Self-supervised based source domain and target domain joint training based overlay manifold estimation method

Abstract

The application discloses a kind of based on source domain and target domain joint training of self-supervision's covering manifold estimation method, comprising: the base station position sample corresponding to each region, topographic information sample and covering manifold sample are obtained to the specified region division;Source domain and target domain sample are sequentially subjected to data enhancement processing and dynamic mask processing;The neural network prediction model framework consisting of feature coding fusion module and covering manifold generation module is constructed, for sample input is executed progressive mask generation strategy after, until generating full mask, finally obtain covering manifold prediction result;Based on prediction result, calculate its reconstruction error with real covering manifold in mask area, and update model network parameters by back propagation.The application has reconstruction ability to target domain feature and generation ability based on source domain information simultaneously in training process, and finally model only needs to input source domain information to generate covering manifold prediction result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network optimization, in particular to a source domain and target domain joint training coverage manifold estimation method based on self-supervised learning, which is suitable for coverage manifold estimation of cellular communication networks and wireless channel spatial feature estimation scenarios. BACKGROUND

[0002] With the continuous expansion of the scale of cellular mobile communication networks and the continuous refinement of cell structure, the spatial distribution of coverage performance has become a core reference index in network planning, parameter optimization and intelligent operation and maintenance. In order to realize high-precision coverage manifold estimation, in recent years, more and more researches try to use deep learning models to learn the coverage distribution law from urban environment structure and base station layout to replace the traditional way of relying on artificial experience or rule modeling.

[0003] Most of the existing deep learning methods use the standard supervised learning paradigm, that is, by inputting source domain information (such as base station location map, terrain structure map, etc.), using a large number of target domain coverage manifold samples for supervision, training a neural network to realize the mapping from the source domain to the target domain.

[0004] In recent years, supervised learning cross-domain mapping methods based on deep learning have been widely used, such as CNN architecture, which has greatly improved performance compared to random geometry methods, but the effect is still not good. However, works represented by cGAN use an adversarial training mechanism to enhance the distribution alignment of the environmental features and coverage manifold features mapped by the generator, which can improve the accuracy of coverage manifold estimation. However, considering the large difference between the features of the environment and the coverage manifold of different domains, the stability and generalization of this supervised training cross-domain mapping cannot meet the application requirements.

[0005] In the coverage manifold estimation task, supervised learning has the following technical bottlenecks:

[0006] (1) There is a significant "cross-domain difference" between the source domain features and the target domain features, and it is difficult for supervised mapping to learn stably. The source domain is usually composed of structural geometric features such as geographic spatial information and base station distribution, while the target domain coverage manifold reflects the continuous distribution characteristics of wireless signals in space. Both have significant heterogeneity in distribution form, expression method and semantic structure. Specifically, supervised learning requires the model to directly learn this high-dimensional, nonlinear, cross-modal mapping relationship, which itself has high instability, and is prone to: insufficient gradient learning, weak feature alignment capability and mapping relationship sensitive to slight changes in input, resulting in large coverage estimation error, especially in complex urban environments.

[0007] (2) Supervised learning is difficult to have both "reconstruction ability" and "generation ability", in supervised reconstruction training, the network tends to learn "how to recover the covering manifold features given the target domain information", so it is good at reconstructing details, but lacks the ability to generate covering manifolds from pure source domain information;

[0008] (3) The multi-base station cooperation scene further magnifies the structural complexity of cross-domain mapping, under multi-base station cooperation strategies such as coherent combination and incoherent combination, the target domain covering manifold will present strong correlation, multi-scale change and spatial jump, etc. Therefore, the relationship between the source domain features (base station layout, spatial structure) and the target domain response (coverage probability) becomes more complex, and the supervised learning model is more difficult to establish a stable mapping, and its performance is significantly limited.

[0009] Therefore, the present application provides a source domain and target domain joint training based on self-supervised coverage manifold estimation method, which can simultaneously utilize source domain information and partial target domain information for joint training, to improve the feature extraction ability and generalization robustness of the model in the cross-domain mapping task, to solve the above technical problems. SUMMARY

[0010] The main purpose of the present application is to provide a source domain and target domain joint training based on self-supervised coverage manifold estimation method, by introducing a dynamic mask mechanism, the model has reconstruction ability and generation ability based on source domain information in the training process, thereby greatly improving the coverage manifold estimation precision and stability, to solve the technical problems of the existing supervised learning method in the background art in the city cellular network coverage manifold estimation, such as weak cross-domain feature alignment ability, dependence on complete target domain label, insufficient generation ability and poor generalization.

[0011] The present application solves the above technical problems by adopting the following technical solutions:

[0012] A source domain and target domain joint training based on self-supervised coverage manifold estimation method is executed by a computer device, comprising the following steps:

[0013] S1. Divide the specified area, obtain the base station position sample, terrain height information sample and coverage manifold sample corresponding to each region, and construct the source domain and target domain sample data set;

[0014] S2. Data enhancement processing and dynamic mask processing are performed on the source domain and target domain samples in turn;

[0015] S3. Construct a neural network prediction model framework composed of a feature coding fusion module and a coverage manifold generation module, which is used to input base station position samples, terrain height samples and masked coverage manifold samples, and perform a progressive masking generation strategy until a full mask is generated, and finally obtain a coverage manifold prediction result;

[0016] S4. Based on the prediction result, calculate the reconstruction error of the mask area with the real coverage manifold, use L1 loss as the training loss, and update the network parameters of the prediction model through back propagation;

[0017] S5. Finally, input only the source domain information into the prediction model to generate the corresponding area coverage manifold prediction result. Specifically, test the trained model by inputting only the source domain information, and the model can directly generate the coverage manifold information of the corresponding area.

[0018] Preferably, the specific construction process of the source domain and target domain sample data set in S1 includes:

[0019] Divide the city map into fixed-size areas by grid, get areas of , where for area , form the corresponding base station position information sample and terrain information sample , finally get containing base station position information samples and containing terrain information samples;

[0020] Get the coverage manifold sample corresponding to the city map by modeling simulation or field measurement and containing the coverage manifold sample;

[0021] Form the above three types of data into a training database, and divide it into a training set and a test set according to a predetermined proportion to ensure that the two do not intersect.

[0022] Preferably, in step S2, before training, the following multi-angle flipping data augmentation operation is simultaneously performed on the base station map, topographic map, and coverage manifold map: 90°, 180°, and 270° rotations are performed. This data augmentation can improve data diversity and enable the model to achieve better generalization performance. Dynamic masking is used to partially or fully mask the coverage manifold samples, allowing the model to learn reconstruction and generation capabilities simultaneously under different visibility conditions. At this time, a random mask matrix is ​​generated for each coverage manifold sample, and the target domain samples are partially or fully masked. Low masking rate: retains some coverage manifold features; high masking rate or full masking: hides the coverage manifold, preventing the model from directly accessing the target domain information. The dynamic masking mechanism is used to regulate the model's dependence on target domain information during training, enabling the model to learn cross-domain alignment and target domain generation capabilities under different masking conditions. Specifically, this includes:

[0023] Masking is performed on the samples covering the manifold, and the first... each region The mask attribute is There exists the first Grid in each region The mask attribute is ,like A value of 0 indicates that no masking is performed for the region. The grid is preserved if A value of 1 indicates that a mask is applied to that position, and a definition is made. , for The number of non-zero elements is used to measure the number of dynamic mask grids. The masking operation is defined as follows:

[0024]

[0025] in, This represents element-wise multiplication in a matrix. Indicates the first The overlay manifold sample after masking each region Represents a matrix consisting entirely of 1s. Indicates the first The overlay manifold sample in front of the region mask;

[0026] in Dynamic changes during training, when near Indicates a high mask rate. Indicates a complete cover;

[0027] In the early stage of training, the mask ratio gradually increases from 0% to 30%. In the middle stage of training, the mask ratio gradually increases from 30% to 70%. In the later stage of training, the mask ratio gradually increases to 100%, i.e., full mask.

[0028] Preferably, the neural network prediction model framework construction process in S3 includes:

[0029] S31. Input the base station location sample, terrain information sample and masked coverage manifold sample into a set of Unet encoder networks to obtain base station location features, terrain height features and masked coverage manifold features respectively.

[0030] S32. Feature splicing and convolution fusion are performed on base station location features, terrain height features, and mask coverage manifold features to obtain fused features with a unified cross-domain expression, which is used to improve the consistency of feature spaces between the source domain and the target domain.

[0031] S33. Using the covering manifold generated by the Unet decoder, the feature input Unet encoder network module is fused, and the spatial structure information is recovered through stepwise upsampling and skip connections. The predicted covering manifold result is output, and finally the mapping from the source domain to the target domain is realized.

[0032] Preferably, the Unet encoder network in S31 It contains 3 layers of convolutional downsampling, where the first layer is... The layer encoder output features are The calculation process is expressed as follows:

[0033]

[0034] in, Indicates the encoder's first... The output features of the layer This indicates the computation of the convolutional module, which consists of two layers. It consists of convolution, ReLU activation function, and batch normalization operation. This indicates a max pooling operation.

[0035] Preferably, the calculation formula for feature splicing in S32 is as follows:

[0036]

[0037] in, This is represented as after feature splicing and fusion operation. Subsequent sample output data, Depend on It consists of convolution, ReLU activation function, and batch normalization operation. Indicates a serial connection operation. , and These represent base station location samples respectively. Topographic information samples Overlay manifold sample after masking through the Unet encoder network processed final output.

[0038] Preferably, the fusion features in S33 are input into the Unet decoder network to output the cover manifold prediction, wherein the decoder network includes three layers of convolutional upsampling, the first layer upsampling module receives the output from the first layer of the encoder, and the output of the first layer upsampling module is fused with the skip connection features of the first layer of the encoder, and the output of the first layer upsampling module is fused with the skip connection features of the first

[0039]

[0040] wherein, represents the output features of the first layer of the decoder, represents the upsampling operation, represents the skip connection features of the first layer;

[0041] The final output of the decoder is the cover manifold prediction:

[0042]

[0043] wherein, is the decoder network, is the first cover manifold sample prediction output by the decoder.

[0044] Preferably, the mask ratio of the progressive mask generation strategy in S3 is dynamically adjusted according to the training stage, and the whole process is divided into three stages:

[0045] 1) Early training stage: the model mainly uses a low mask rate (0-30%), and at this time most of the target domain information is contained in the input. With the assistance of the source domain information, the model focuses on learning the reconstruction and completion of the missing part of the target domain, and preliminarily establishes the cross-domain feature alignment capability.

[0046] 2) Mid-training stage: as the reconstruction capability of the model improves, the mask rate is gradually increased to 30-70%. The input of the target domain information is correspondingly reduced, which promotes the model to complete more part reconstruction while deepening the understanding of the complex nonlinear relationship between the source domain and the target domain, and further strengthens the stability of feature fusion and cross-domain mapping.

[0047] 3) Late training stage: the model has strong reconstruction ability and fully understands the source domain-target domain relationship. At this time, the mask rate is further increased to 70-100%, until full masking (i.e. full masking state). The model only relies on source domain information for prediction at this stage, realizing the ability conversion from "reconstruction depending on part of the target domain" to "generation based on the source domain".

[0048] Preferably, the S4 adopts L1 loss back propagation to update the network parameters of the prediction model, and the loss function is:

[0049]

[0050] Wherein, Indicates the number of samples, Indicates the number of samples, Indicates the number of samples, Indicates the number of samples, Indicates the number of samples, Indicates the number of samples, Indicates the number of samples, Indicates the number of samples, Indicates the number of samples,

[0051] From the above technical solution, the present application provides a kind of based on self-supervision's source domain and target domain joint training's covering manifold estimation method.Compared with prior art, the present application has the following advantages:

[0052] 1.The present application processes part or full mask processing to the target domain covering manifold through dynamic mask mechanism, so that the model learns reconstruction ability and generation ability at the same time during training, can strengthen the recovery of target domain details under low mask rate, enhance the generation ability based on source domain under high mask rate, and effectively improve the spatial reconstruction accuracy of covering manifold by cooperating with multi-modal feature fusion and Unet decoding structure, so that the shape and details of covering manifold are closer to the real distribution.

[0053] 2.The present application introduces dynamic mask self-supervised training mechanism into source domain and target domain joint learning framework, constructs multiple training scenes through random mask, so that the model can optimize cross-domain feature alignment ability and target domain generation ability under different visibility conditions, and finally the model can generate covering manifold prediction results only by inputting source domain information, which solves the consistency problem such as "training depends on target domain, inference lacks target domain" in supervised learning, not only enhances the stability of cross-domain mapping, but also makes the model have strong robust generation ability.

[0054] 3. The application can maintain stable estimation performance under different base station densities, different urban areas and different coverage complexities, can be applied to multi-base station cooperation scenes such as coherent combination and incoherent combination, and since the method is not sensitive to specific propagation models, it can also be extended to other types of cross-domain learning tasks or other spatial prediction tasks, and has strong engineering application potential and universality.

[0055] It should be understood that the matters described in this section are not intended to identify key or important features of the embodiments of the application, nor are they used to limit the scope of the application. Other features of the application will become apparent from the following description. Of course, any product implementing the application does not necessarily need to achieve all the advantages mentioned above. BRIEF DESCRIPTION OF DRAWINGS

[0056] The drawings accompanying the specification of this application form a part thereof, serve to provide further understanding of the application, and together with the description of the exemplary embodiments of the application given below, serve to explain the application, and do not constitute an improper limitation on the application. In the drawings:

[0057] Figure 1 It is a schematic diagram of the overall process of the application;

[0058] Figure 2 It is a schematic block diagram of the model training process of the application;

[0059] Figure 3 It is a schematic diagram of the operation process of the dynamic mask module of the application;

[0060] Figure 4 It is a schematic diagram of the operation process of the feature fusion module of the application;

[0061] Figure 5 It is a schematic block diagram of the model testing process of the application;

[0062] Figure 6 It is a visualization schematic diagram of the coverage manifold estimation of the application, wherein: (a) is a visualization schematic diagram of the coverage manifold sample Groundtruth, (b) is a visualization schematic diagram of the sample of the application (JoSS) algorithm in the coverage manifold estimation, (c) is a visualization schematic diagram of the sample of the coverage manifold cGAN algorithm, and (d) is a visualization schematic diagram of the sample of the coverage manifold CNN-AE algorithm. DETAILED DESCRIPTION

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] For details in the embodiments, please refer to Figures 1 to 5 .

[0065] refer to Figure 1 The covering manifold estimation method based on self-supervised joint training of the source and target domains proposed in this invention includes the following steps:

[0066] S1. Through The grid divides the city map of a specified area into fixed-size grids. The region was ultimately obtained containing each region of Corresponding to each region There are corresponding base station location samples (one of the source domain features, obtained by marking the grid location of the base station in a two-dimensional plane, used to reflect the base station deployment structure), terrain height information samples (one of the source domain features, which is two-dimensional regular rasterized data used to characterize spatial elevation changes), and coverage manifold samples (target domain features, the coverage probability value for each grid point). Among them, the base station location information samples... and terrain information samples Finally, samples containing base station location information are obtained. and samples containing terrain information This is to construct a source domain and target domain sample dataset that can be used for model training, and to simultaneously divide the dataset into training and test sets.

[0067] At this point, a coverage manifold sample corresponding to the city map is obtained through modeling simulation or field measurement. And samples containing the covered manifold The three types of data collected above are used to form a training database, which is then divided into a training set and a test set according to a predetermined ratio. The training set is used for model parameter learning, and the test set is used for performance verification. The two sets do not overlap. At this time, data with different numbers of base stations and different cooperation strategies (coherent merging and incoherent merging) can form different types of coverage manifold samples, thereby improving the diversity of training data.

[0068] S2. Perform data augmentation and dynamic masking processing on the source and target domain samples in sequence.

[0069] At this time, before training, the following multi-angle flipping data enhancement operations are synchronously performed on the base station graph, topographic map and coverage manifold graph: including performing 90°, 180°, 270° rotation, here the data enhancement can expand the sample quantity without changing the essential characteristics of the coverage pattern, thereby being able to improve the data diversity, so that the coverage manifold estimation model obtained by training has better generalization performance and generation capability, and the dynamic mask processing is used for partial masking or full masking of the coverage manifold sample, so that the model simultaneously learns the reconstruction capability and generation capability under different visibility conditions, at this time, a random mask matrix is generated for each coverage manifold sample, and the target domain sample is partially masked or fully masked, the random mask matrix is generated in a uniform distribution or Poisson sampling manner, so that the mask position is irregularly distributed, and the mask proportion is randomly set in the range of 0-1, wherein, low mask rate: retaining part of the coverage manifold features, and enhancing the reconstruction capability of the model by retaining part of the target domain features; high mask rate or full mask: hiding the coverage manifold, so that the model cannot directly access the target domain information, and making the model rely on the source domain to generate the coverage map; at this time, the positions of 1 in the mask matrix represent the masked area, and the positions of 0 represent the retained area. The mask operation is applied to the coverage manifold sample by element-wise Hadamard product, so that the masked area does not contain target domain information at the input end.

[0070] Further, the dynamic mask mechanism randomly changes at each epoch to ensure that the model simultaneously learns the cross-domain mapping relationship under multiple mask conditions during training, thereby being able to be used to regulate the degree of dependence of the model on the target domain information during training, so that the model learns the cross-domain alignment capability and the target domain generation capability under different mask conditions, respectively, and specifically includes:

[0071] Mask processing is performed on the coverage manifold sample, and the mask attribute of the first area is defined as , if is 0, it means that no mask is performed, and the grid in the area is retained, if is 1, it means that the position is masked, and , is defined as the number of non-zero elements in , which is used to measure the number of dynamic mask grids, and the mask operation is defined as follows:

[0072]

[0073] Among them, represents element-wise multiplication in the matrix, represents the first a coverage manifold sample after a region mask, a matrix representing all 1s;

[0074] wherein dynamically changes during the training process, when close to represents a high mask rate, represents full mask;

[0075] Reference Figure 3 , at the beginning of training, the mask rate gradually increases from 0 to 30%, in the middle of training, the mask rate gradually increases from 30% to 70%, and in the later stage of training, the mask rate gradually increases to 100%, that is, full mask.

[0076] At this time, the target domain coverage manifold is partially or fully masked by the dynamic mask mechanism, so that the model learns the reconstruction ability and generation ability at the same time during training; under low mask rate, the details of the target domain are strengthened to recover, and under high mask rate, the generation ability based on the source domain is enhanced, while cooperating with multi-modal feature fusion and Unet decoding structure, effectively improving the spatial reconstruction accuracy of the coverage manifold.

[0077] S3. Construct a neural network prediction model framework composed of a feature encoding fusion module and a coverage manifold generation module, which is used to input the base station position sample, the terrain height sample and the coverage manifold sample after mask processing, and execute the progressive mask generation strategy until the full mask is generated, and finally obtain the coverage manifold prediction result;

[0078] The neural network prediction model framework construction process includes:

[0079] S31. Input the base station position sample, the terrain information sample and the coverage manifold sample after mask into a group of Unet encoder networks at the same time, at this time, the three input channels are input into the Unet encoder network respectively, local spatial features are extracted through the convolution layer, and finally the base station position feature, the terrain height feature and the mask coverage manifold feature are obtained respectively;

[0080] In actual use, the Unet encoder network contains 3 layers of convolution downsampling, wherein the first layer encoder output feature is , and its calculation process is represented as:

[0081]

[0082] wherein, represents the output feature of the first layer of the encoder, represents the convolution module calculation, and the convolution module is composed of two layers of It consists of convolution, ReLU activation function, and batch normalization operation. This represents the max pooling operation;

[0083] S32. Feature splicing and convolution fusion are performed on base station location features, terrain height features, and mask coverage manifold features to obtain fused features with a unified cross-domain expression, which is used to improve the consistency of feature spaces between the source domain and the target domain.

[0084] In practical use, refer to Figure 4 The formula for calculating feature splicing is:

[0085]

[0086] in, This is represented as after feature splicing and fusion operation. Subsequent sample output data, Depend on It consists of convolution, ReLU activation function, and batch normalization operation. Indicates a serial connection operation. , and These represent base station location samples respectively. Topographic information samples Overlay manifold sample after masking via Unet encoder network The final output after processing;

[0087] S33. Using the covering manifold generated by the Unet decoder, the feature input Unet encoder network module is fused, and the spatial structure information is recovered through stepwise upsampling and skip connections. The predicted covering manifold result is output, and finally the mapping from the source domain to the target domain is realized.

[0088] In practical use, refer to Figure 2 The fused features are input into the Unet decoder network to output a manifold-covering prediction, where the decoder network... It includes three layers of convolutional upsampling, the first... The layer upsampling module receives from The output of the upsampling module of the layer and the corresponding encoder of the layer The skip connection features of the layers are fused, the first layer Upsampling module output features for:

[0089]

[0090] in, Indicates the decoder's first The output features of the layer denotes an up-sampling operation, denotes the layer of the Unet encoder network;

[0091] The decoder finally outputs the coverage manifold prediction as:

[0092]

[0093] wherein, is the decoder network, is the th coverage manifold sample prediction output by the decoder.

[0094] Further, in an embodiment, each layer of the Unet encoder network comprises a convolution, an activation and a down-sampling operation, so that features are extracted layer by layer and form deep representations of different scales, and the three types of output features are spliced in the channel dimension to form a fusion feature containing multi-modal information; the fusion feature is processed by a set of convolution fusion modules for feature unification, so that the source domain feature and the target domain residual feature are aligned in the same semantic space, and then a feature fusion operation is performed through a feature fusion module, which can include a 1x1 convolution for compressing the number of channels.

[0095] In the coverage manifold generation process, the fusion feature gradually recovers the spatial resolution through multiple deconvolution or up-sampling modules; in the up-sampling process, the jump connection features corresponding to the encoder layers are spliced with the current decoding features to compensate for the loss of details in the up-sampling process; after multiple layers of convolution processing, the decoder finally outputs a coverage manifold prediction map of the same size as the input region;

[0096] The final output layer uses a ReLU activation function to generate a continuous value of coverage intensity or a coverage probability distribution;

[0097] At this time, the jump connection mechanism can also significantly enhance the recovery ability of edge structure and spatial details, and improve the accuracy of coverage manifold prediction

[0098] S4. Based on the prediction result, the reconstruction error of the prediction result and the real coverage manifold in the mask region is calculated, L1 loss is used as the training loss, the network parameters of the prediction model are updated through back propagation, wherein only the L1 loss between the prediction result and the real coverage manifold in the mask region is calculated, so that the model focuses on learning the completion ability of the masked region, and finally the model parameters are updated through the AdamW optimizer for back propagation, and the dynamic mask changes randomly in each iteration, so that the model obtains robust cross-domain mapping ability, and the loss function is specifically:

[0099]

[0100] wherein, denotes the number of samples, denotes the number of masks in the first sample, and the loss function uses the mean absolute error to measure the estimation accuracy of the mask area.

[0101] S5. Finally, only the source domain information is input into the prediction model to generate the corresponding regional coverage manifold prediction result, please refer to Figure 5 , by testing the trained model, only the source domain information is input, and the model can directly generate the coverage manifold information of the corresponding region.

[0102] The model uses a clear Unet encoding-decoding framework, and through mask operation, feature fusion and self-supervised loss, a complete training process is formed. The overall structure is simple and easy to understand, easy to implement and deploy. The model training and inference process is stable, and the hardware and computing power requirements are moderate, which is convenient for integration and application in actual network planning and optimization system.

[0103] In summary, this method introduces a dynamic mask self-supervised training mechanism into the source domain and target domain joint learning framework for the first time. By constructing multiple training scenarios through random masks, the model can optimize the cross-domain feature alignment capability and target domain generation capability under different visibility conditions, solving the consistency problem of "training dependent on target domain, inference lacking target domain" in supervised learning. This mechanism not only enhances the stability of cross-domain mapping, but also enables the model to have strong robustness and generation capability. Finally, it can be applied to coherent integration and incoherent integration and other multi-base station cooperation scenarios, and can maintain stable estimation performance under different base station densities, different city areas and different coverage complexities. At the same time, since this method is not sensitive to specific propagation models, it can also be extended to other types of cross-domain learning tasks or other spatial prediction tasks, and has strong engineering application potential and universality.

[0104] Further, the method also adopts the following implementation scheme in the specific implementation process: the design of the gradual mask generation strategy is based on the requirement of the method for target domain prior information in the training process, and adopts a gradual learning mechanism from more to less to gradually enhance the model's ability to generate target domains from source domains. Specifically, the mask ratio is dynamically adjusted with the training stage, and the whole process is divided into three stages:

[0105] 1) Early training: the model mainly uses low mask rate (0-30%), and at this time most of the target domain information is contained in the input. The model learns to reconstruct and complete the missing part of the target domain with the help of the source domain information, and initially establishes the cross-domain feature alignment capability.

[0106] 2) Mid-training: As the model's reconstruction ability improves, the mask rate is gradually increased to 30-70%. The target domain information input is correspondingly reduced, which promotes the model to complete more partial reconstruction while deepening the understanding of the complex nonlinear relationship between the source domain and the target domain, further strengthening the stability of feature fusion and cross-domain mapping.

[0107] 3) Late training: The model has strong reconstruction ability and a full understanding of the source domain-target domain relationship. At this time, the mask rate is further increased to 70-100%, and even to full masking (i.e., full masking state). The model relies only on source domain information for prediction in this stage, realizing the conversion from "reconstruction relying on part of the target domain" to "generation based on the source domain".

[0108] This strategy achieves a smooth transition from reconstruction completion to full generation by dynamically adjusting the mask ratio. Experiments show that the gradual mask setting of "0-30% in the early stage, 30-70% in the middle stage, and 70-100% in the late stage" can effectively balance the learning goals at different stages, and ultimately enable the model to have robust coverage manifold generation capability.

[0109] In a specific embodiment, by comparing the mainstream coverage manifold prediction methods respectively, such as PPP, BFSG, CNN-AE, cGAN, the experiment uses a unified urban cellular coverage manifold dataset, the signal-to-noise ratio threshold value for calculating the coverage rate is set to 0dB, 5dB, 10dB, 15dB and 20dB, the number of base stations in each sample is 8-12, the terrain area size is 1000m x 1000m, the mean absolute error (MAE) index, the root mean square error (RMSE) index and the visualization comparison are used as the main evaluation index, and the coverage manifold estimation performance (the algorithm proposed in the application is referred to as JoSS) is as follows Table 1 (MAE) and Table 2 (RMSE) show:

[0110] Table 1: Performance comparison table of coverage manifold algorithm (MAE)

[0111]

[0112] Based on the above Table 1, for the mean absolute error (MAE) index, the result of the present application is the optimal value, and the result of cGAN is the suboptimal value.

[0113] Table 2: Performance comparison table of coverage manifold algorithm (RMSE)

[0114]

[0115] Based on the comparison of the above Tables 1 and 2, PPP and BFSG are random geometry-based methods, which predict the entire study area as a spatial average when estimating the coverage manifold, and thus do not fully utilize the base station locations and terrain environmental features, and the estimation effect is poor.

[0116] In addition, the CNN-AE and the cGAN use the base station locations and the terrain information as the model input, which is essentially a cross-domain supervised training. Since the feature gap between the source domain and the target domain is large, the performance still has room for improvement.

[0117] Therefore, JoSS encodes the base station location information, the terrain information and the coverage manifold information after masking, and performs multi-modal feature fusion at the encoding layer. Through the dynamic masking manner, the decoder trains the generation capability of the target domain and realizes the feature alignment of the source domain and the target domain through the reconstruction training, so that the performance of JoSS is improved by up to 30% compared with the cGAN algorithm of the cross-domain supervised training. At this time, the present application also visualizes the CNN-AE and the cGAN in the above experiment and the present application to intuitively show the effect of the present application, which is specifically as shown in Figure 6 , Figure 6 Fig. 4 is a visual diagram of the coverage manifold sample Ground truth, the coverage manifold sample of the present application (JoSS) algorithm, the coverage manifold sample of the cGAN algorithm and the coverage manifold sample of the CNN-AE algorithm.

[0118] It can be observed from Figure 6 that the JoSS algorithm estimation graph is closest to the database coverage manifold sample, the cGAN algorithm estimation graph can estimate the general outline of the coverage manifold, but the estimation effect in the details is poor, the CNN-AE is poor in estimating the outline of the coverage manifold, and the performance is the worst.

[0119] In addition, according to the above experimental results, the present application achieves lower estimation error under different cooperation strategies, different signal-to-noise ratio thresholds and different base station densities, the coverage manifold shape and details are closer to the real distribution, and the overall performance is better than that of the traditional supervised learning method.

[0120] In another embodiment provided by the present application, a computer program product containing instructions is also provided, which, when running on a computer, causes the computer to execute the coverage manifold estimation method based on self-supervised source domain and target domain joint training in any of the above embodiments.

[0121] It can be understood that the system provided by the embodiments of the present application corresponds to the method provided by the embodiments of the present application, and the explanation, examples and beneficial effects of the related content can refer to the corresponding parts in the above method.

[0122] The embodiment of the present application also provides an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete mutual communication through the communication bus,

[0123] The memory is used for storing a computer program.

[0124] The processor is used for executing the program stored on the memory, and realizing the above-mentioned self-supervised source domain and target domain joint training based on the covering manifold estimation method.

[0125] The communication bus mentioned in the above-mentioned electronic device can be a peripheral component interconnect bus or an extended industry standard architecture bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0126] The communication interface is used for communication between the above-mentioned electronic device and other devices.

[0127] The memory can include a random access memory, and can also include a non-volatile memory, for example, at least one disk memory. Optionally, the memory can also be at least one storage device located away from the processor.

[0128] The processor mentioned above can be a general-purpose processor, including a central processing unit, a network processing unit, etc.; can also be a digital signal processor, an application specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0129] In the above-mentioned embodiment, all or part of it can be realized by software, hardware, firmware or any combination thereof. When realized by software, all or part of it can be realized in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiment of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable devices. The computer instructions can be stored in a computer readable storage medium or transferred from one computer readable storage medium to another, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired or wireless mode. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available medium can be a magnetic medium, an optical medium or a semiconductor medium, etc.

[0130] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0131] In addition, it should be noted that if the embodiments of the present application involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative position relationship, movement condition, etc. between components in a certain posture, and if the certain posture changes, the directional indications also change accordingly.

[0132] In addition, if the embodiments of the present application involve descriptions of "first", "second", etc., the descriptions of "first", "second", etc. are only for description purposes and cannot be understood as indicating or implying the relative importance of the technical features or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include at least one of the features. In addition, the meaning of "and / or" appearing throughout the text includes three parallel solutions. Taking "A and / or B" as an example, it includes A solution, or B solution, or A and B solutions. In addition, in the embodiments of the present application, "a plurality of" means two or more. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the realization of ordinary skilled in the art, when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, nor is it within the protection scope required by the present application.

Claims

1. A self-supervised based source and target domain joint training method for covering manifold estimation, characterized in that, The application relates to a method for predicting coverage manifold of a target area based on a source area, comprising the following steps: S1. Dividing a designated area to obtain base station position samples, terrain information samples and coverage manifold samples corresponding to each region, and constructing a source domain and target domain sample data set; S2. Sequentially performing data enhancement processing and dynamic mask processing on the source domain and target domain samples; S3. Constructing a neural network prediction model framework composed of a feature encoding fusion module and a coverage manifold generation module, which is used for performing a progressive mask generation strategy after inputting the base station position samples, the terrain height samples and the mask-processed coverage manifold samples, until a full mask is generated, and finally obtaining a coverage manifold prediction result; S4. Based on the prediction result, the reconstruction error of the prediction result and the real coverage manifold in the mask area is calculated, L1 loss is used as the training loss, and the network parameters of the prediction model are updated through back propagation; S5. Finally, the source domain information is input into the prediction model to generate a corresponding region coverage manifold prediction result; The dynamic mask processing in S2 is used for partial mask or full mask of the coverage manifold samples, and specifically comprises: The mask processing is performed on the coverage manifold sample, and the mask attribute of the first region is defined as , the mask attribute of the grid in the second region is defined as , if the mask attribute is 0, it means that no mask is performed, and the grid in the region is reserved, if the mask attribute is 1, it means that the position is masked, and the mask attribute is defined as , the number of non-zero elements in is , which is used to measure the number of dynamic mask grids, and the mask operation is defined as follows: , , , , , , . wherein, denotes element-wise multiplication of matrices, denotes the th region masked coverage manifold sample, denotes a matrix of all ones, denotes the th region unmasked coverage manifold sample.

2. The self-supervision based source domain and target domain jointly trained overlay manifold estimation method of claim 1, wherein, The specific construction process of the source domain and target domain sample data set in S1 comprises: By dividing the city map into fixed-size areas using a grid, a area containing samples of base station location information and samples of terrain information is obtained for each area ; obtaining a sample of a coverage manifold corresponding to the city map by modeling simulation or field measurements ; The three types of data obtained above form a training database, which is divided into a training set and a test set according to a predetermined proportion.

3. The self-supervision based source domain and target domain jointly trained overlay manifold estimation method of claim 1, wherein, The construction process of the neural network prediction model framework in S3 comprises: S31. The base station position samples, the terrain information samples and the mask-processed coverage manifold samples are simultaneously input into a group of Unet encoder networks to obtain base station position features, terrain height features and mask coverage manifold features respectively; S32. The base station position features, the terrain height features and the mask coverage manifold features are subjected to feature splicing and convolution fusion to obtain fusion features in a cross-domain unified expression; S33. The coverage manifold generated by the Unet decoder is input into the Unet encoder network module, the spatial structure information is recovered through step-by-step upsampling and jump connection, the predicted coverage manifold result is output, and finally the mapping from the source domain to the target domain is realized.

4. The self-supervision based source domain and target domain jointly trained overlay manifold estimation method of claim 3, wherein, The Unet encoder network in S31 contains 3 layers of convolutional downsampling, where the first layer is... The layer encoder output features are The calculation process is expressed as follows: wherein, represents the output of the encoder first layer, represents the convolution module computation, the convolution module being composed of two layers of convolution, ReLU activation function and batch normalization operations, represents the max-pooling operation.

5. The self-supervision based source domain and target domain jointly trained overlay manifold estimation method of claim 4, wherein, The calculation formula of the feature splicing in S32 is: in, This is represented as after feature splicing and fusion operation. Subsequent sample output data, Depend on It consists of convolution, ReLU activation function, and batch normalization operation. Indicates a serial connection operation. , and These represent base station location samples respectively. Topographic information samples Overlay manifold sample after masking The final output after network processing by the Unet encoder.

6. The self-supervision based source domain and target domain jointly trained overlay manifold estimation method of claim 5, wherein, The fusion features in S33 are input into a Unet decoder network to output the overlay manifold prediction, where the decoder network includes three layers of convolutional up-sampling, the first layer up-sampling module receives the features from the output of the up-sampling module of the first layer and fuses the features with the corresponding encoder first layer skip connection features, the output of the first layer up-sampling module is wherein, denotes the output of the decoder first layer, denotes an up-sampling operation, denotes the skip connection feature of the first layer. The final output of the decoder is the coverage manifold prediction: wherein, is a decoder network, is a first cover manifold sample prediction.

7. The self-supervision based source domain and target domain jointly trained overlay manifold estimation method of claim 6, wherein, In S4, the network parameters of the prediction model are updated through back propagation of L1 loss, and the loss function is: in, Indicates the number of samples. Indicates the first The number of samples with a mask of 1. Indicates the first Grid in each region Overlay manifold sample in front of the mask, Indicates the first Grid in each region The decoder's final output covers the manifold prediction results.

Images