction stage based on multi-scale dilated convolution includes three main methods: (2.1) multi-scale dilated convolution encoder method, (2.2) residual connection method, (2.3) same-padding convolution method.
[0094] (2.1) The specific process of the multi-scale dilated convolution encoder method is as follows:
[0095] (2.1.1) Initialize 1 standard one-dimensional convolutional layer to increase the continuity of the 6G network delay sample pairs and obtain their global features from a global perspective.
[0096] (2.1.2) Initialize three multi-scale dilated convolutions with dilation coefficients of 1, 3, and 5 respectively to capture the features of the 6G network delay data from multiple levels.
[0097] (2.1.3) Initialize a fully connected layer, where the input dimension of the fully connected layer is the number of output channels of the multi-scale dilated convolution * the number of convolutional kernels * the number of dilation rates, and the output dimension is 1.
[0098] (2.1.4) Initialize an empty list conv_outputs to store the outputs of different multi-scale dilated convolutions.
[0099] (2.1.5) For the three multi-scale dilated convolutions with dilation factors of 1, 3, and 5 respectively, call the (2.2) residual connection method, and add the output results to the conv_outputs list.
[0100] (2.1.6) For the three multi-scale dilated convolutions with dilation factors of 1, 3, and 5 respectively, and the time series X of multi-scale features has been extracted through step (2.2.6) res Concatenate them along the channel dimension to capture the features of 6G network latency data from multiple levels, and then apply a max pooling operation to the concatenated output to reduce the size of the feature map. Finally, flatten the pooled output into a two-dimensional tensor for input to the fully connected layer.
[0101] (2.1.7) Input the flattened tensor into the fully connected layer to obtain the final output.
[0102] The specific process of adopting the (2.2) residual connection method in the above step (2.1.5) is as follows:
[0103] (2.2.1) Obtain the multi-scale 6G network latency sample pair Y with some timestamps masked, which has been processed by the multi-scale random cropping method and the multi-scale timestamp masking method in the (1) data preprocessing stage, and apply it to the standard one-dimensional convolution. masked , and apply it to the standard one-dimensional convolution.
[0104] X conv1d = Conv1D(Y masked )
[0105] (2.2.2) Obtain the output of (2.2.1) and apply it to the activation function Gaussian error linear unit and obtain the corresponding output.
[0106] X gelu1 = GELU(X conv1d )
[0107] (2.2.3) Apply the output obtained in (2.2.2) to the (2.3) same padding convolution method to obtain a one-dimensional tensor X padded to the specified size padding .
[0108] (2.2.4) Apply the activation function Gaussian error linear unit to the output obtained in (2.2.3) and obtain the corresponding output.
[0109] X gelu2 = GELU(X padding )
[0110] (2.2.5) Apply the output obtained in (2.2.4) to the (2.3) same padding convolution method and obtain a one-dimensional tensor X padded to the specified size padding2 ;
[0111] (2.2.6) Perform a residual connection on the output obtained in (2.2.5) and the output obtained in (2.2.1) to obtain the time series X that has extracted multi-scale features res .
[0112] X res = X conv1d + X padding2
[0113] The specific processes of the above steps (2.2.3) and (2.2.5) using the same padding convolution method in the feature extraction stage based on multi-scale dilated convolution are as follows:
[0114] (2.3.1) Calculate the size of the receptive field
[0115] receptive_field = (kernel_size - 1) * dilation + 1
[0116] In the formula, kernel_size is the size of the convolution kernel, which is defaulted to 3 in the present invention; dilation is the dilation coefficient, which are 1, 3, and 5 respectively in the present invention
[0117] (2.3.2) Calculate the size of the padding according to the receptive field, and the size of the padding should be half of the receptive field
[0118]
[0119] (2.3.3) Judge whether it is necessary to remove the redundant padding value according to the size of the receptive field. If the receptive field is even, one position needs to be removed
[0120]
[0121] (2.3.4) Define a padding convolution layer according to the convolution kernel size, padding size, dilation rate, etc., and apply it to the specified multi-scale 6G network delay sample pair
[0122] The above (3) Based on the improved mixer delay prediction stage obtains the 6G network delay multi-scale sample pairs processed by the (2) Based on multi-scale dilated convolution feature extraction stage, and captures the time pattern of the 6G network delay by calling the corresponding methods in the (3) Based on the improved mixer delay prediction stage. The (3) Based on the improved mixer delay prediction stage mainly includes three methods: (3.1) Temporal Mixing Multilayer Perceptron method, (3.2) Feature Mixing Multilayer Perceptron method, (3.3) Temporal Projection method
[0123] (3) The specific process of the (3.1) Time Mixing Multilayer Perceptron method in the delay prediction stage based on the improved mixer is as follows:
[0124] (3.1.1) Obtain sample pairs of multi-scale 6G network delay features processed by the (2) feature extraction stage based on multi-scale dilated convolution, and complete the multi-scale 6G network delay normalization operation through a two-dimensional batch normalization layer to obtain the corresponding output.
[0125] X normalized = BatchNorm2D(X padded )
[0126] (3.1.2) Transpose the output of (3.1.1) and pass it through a fully connected layer, that is, map the multi-scale 6G network delay to a high-dimensional latent vector and obtain the corresponding output.
[0127] X latent = FC(Transpose(X normalized ))
[0128] (3.1.3) Pass the output of (3.1.2) through an improved activation function - the parametric rectified linear unit to complete the non-linear transformation of the multi-scale 6G network delay. Because compared with the rectified linear unit, the activation function used in the original time series mixing prediction method, the parametric rectified linear unit is more suitable for the convolution operation in the (2) feature extraction stage based on multi-scale dilated convolution, and can ensure that the present invention can adjust its non-linear degree during the training process, so as to better adapt to the different data distributions of different 6G network delay data sets.
[0129] X activated = PReLU(X latent )
[0130] (3.1.4) Pass the output of (3.1.3) through a dropout layer, that is, set the outputs of some neurons during the training process to zero, so as to reduce overfitting in the training process of the present invention, thereby improving the generalization ability of the present invention.
[0131] X dropped = Dropout(X activated )
[0132] (3.1.5) Perform a residual connection between the output of (3.1.4) and the multi-scale 6G network delay processed by the (2) feature extraction stage based on multi-scale dilated convolution, so as to complete the capture of the time pattern in the multi-scale 6G network delay.
[0133] X residual = X dropped + X padded
[0134] (3) The specific process of the (3.2) Feature Mixing Multilayer Perceptron method in the delay prediction stage based on the improved mixer is as follows:
[0135] (3.2.1) Transpose the multi-scale 6G network delay processed by the (3.1) Temporal Mixing Multilayer Perceptron method in the delay prediction stage based on the improved mixer and complete normalization through a two-dimensional batch normalization layer;
[0136] X normalized_2 = RatchNorm2D(Transpose(X residual ))
[0137] (3.2.2) Map the multi-scale 6G network delay processed in (3.2.1) to a high-dimensional latent vector through a fully connected layer;
[0138] X latent_2 = FC(X normalized_2 )
[0139] (3.2.3) Process the multi-scale 6G network delay processed in (3.2.2) through an improved activation function - parametric rectified linear unit to ensure that the present invention adaptively adjusts its own non-linearity during the training process, so as to better use different data distributions of different 6G network delay datasets;
[0140] X activated_2 = PReLU(X latent_2 )
[0141] (3.2.4) Process the multi-scale 6G network delay processed in (3.2.3) through a dropout layer, that is, set the outputs of some neurons during the training process to zero to reduce overfitting during the training process of the present invention, thereby improving the generalization ability of the present invention;
[0142] X dropped_2 = Dropout(Xactivalted _2 )
[0143] (3.2.5) Remap the multi-scale 6G network delay processed in (3.2.4) to a high-dimensional latent vector through a fully connected layer;
[0144] X latent_3 = FC(X dropped_2 )
[0145] (3.2.6) Process the multi-scale 6G network delay processed in (3.2.5) through a dropout layer, that is, set the outputs of some neurons during the training process to zero to improve the generalization ability of the present invention;
[0146] Xdropped_3 = Dropout(X latent_3 )
[0147] (3.2.7) Residually connect the multi-scale 6G network latency processed by (3) the time mixing multi-layer perceptron method in the latency prediction stage based on the improved mixer with the multi-scale 6G network latency processed by (3) the feature mixing multi-layer perceptron method in the latency prediction stage based on the improved mixer, so as to obtain the time pattern and features of the multi-scale 6G network latency.
[0148] X final = X residual + X dropped_3
[0149] In (3) the time projection method in the latency prediction stage based on the improved mixer, it obtains the multi-scale 6G network latency processed by (3) the feature mixing multi-layer perceptron method in the latency prediction stage based on the improved mixer, and applies a fully connected layer to project this 6G network latency to the predicted latency length required in the 6G network latency prediction task.
[0150] Finally, the present invention performs latency prediction on the test set in the 6G network latency dataset under the network latency prediction structure according to the time pattern and features of the multi-scale 6G network latency obtained in (1)(2)(3) stages.
[0151] In summary, to meet the requirements of 6G network latency prediction, the present invention proposes a method based on multi-scale contrast learning, which combines an improved method for general time series representation and an improved time series hybrid prediction method to better capture the complex and dynamic temporal features in the 6G network, improving the prediction accuracy and robustness. The method has the following characteristics: 1) In the data preprocessing stage, a multi-scale random cropping method and a multi-scale timestamp masking method are introduced to obtain overlapping multi-scale sub-latency sequences, reducing the complexity of the training process of the present invention and enhancing the robustness of the present invention; 2) In the feature extraction stage based on multi-scale dilated convolution, dilated convolutions with dilation coefficients of 1, 3, and 5 are introduced to capture the features of 6G network latency data from multiple levels, and the continuity of 6G network latency data is increased through a one-dimensional convolutional layer to obtain global features from a global perspective; 3) In the method for general time series representation, the above data preprocessing method, multi-scale dilated convolution method, and time series hybrid prediction method with an improved activation function are introduced, enhancing the ability of the present invention to capture complex temporal features and time patterns and improving the accuracy of 6G network latency prediction.
[0152] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A 6G network delay prediction method based on multi-scale contrastive learning, characterized in that: The following steps are involved: Step 1: Obtain the original 6G network delay dataset; Step 2: Data preprocessing: Introduce multi-scale random cropping method and multi-scale timestamp masking method to obtain overlapping multi-scale time delay sequence sample pairs Y masked ; Step 3: Feature extraction stage; The multi-scale 6G network delay sequence is subjected to the multi-scale dilated convolution feature extraction stage to obtain the local and global features of the 6G network delay data; Step 4: Delay prediction stage; the 6G network delay data with local and global features extracted will be subjected to the delay prediction stage based on the improved mixer to obtain the time pattern and characteristics of the 6G network delay; the time pattern and characteristics of the 6G network delay obtained and the 6G network delay data set to be tested will be used in multi-scale 6G network delay prediction to predict the network delay.
2. The 6G network delay prediction method based on multi-scale contrastive learning according to claim 1 is characterized in that: In step 2, first, the multi-scale random cropping method is used to randomly cut the data used for 6G network delay prediction to obtain sample pairs of overlapping sub-time series; then the sample pairs of sub-time series are input into the fully connected layer; finally, the multi-scale timestamp masking method is used to mask part of the timestamps to obtain the enhanced multi-scale network delay sequence sample pairs Y masked .
3. The 6G network delay prediction method based on multi-scale contrastive learning according to claim 1 is characterized in that: In step 3, firstly, a residual connection method is used for multi-scale dilated convolutions with different expansion coefficients to obtain a time series X from which multi-scale features are extracted. res ; Then extract the multi-scale features of the time series X with different expansion coefficients res Splicing is performed on the channel dimension to capture the features of 6G network delay data from multiple levels, and then the maximum pooling operation is applied to the spliced output to reduce the size of the feature map; finally, the pooled output is flattened into a two-dimensional tensor and input into the fully connected layer.
4. The 6G network delay prediction method based on multi-scale contrastive learning according to claim 3 is characterized in that: The residual connection method comprises: Step 3.1: The multi-scale network delay samples Y masked Applying the standard one-dimensional convolution yields X conv1d ; X conv1d =Conv1D(Y masked ) Step 3.2: X conv1d Applying the activation function Gaussian error linear unit to get X gelu1 , and then use the same padding convolution method to obtain a one-dimensional tensor X padding ; X gelu =GELU(X conv1d ) Step 3.3: Repeat step 3.2 and change X padding Apply the activation function Gaussian error linear unit and obtain the corresponding output X gelu2 Then, the same padding convolution method is used to obtain a one-dimensional tensor X padding2 ; X gelu2 =GELU(X padding ) Step 3.4: Convert the X obtained in step 3.1 to conv1d The X obtained in step 3.3 paddin2 Perform residual connection to obtain the time series X from which multi-scale features have been extracted res ; X res =X conv1d +X padding2 。 5. The 6G network delay prediction method based on multi-scale contrastive learning according to claim 4 is characterized in that: The padded convolution layer in the same padded convolution method is defined according to the convolution kernel size, the padding size, and the dilation rate, and specifically includes the following steps: Step 3.2.1: Calculate the size of the receptive field; receptive_field=(kernel_size-1)*dilation+1 In the formula, kernel_size is the convolution kernel size, and dilation is the expansion coefficient; Step 3.2.2: Calculate the padding size based on the receiving field. The padding size should be half of the receiving field. Step 3.2.3: Determine whether to remove extra padding values based on the size of the receiving domain. If the receiving domain is an even number, remove one position.
6. The 6G network delay prediction method based on multi-scale contrastive learning according to claim 1 is characterized in that: The step 4 specifically includes: Step 4.1: Use time-mixed multi-layer perception to extract the time features of multi-scale 6G network delay sample pairs; Step 4.2: Extract temporal patterns from the sample pairs processed in step 4.1 using a feature multi-layer perception method; Step 4.3: Use the time projection method to apply the time pattern and characteristics of the multi-scale 6G network delay obtained in steps 4.1 and 4.2 to the fully connected layer to project it to the predicted delay length required in the 6G network delay prediction task.
7. The 6G network delay prediction method based on multi-scale contrastive learning according to claim 6 is characterized in that: The step 4.1 specifically includes: Step 4.1.1: Take the sample pairs of multi-scale 6G network delay features output in step 2 and complete the multi-scale 6G network delay normalization operation through a two-dimensional batch normalization layer; X normalized =BatchNorm2D(X padded ) Step 4.1.2: Transpose the output of step 4.1.1 and map the multi-scale 6G network delay to a high-dimensional latent vector through a fully connected layer; X latent =FC(Transpose(X normalized )) Step 4.1.3: The output of step 4.1.2 is parameterized into a linear rectifier unit through an improved activation function to complete the nonlinear transformation of the multi-scale 6G network delay; X activated !!PRELU(X latent ) Step 4.1.4: Pass the output of step 3.1.4 through a random dropout layer to set the output of the neurons in the training process to zero; X dropped =Dropout(X activated ) Step 4.1.5: Perform a residual connection between the output of step 4.1.4 and the multi-scale 6G network delay processed by the feature extraction stage based on multi-scale dilated convolution in step 3 to complete the capture of the time pattern in the multi-scale 6G network delay; X residual =X dropped +X padded 。 8. The 6G network delay prediction method based on multi-scale contrastive learning according to claim 7 is characterized in that: The step 4.2 specifically includes: Step 4.2.1: Transpose the multi-scale 6G network delay processed by the time-mixed multi-layer perception method output in step 4.1 and normalize it through a two-dimensional batch normalization layer; X normalized_2 =BatchNorm2D(Transpose(X residual )) Step 4.2.2: Map the multi-scale 6G network delay processed in step 4.2.1 to a high-dimensional latent vector through a fully connected layer; X latent_2 =FC(X normalized_2 ) Step 4.2.3: Process the multi-scale 6G network delay processed in step 4.2.2 by parameterizing the linear rectifier unit with an improved activation function, and adjust its own nonlinearity; X activated_2 !!PRELU(X latent_2 ) Step 4.2.4: The multi-scale 6G network delay processed in step 4.2.3 is processed through a random dropout layer, a fully connected layer, and a random dropout layer in sequence; X dropped_2 =Dropout(X activated_2 ) X latent_3 =FC(X dropped_2 ) X dropped_3 =Dropout(X latent_3 ) Step 4.2.5: Perform residual connection on the multi-scale 6G network delay processed by the inter-hybrid multi-layer perception method in step 4.1.5 and the multi-scale 6G network delay processed by the feature hybrid multi-layer perception method in step 4.2.4 to obtain the time pattern and characteristics of the multi-scale 6G network delay; X final =X residual +X dropped_3 。