Wireless network resource joint optimization method and system based on federal feature rarefaction

Abstract

The invention discloses a wireless network resource joint optimization method and system based on federal feature rarefaction, relates to the technical field of wireless communication network resource management, and aims to reduce the transmission data volume through feature rarefaction processing and improve the resource utilization rate. The privacy is protected in combination with noise disturbance, and the resource allocation strategy is dynamically generated based on global feature analysis, so that the resource allocation efficiency can be improved, the communication overhead can be reduced, and the privacy protection capability is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication network resource management, in particular to a wireless network resource joint optimization method and system based on federated feature sparsification. BACKGROUND

[0002] In the current field of wireless network communication, with the rapid growth of the number of terminal devices and the diversification of network applications, the management and optimization of wireless network resources are facing unprecedented challenges. Traditional wireless network resource allocation methods often use centralized processing, and the state information of all terminal devices needs to be uploaded to the central node for unified processing, which not only leads to huge communication overhead, but also easily causes privacy leakage problems. At the same time, due to the spatio-temporal difference of the network state of different terminal devices, centralized processing is difficult to reflect the actual demand of each terminal in real time and accurately, thereby causing unbalanced resource allocation and low efficiency. In addition, the existing technology lacks effective feature screening mechanism when processing massive terminal device data, resulting in too much redundant data transmission, further aggravating network congestion. In terms of privacy protection, traditional methods usually use simple data desensitization or encrypted transmission, which is difficult to realize effective privacy protection while ensuring data usability. For the problem of dynamic allocation of network resources, most existing solutions are based on static models for decision-making, which cannot adapt to the rapid changes of wireless network environment, resulting in low resource utilization.

[0003] The above content is only used to assist in understanding the technical solutions of the present application, and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0004] The main purpose of the present application is to provide a wireless network resource joint optimization method and system based on federated feature sparsification, aiming to improve resource allocation efficiency and reduce communication overhead.

[0005] To achieve the above purpose, the present application provides a wireless network resource joint optimization method based on federated feature sparsification, which comprises: The terminal device collects the local network state original data set, and performs feature extraction and standardization processing on the original data set to generate a standardized feature vector; The terminal device performs feature importance analysis operation based on the standardized feature vector to generate a feature importance vector; The terminal device receives the channel capacity indication information issued by the base station, and performs sparse mask generation operation combined with the feature importance vector to generate a binary feature mask; The terminal device extracts a key feature subset from the standardized feature vector according to the binary feature mask, generates privacy protection feature data after adding noise disturbance to the key feature subset, and uploads the privacy protection feature data to an edge aggregation node; The edge aggregation node receives the privacy protection feature data uploaded by multiple terminals, performs a cross-device feature alignment operation on the privacy protection feature data, and generates an aligned feature data set; Perform weighted fusion processing on the aligned feature data set to generate a global feature space matrix; Based on the global feature space matrix, analyze the space-time correlation of the device, and generate a network resource strategy containing channel allocation parameters and terminal scheduling sequence; The base station controller executes the network resource strategy and broadcasts the updated channel state parameters to the terminal device.

[0006] In an embodiment, the terminal device collects a local network state original data set, and the step of generating a standardized feature vector by performing feature extraction and standardization processing on the original data set includes: The terminal device collects signal strength original data, transmission delay original data, and device energy consumption original data of the local network; Divide the original data into a preset time window, and calculate the time domain statistical features within the time window; Perform Fourier transform on the data within the time window to extract frequency domain principal component energy features; Respectively normalize the time domain statistical features and the frequency domain principal component energy features by maximum and minimum, and splice the normalized time domain statistical features and frequency domain principal component energy features to generate the standardized feature vector.

[0007] In an embodiment, the terminal device performs a feature importance analysis operation based on the standardized feature vector to generate a feature importance vector, which includes the following steps: Copy the standardized feature vector to generate a test vector copy, and set the specified feature dimension values in the test vector copy to zero to generate a shielded feature vector; Based on the preset network performance prediction function, calculate the network throughput prediction value corresponding to the shielded feature vector; Calculate the deviation amount of the reference network throughput prediction value of the standardized feature vector and the network throughput prediction value of the shielded feature vector; Generate a feature importance initial sequence according to the deviation amount, and perform sliding window smoothing processing on the feature importance initial sequence to generate the feature importance vector.

[0008] In an embodiment, the terminal device receives channel capacity indication information issued by a base station, and performs a sparse mask generation operation in combination with the feature importance vector, and the step of generating a binary feature mask comprises: Resolving the number of available transmission bits from the channel capacity indication information, and calculating the average number of encoding bits per feature dimension based on the locally stored historical transmission records; According to the number of available transmission bits and the average number of encoding bits per feature dimension, calculating the maximum number of transmittable dimensions; The feature dimensions are sorted in descending order according to the feature importance vector, and the top K dimensions are selected as the reserved feature dimension set, K is equal to the maximum number of transmittable dimensions; A binary mask matrix is constructed, the reserved feature dimension set is set to 1 at the corresponding position, and the rest is set to 0, to generate the binary feature mask.

[0009] In an embodiment, the step of generating privacy-protected feature data after adding noise disturbance to the key feature subset comprises: According to the feature type of the key feature subset, query the preset noise intensity coefficient mapping table to obtain the noise intensity coefficient of the current feature, and generate a Laplace distribution random number based on the noise intensity coefficient; The random number is superimposed on the numerical part of the key feature subset; Performing boundary truncation processing on the superimposed feature value to generate the privacy-protected feature data.

[0010] In an embodiment, the edge aggregation node receives the privacy-protected feature data uploaded by multiple terminals, and performs a cross-device feature alignment operation on the privacy-protected feature data to generate an aligned feature data set, and the step comprises: Resolving the dimension description information of the privacy-protected feature data uploaded by each terminal, extracting the correspondence between the feature name and the dimension index; according to the similarity of the feature name, cross-terminal dimension matching is performed to generate a terminal feature dimension mapping relationship table; Adjusting the dimension order of the privacy-protected feature data of each terminal according to the terminal feature dimension mapping relationship table; Detecting the missing dimension position in the adjusted data, and filling the missing dimension position with the global feature historical mean of the nearest preset first number of historical data collection periods to generate the aligned feature data set.

[0011] In an embodiment, the step of performing weighted fusion processing on the aligned feature data set to generate a global feature space matrix comprises: Statistical analysis of the number of local data samples of each terminal, and determining the aggregation weight coefficient of each terminal according to the number of local data samples and the total number of samples; performing a weighted calculation on the feature values of each feature dimension in the alignment feature data set; performing standard deviation normalization processing on the weighted calculation results, and integrating the normalized values of all dimensions to generate the global feature space matrix.

[0012] In an embodiment, the step of generating a network resource policy including a channel allocation parameter and a terminal scheduling sequence based on the analysis of the space-time correlation of the device based on the global feature space matrix includes: obtaining geographic position coordinate data reported by a terminal device, and calculating a Euclidean distance matrix between terminal devices based on the geographic position coordinate data; constructing a space-time correlation matrix according to the Euclidean distance matrix and the service type of the device; performing cluster analysis on the space-time correlation matrix, identifying a communication device group, and assigning a channel bandwidth ratio to each group based on the communication device group; generating a terminal scheduling priority sequence according to the service urgency; fusing the channel bandwidth allocation ratio and the terminal scheduling priority sequence to generate the network resource policy.

[0013] In an embodiment, the method further includes: calculating the cosine similarity between the network resource policies of a continuous preset second number of decision cycles; when the cosine similarity is less than a preset threshold, generating a sparsity adjustment instruction according to the cosine similarity and sending the instruction to a terminal device; the terminal device updates the size of the sliding window in the feature importance analysis according to the sparsity adjustment instruction.

[0014] In addition, to achieve the above-mentioned purpose, the present application also proposes a wireless network resource joint optimization system based on federal feature sparsification, which comprises a memory, a processor, and a wireless network resource joint optimization program based on federal feature sparsification stored on the memory and executable on the processor. The wireless network resource joint optimization program based on federal feature sparsification is configured to implement the steps of the wireless network resource joint optimization method based on federal feature sparsification.

[0015] The wireless network resource joint optimization method and system based on federal feature sparsification proposed in the present application can reduce the amount of transmission data through feature sparsification processing, protect privacy combined with noise disturbance, and dynamically generate resource allocation strategies based on global feature analysis, which can improve resource allocation efficiency and reduce communication overhead, while improving privacy protection capability. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings, which are incorporated herein and constitute part of the specification, illustrate embodiments consistent with the application and, together with the description, further serve to explain the principles of the application.

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, for those of ordinary skill in the art, the other drawings can also be obtained based on these drawings without any creative work.

[0018] Figure 1 The flowchart provided by an embodiment of the wireless network resource joint optimization method based on federal feature sparsification of the present application; Figure 2 The flowchart provided by an embodiment of the wireless network resource joint optimization method based on federal feature sparsification of the present application; Figure 1 The detailed flowchart of step S100 in the embodiment of the present application; Figure 3 The detailed flowchart of step S200 in the embodiment of the present application; Figure 1 The detailed flowchart of step S300 in the embodiment of the present application; Figure 4 Figure 1 The detailed flowchart of step S400 in the embodiment of the present application; Figure 5 The detailed flowchart of step S500 in the embodiment of the present application; Figure 1 The detailed flowchart of step S600 in the embodiment of the present application; Figure 6 Figure 1 The detailed flowchart of step S700 in the embodiment of the present application; Figure 7 The detailed flowchart of step S700 in the embodiment of the present application; Figure 1 The detailed flowchart of step S700 in the embodiment of the present application; Figure 8 Figure 1 The detailed flowchart of step S700 in the embodiment of the present application; Figure 9 The flowchart provided by another embodiment of the wireless network resource joint optimization method based on federal feature sparsification of the present application; Figure 10 The structural diagram provided by an embodiment of the wireless network resource joint optimization system based on federal feature sparsification of the present application.

[0019] Explanation of reference signs: 10, memory; 20, processor.

[0020] The purposes, functional features and advantages of the present application will be further illustrated with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0021] ​​​The technical solutions in the present application will be described clearly and completely in the present application combined with the drawings in the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. The components of the present application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0022] It should be understood that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0023] In the prior art, the field of wireless network communication faces the challenge brought by the explosive growth of the number of terminal devices and the diversification of network applications. The traditional method relies on centralized processing, which requires all terminal devices to upload complete state information to the center node, resulting in excessive communication overhead and the risk of privacy leakage. The network status of different terminal devices has spatiotemporal differences, and centralized processing cannot reflect the actual needs of each terminal in real time, resulting in uneven resource allocation and low efficiency. For example, in a dense urban environment, when a large number of mobile terminals simultaneously request network resources, the center node cannot quickly process the dynamically changing device status, resulting in increased communication delay for some terminals.

[0024] In order to solve the above problems, it is necessary to design a distributed resource optimization method that can reduce communication overhead and protect privacy. Considering the local data processing capability of terminal devices, the amount of transmitted data can be reduced through feature screening, but it is necessary to ensure that the screened features can still accurately reflect the network status; if a federated learning framework is used, each terminal independently processes data and only uploads key information, it is also necessary to protect privacy while achieving global resource optimization. Therefore, by introducing a dynamic feature sparsification mechanism, the feature dimension is adaptively adjusted according to the channel capacity to balance transmission efficiency and information integrity; in addition, cross-device feature alignment and fusion are performed at the edge node, which can effectively capture spatiotemporal correlation and provide decision basis for resource allocation.

[0025] Based on this, the embodiments of the present application provide a wireless network resource joint optimization method based on federated feature sparsification, referring to Figure 1 The wireless network resource joint optimization method based on federated feature sparsification includes steps S100-S800, wherein: In step S100, the terminal device collects a local network state original data set, and performs feature extraction and standardization processing on the original data set to generate a standardized feature vector. In step S200, the terminal device performs feature importance analysis operation based on the standardized feature vector to generate a feature importance vector. In step S300, the terminal device receives channel capacity indication information issued by the base station, and performs sparse mask generation operation in combination with the feature importance vector to generate a binary feature mask. In step S400, the terminal device extracts a key feature subset from the standardized feature vector according to the binary feature mask, generates privacy protection feature data after adding noise disturbance to the key feature subset, and uploads the privacy protection feature data to an edge aggregation node. In step S500, the edge aggregation node receives the privacy protection feature data uploaded by multiple terminals, performs cross-device feature alignment operation on the privacy protection feature data to generate an aligned feature data set. In step S600, weighted fusion processing is performed on the aligned feature data set to generate a global feature space matrix. In step S700, the global feature space matrix is used to analyze the space-time correlation of devices to generate a network resource strategy containing channel allocation parameters and terminal scheduling sequence. In step S800, the base station controller executes the network resource strategy and broadcasts updated channel state parameters to the terminal device.

[0026] In this embodiment, the standardized feature vector refers to the unified scale data formed after the original network state data is subjected to time domain statistical feature extraction and frequency domain energy feature conversion. Specifically, methods such as time window segmentation to calculate mean and variance, Fourier transform to extract principal component energy, etc. can be used to eliminate the dimensional differences of different features. Feature importance analysis refers to the degree of influence on network performance prediction by interfering with a specific feature dimension and evaluating its impact. Specifically, the method of comparing the prediction result deviation after shielding part of the feature can be used to identify the feature dimension that has a significant impact on resource allocation decision. Binary feature mask refers to a selector that dynamically determines the selection of transmittable feature dimensions according to channel capacity constraints. Specifically, the method of truncating low-value dimensions after sorting by importance can be used to generate, which is used to prioritize the transmission of high-value features under limited bandwidth. Privacy-protected feature data refers to the data form of adding noise to the screened key features. Specifically, the method of generating perturbation values that meet the differential privacy requirements using Laplace mechanism can be used to prevent the original feature information from being reverse cracked during transmission. Cross-device feature alignment refers to the operation of unifying the feature dimensions and order of data uploaded by different terminals. Specifically, the method of feature name matching and historical mean filling missing dimensions can be used to eliminate the influence of device heterogeneity on global analysis. Global feature space matrix refers to the unified representation formed by aggregating multi-terminal feature data. Specifically, the method of sample size weighted fusion and standard deviation normalization can be used to generate, which is used to fully reflect the spatio-temporal distribution characteristics of network state. Device spatio-temporal correlation analysis refers to the operation of mining the communication relationship between devices combined with geographic location and service type. Specifically, the method of Euclidean distance matrix construction and clustering analysis can be used to identify device groups with strong correlation to optimize channel allocation.

[0027] In this embodiment, the terminal device first collects local signal strength, transmission delay and other raw data, extracts time domain statistical features through time window segmentation, and extracts frequency domain energy features using Fourier transform. After normalization processing, the standardized feature vector is formed, which eliminates the dimensional difference of different features. Then, by shielding part of the feature dimension and evaluating its impact on network throughput prediction, a vector reflecting the importance of each dimension is generated. According to the current channel capacity information issued by the base station, the maximum number of transmittable feature dimensions is calculated, and the key dimensions are selected according to the importance and sorted to generate a binary mask. Using the mask to extract the feature subset and superimposing noise that meets the privacy protection requirements, irreversible perturbation data is uploaded. After receiving the multi-terminal data, the edge aggregation node adjusts the dimension order through feature name matching, fills in the missing dimensions using historical mean, and forms a unified structure of the aligned dataset. The aligned data is subjected to sample weight fusion and normalization processing to generate a global feature matrix. Combined with the device geographical location, the spatial distance matrix is calculated, the spatiotemporal correlation matrix is constructed by superimposing the business type similarity, and the device grouping is divided and the channel resources are allocated through clustering analysis. According to the business emergency level, the scheduling priority sequence is generated to form the network strategy containing the channel allocation proportion and the scheduling order. After the base station executes the strategy, the updated channel state parameters are broadcast to the terminal device to form a closed-loop optimization.

[0028] In this embodiment, the scheme significantly reduces the amount of transmission data through dynamic feature screening, performs feature perturbation on the terminal side to ensure that the original data does not leave the local device, accurately identifies the device grouping features through global feature fusion and clustering analysis, improves channel utilization, adjusts the feature selection threshold according to the real-time channel capacity, and balances the transmission efficiency and information integrity. In this way, the application effectively reduces the communication burden between the terminal and the base station and avoids transmitting redundant feature data. Under the premise of protecting user privacy, the spatiotemporal distribution law of the network state is accurately captured through cross-device feature alignment and fusion. The channel allocation strategy based on device grouping features improves the spectral resource utilization rate in dense scenarios. The dynamically adjusted feature sparsification mechanism enables the system to adapt to changes in channel capacity, maintaining the real-time and accuracy of resource allocation. The synergistic effect of noise perturbation and feature screening reduces the risk of privacy leakage while ensuring the effective transmission of key information.

[0029] In a possible implementation, with reference to Figure 2 , step S100 includes steps S110-S140, wherein: In step S110, the terminal device collects signal strength raw data, transmission delay raw data and device energy consumption raw data of the local network. In step S120, the raw data is segmented with a preset time window, and time domain statistical features are calculated within the time window. Step S130, Fourier transform is performed on the data in the time window to extract the frequency domain principal component energy feature; Step S140, maximum and minimum normalization is respectively performed on the time domain statistical feature and the frequency domain principal component energy feature, and the normalized time domain statistical feature and the frequency domain principal component energy feature are spliced to generate the standardized feature vector.

[0030] In this embodiment, the preset time window refers to an interval divided according to a fixed time length for the continuously collected original data, and can be specifically implemented by using a sliding window mechanism, for example, the data is divided into 5-second or 10-second segments for capturing the dynamic changes of the network state. The time domain statistical feature refers to statistical indexes such as mean, variance and extreme value calculated for the data in the time window, and can be specifically implemented by using a sliding average algorithm or difference calculation, for representing the short-term fluctuation characteristics of signal strength, transmission delay and device energy consumption. The Fourier transform refers to a mathematical method for converting a time domain signal into a frequency domain signal, and can be specifically implemented by using a fast Fourier transform algorithm, for extracting periodic components in the signal. The frequency domain principal component energy feature refers to a feature vector of a frequency energy concentrated region selected by principal component analysis, and can be specifically implemented by using a covariance matrix decomposition method, for representing the energy distribution characteristics of the signal in the frequency domain. The maximum and minimum normalization refers to a standardization method for linearly mapping a feature value to the interval [0, 1], and can be specifically implemented by calculating the proportional relationship between the maximum value and the minimum value of the feature dimension, for eliminating the dimension difference of different features.

[0031] In this embodiment, after the signal strength, transmission delay and device energy consumption data are divided into time windows of a fixed length, statistical features are calculated in the time domain to capture the short-term fluctuation law, and frequency domain energy distribution features are extracted by Fourier transform to reflect the periodic change mode. The time domain features and the frequency domain features are spliced after normalization to form a multi-dimensional standardized feature vector. For example, the mean and variance of the signal strength are calculated in the time window, and the frequency domain principal component energy proportion is extracted, and the two are normalized and combined into a unified feature representation. This processing method can simultaneously retain the time-varying characteristics and periodicity of the network state, and provides a multi-dimensional data basis for subsequent feature selection.

[0032] In this embodiment, the scheme can more completely describe the space-time variation law of network state through time-frequency dual-domain feature fusion, and can eliminate the dimensional difference of multi-source data through normalization processing, improve the comparability of features and the stability of subsequent processing, and solve the problem of single dimension of traditional network state feature extraction and insufficient representation ability. Through the cooperative extraction of time domain statistical features and frequency domain energy features, the representation ability of network state data is enhanced, and a more comprehensive data basis is provided for subsequent feature sparsification processing. At the same time, the normalization operation reduces the compatibility processing difficulty of multi-source heterogeneous data, avoids the interference of different dimensional features on model training, and improves the accuracy of network resource optimization decision.

[0033] In one possible implementation, with reference to Figure 3 , step S200 includes steps S210-S240, in which: Step S210, copying the standardized feature vector to generate a test vector copy, and setting the specified feature dimension value in the test vector copy to zero to generate a shielded feature vector; Step S220, calculating the network throughput prediction value corresponding to the shielded feature vector based on the preset network performance prediction function; Step S230, calculating the deviation amount of the reference network throughput prediction value of the standardized feature vector and the network throughput prediction value of the shielded feature vector; Step S240, generating an initial sequence of feature importance according to the deviation amount, and performing sliding window smoothing processing on the initial sequence of feature importance to generate the feature importance vector.

[0034] In this embodiment, the test vector copy refers to a temporary data copy generated by copying the original feature vector, which can be implemented by memory copying, and is used to isolate the influence of the zero setting operation on the original data. The shielded feature vector refers to the feature data after setting the specified feature dimension to zero, which can be implemented by assigning a zero value to the target dimension through index positioning, and is used to simulate the change of network performance when the feature dimension is missing. The network performance prediction function refers to a mathematical model for evaluating network throughput, which can be implemented by a regression model based on gradient boosting tree, and outputs the predicted throughput by inputting the feature vector. The deviation amount refers to the difference value between the reference prediction value and the shielded prediction value, which can be calculated by absolute difference or relative percentage, and is used to quantify the influence of the feature dimension on the network performance. The sliding window smoothing processing refers to the operation of local averaging of sequence data, which can be implemented by a fixed window length mean filtering algorithm, and is used to eliminate random fluctuations in the feature importance evaluation process.

[0035] In the feature importance analysis process in this embodiment, a test copy is first generated by copying the standardized feature vector to ensure the integrity of the original data. By sequentially setting different feature dimensions in the test copy to zero, a plurality of shielded feature vectors are generated and input into the network performance prediction function to calculate the corresponding network throughput prediction values. The baseline prediction value of the original feature vector is compared with each shielded prediction value to form an initial sequence reflecting the importance of each feature dimension. Due to the dynamic characteristics of the wireless network environment, short-term fluctuations may exist in the initial sequence. By using a sliding window smoothing process, the evaluation results of adjacent multiple periods are averaged to finally generate a stable and reliable feature importance vector.

[0036] In this embodiment, the scheme can dynamically capture the actual influence of different features on network performance through the quantitative evaluation mechanism of shielded feature dimensions, effectively suppresses transient interference by combining sliding window processing, makes the feature importance evaluation results more suitable for real-time network environment, can accurately identify the key feature dimensions that affect network resource allocation, and avoids the waste of communication resources caused by the transmission of redundant feature data. By dynamically adjusting the feature importance evaluation results, the network state differences of different terminal devices can be adapted, providing a reliable basis for subsequent sparsification processing, thereby improving the effectiveness of the overall resource optimization strategy.

[0037] In a feasible implementation manner, referring to Figure 4 , step S300 includes steps S310-S340, wherein: Step S310, parsing the available transmission bit number from the channel capacity indication information, and calculating the single feature dimension average encoding bit quantity based on the locally stored historical transmission record; Step S320, calculating the maximum transmittable dimension number according to the available transmission bit number and the single feature dimension average encoding bit quantity; Step S330, sorting the feature dimensions in descending order according to the feature importance vector, selecting the top K dimensions as the reserved feature dimension set, and K is equal to the maximum transmittable dimension number; Step S340, constructing a binary mask matrix, setting the corresponding positions of the reserved feature dimension set to 1 and the remaining positions to 0 to generate the binary feature mask.

[0038] In this embodiment, the channel capacity indication information refers to the quantized numerical index of the current available wireless transmission resource quantity issued by the base station through the control channel. Specifically, it can be realized by mapping the modulation and coding scheme index value carried by the physical layer control signaling to the bit number. Its function is to provide the terminal device with the data transmission capacity constraint under the current channel condition. The single feature dimension average encoding bit quantity refers to the statistical mean of the actual consumed encoding bit number of a single feature dimension in the historical transmission process. Specifically, it can be realized by using a sliding window to count the actual used bit number of each feature dimension in the last N transmissions and calculating the average value. Its function is to provide a quantitative basis for dynamically adjusting the feature transmission number. The maximum transmittable dimension number refers to the upper limit of the number of feature dimensions allowed to be transmitted under the current channel capacity. Specifically, it can be realized by using the available transmission bit number divided by the single feature dimension average encoding bit quantity and taking the integer part. Its function is to ensure that the transmission data quantity does not exceed the channel carrying capacity. The feature importance vector descending order sorting refers to arranging the feature dimensions from high to low according to the influence degree of the feature on the network performance prediction. Specifically, it can be realized by using the quicksort algorithm to sort the feature importance vector. Its function is to preferentially retain the features that have greater influence on the resource optimization decision.

[0039] In this embodiment, after the channel capacity indication information is parsed, the terminal device queries the locally stored historical transmission records to calculate the average encoding bit quantity consumed by each feature dimension in the past transmission process. For example, if the historical records show that a certain feature dimension needs to be encoded with an average of 3 bits, and the current available transmission bit number is 120 bits, then the maximum transmittable dimension number is 40. Subsequently, the feature importance vector is sorted, for example, a feature importance vector contains the weight values of 50 dimensions, and after descending order arrangement, the first 40 dimensions are selected as the reserved set. In the generation process of the binary mask matrix, the matrix positions corresponding to the reserved feature dimensions are set to 1, and the remaining positions are set to 0. For example, a binary sequence with a length of 50 is generated, the first 40 bits are 1, and the last 10 bits are 0.

[0040] Compared with the prior art, the traditional method usually adopts a fixed feature selection strategy or ignores the dynamic change of channel capacity, resulting in that the amount of transmitted data exceeds the carrying capacity of the channel or important features are missed. For example, the prior art can only select the first 30 dimensions according to the importance of the features, but cannot fully utilize the bandwidth when the channel capacity is sufficient, or is forced to discard key features when the capacity is insufficient. The scheme dynamically calculates the maximum number of transmittable dimensions, combines the real-time channel capacity and the importance ranking of the features, and realizes the adaptive matching of the number of transmitted features and the channel conditions. Through the above technical scheme, the technical problem of how to balance the feature transmission integrity and the communication resource limitation under the limited channel capacity is solved. By dynamically calculating the maximum number of transmittable dimensions, it is ensured that the amount of transmitted data strictly conforms to the current channel capacity constraint; by using the dimension selection mechanism based on the importance ranking of the features, the features that have greater influence on resource optimization decision are preferentially transmitted, and the channel resources are avoided from being occupied by invalid features. For example, in the channel capacity fluctuation scene, when the number of available bits decreases, the number of transmission dimensions is automatically reduced, but the features with high importance ranking are always retained, so that the communication overhead is reduced while the decision accuracy is maintained.

[0041] In a feasible implementation manner, referring to Figure 5 , the step S400 comprises steps S410-S430, wherein: In step S410, the noise intensity coefficient mapping table of the feature type of the key feature subset is queried to obtain the noise intensity coefficient of the current feature, and a Laplace distribution random number is generated based on the noise intensity coefficient; In step S420, the random number is superimposed on the numerical part of the key feature subset; In step S430, boundary truncation processing is performed on the superimposed feature value to generate the privacy protection feature data.

[0042] In this embodiment, the noise intensity coefficient mapping table refers to a query table containing the correspondence between the feature type and the noise intensity, which can be constructed by offline training or expert experience, and is used to dynamically adjust the noise intensity according to the difference in privacy sensitivity of different features. The Laplace distribution random number refers to a random variable conforming to the Laplace probability density function, which can be generated by inverse transformation method or table lookup method, and its probability distribution characteristics can meet the differential privacy protection requirement. The boundary truncation processing refers to the threshold limiting operation on the feature value exceeding the preset numerical range, which can be realized by using piecewise linear function or saturation function, to prevent the feature value from exceeding the reasonable range after superimposing the noise and causing data distortion.

[0043] In this embodiment, before uploading the feature data, first match the corresponding noise intensity coefficient in the preset mapping table according to the feature type, for example, the signal intensity feature can be configured with a smaller coefficient to maintain data availability, and the energy consumption feature can be configured with a larger coefficient to enhance privacy protection. Based on the matched coefficient, generate Laplace noise, for example, for a feature dimension with a coefficient of 0.3, the noise value can come from a Laplace distribution with a scale parameter of 0.3. After superimposing the noise value on the original feature value, through boundary truncation, the values exceeding the preset upper and lower limits are forced to be constrained to the boundary value, for example, the feature values exceeding the [0, 1] interval are set to 0 or 1, to ensure that the processed data meets the privacy requirements and retains effective statistical characteristics.

[0044] In this embodiment, the scheme realizes dynamic balance between privacy protection strength and data availability through a feature type adaptive noise intensity adjustment mechanism, and effectively suppresses abnormal deviation of feature values caused by noise interference through a boundary truncation operation. It can realize fine-grained privacy protection under the federated learning framework, apply differential noise disturbance to different feature types, avoid privacy leakage or data distortion risk caused by uniform noise intensity, and maintain the effective distribution form of feature data through boundary truncation, providing a reliable data foundation for subsequent cross-device feature alignment and global resource optimization.

[0045] In a feasible implementation manner, referring to Figure 6 , step S500 includes steps S510-S530, wherein: Step S510, analyze the dimension description information of the privacy protection feature data uploaded by each terminal, extract the correspondence between the feature name and the dimension index; perform cross-terminal dimension matching according to the feature name similarity, and generate a terminal feature dimension mapping relationship table; Step S520, adjust the dimension order of the privacy protection feature data of each terminal according to the terminal feature dimension mapping relationship table; Step S530, detect the missing dimension position in the adjusted data, and fill the missing dimension position with the global feature historical mean value of the nearest preset first number of historical data acquisition periods, to generate the aligned feature data set.

[0046] In this embodiment, the dimension description information refers to metadata containing feature name, data type and dimension position index, which can be implemented in JSON format or protocol buffer, and is used to identify the semantic information of different terminal feature vectors. Cross-terminal dimension matching refers to establishing the correspondence between features of different devices by calculating the text similarity of feature names, which can be implemented by using cosine similarity algorithm or edit distance algorithm, and is used to solve the problem of dimension misalignment caused by different terminal feature naming rules. The global feature historical mean refers to the average value of all terminal feature dimensions in the previous period stored by the aggregation node, which can be dynamically updated through the sliding window mechanism, and is used to provide a reasonable filling reference value when the feature is missing.

[0047] In this embodiment, in the cross-device feature alignment process, the feature metadata carried in the data uploaded by each terminal is first parsed to extract the mapping relationship between the feature name and the dimension position. By calculating the text similarity between the feature names of different terminals, such as converting the feature names into vectors by using the word vector model and then calculating the cosine similarity, a cross-terminal feature correspondence table is established. According to the relationship table, the dimension order of each terminal feature vector is rearranged so that the same semantic features are in the same dimension position in different terminal data. For the problem of feature missing caused by device heterogeneity, such as the failure of a terminal to collect data from a specific sensor, the global historical mean of the feature dimension in the last several periods is used for filling, such as the mean of the last three historical data collection periods, so as to ensure the integrity of the data set.

[0048] In this embodiment, the scheme effectively eliminates the problem of inconsistent data dimensions caused by device heterogeneity, improves the semantic consistency of the federated feature space, solves the difficulty of dimension alignment caused by inconsistent naming rules of multi-terminal features, and solves the problem of feature missing caused by differences in device sensor configuration, ensuring that the subsequent weighted fusion operation is performed in a unified dimension space, improving the integrity and availability of the global feature matrix, and providing an accurate data basis for network resource policy generation.

[0049] In a feasible implementation manner, with reference to Figure 7 , step S600 includes steps S610-S630, in which: In step S610, the number of local data samples of each terminal is counted, and the aggregation weight coefficient of each terminal is determined according to the number of local data samples and the total number of samples; In step S620, weighted calculation is performed on the feature values of each feature dimension in the aligned feature data set; In step S630, standard deviation normalization processing is performed on the weighted calculation result, and the normalized values of all dimensions are integrated to generate the global feature space matrix.

[0050] In this embodiment, the number of local data samples refers to the total number of data entries collected by the terminal device locally and participating in feature generation, which can be realized by counting the sample quantity field in each terminal uploaded data packet, and is used to reflect the data contribution scale of different terminals. The aggregation weight coefficient refers to the proportion parameter used to adjust the influence of each terminal data in global fusion, which can be calculated by the ratio of the number of local samples to the sum of the number of samples of all terminals, so that the terminal with large data quantity occupies higher weight in the global feature. Weighted calculation refers to multiplying each feature dimension data value of each terminal by its corresponding weight coefficient and then accumulating, which can be realized by matrix multiplication operation, ensuring that different size data sources have a reasonable impact on the global feature. Standard deviation normalization processing refers to scaling operation based on statistical distribution of weighted feature values, which can be realized by performing linear transformation of each feature value by dividing the standard deviation after calculating the global standard deviation, eliminating the dimensional difference between different feature dimensions.

[0051] In this embodiment, after the edge aggregation node completes the cross-device feature alignment, the number of local data samples uploaded by each terminal in this aggregation period is first extracted by traversing all the privacy protection feature data uploaded by the terminal device. The total sample quantity is obtained by adding the sample quantity of each terminal, and the aggregation weight coefficient of a single terminal is obtained by dividing the sample quantity of the terminal by the total sample quantity. For each feature dimension in the aligned feature data set, the feature values of each terminal in that dimension are multiplied by the corresponding weight coefficient and then accumulated to obtain the weighted fusion value of that dimension. The standard deviation of the weighted fusion value of all terminals in that dimension is calculated, and the weighted fusion value of each dimension is divided by the corresponding standard deviation to obtain the normalized value. The normalized values of all feature dimensions are arranged and combined in the original dimension order to form a feature space matrix containing global statistical characteristics.

[0052] In this embodiment, by dynamically calculating the weight coefficient based on the data sample quantity, the data contribution of each terminal can be accurately reflected, and the excessive interference of small data quantity terminals on the global feature can be avoided. At the same time, the standard deviation normalization processing is introduced, which effectively eliminates the fusion deviation caused by the different dimensions of different features, and significantly improves the data consistency of the global feature space compared with the direct splicing feature method in the prior art. In this way, the application can dynamically adjust the fusion weight according to the actual data size of the terminal, improve the representation accuracy of the global feature space to the network resource distribution, and solve the feature deviation problem caused by the traditional average fusion method in the terminal data imbalance scene. By standard deviation normalization processing, the statistical distribution difference between multi-dimensional features is eliminated, the stability of the global feature matrix in subsequent resource allocation decisions is enhanced, and the risk of policy shock caused by different feature scales is reduced.

[0053] In a feasible implementation manner, the reference Figure 8The step S700 includes steps S710-S750, wherein: In step S710, geographical position coordinate data reported by a terminal device is acquired, and a Euclidean distance matrix between terminal devices is calculated based on the geographical position coordinate data. In step S720, a space-time correlation matrix is constructed according to the Euclidean distance matrix and a device service type. In step S730, clustering analysis is performed on the space-time correlation matrix, communication device grouping is identified, and a channel bandwidth ratio is assigned to each group based on the communication device grouping. In step S740, a terminal scheduling priority sequence is generated according to a service urgency degree. In step S750, the channel bandwidth allocation ratio and the terminal scheduling priority sequence are fused to generate the network resource policy.

[0054] In this embodiment, the geographical position coordinate data refers to longitude and latitude coordinates collected by a terminal device through a built-in positioning module, which can be implemented by using a GPS module or a base station triangulation technology, and is used to determine the physical position relationship between devices to provide basic data for subsequent calculation of spatial correlation. The Euclidean distance matrix refers to a symmetric matrix generated by calculating the square root of the sum of squares of the differences in geographical position coordinates between two devices, which can be implemented by using a matrix operation library, and is used to quantify the spatial distance between devices and reflect the potential signal interference degree between devices. The space-time correlation matrix refers to a multi-dimensional matrix formed by weighting and splicing the Euclidean distance matrix and the device service type features, which can be implemented by using a feature splicing and normalization method, and is used to comprehensively represent the correlation of devices in spatial distribution and service demand. The communication device grouping refers to a device set formed by classifying the space-time correlation matrix by using a clustering algorithm, which can be implemented by using a K-means or hierarchical clustering algorithm, and is used to identify a device group with similar communication demand to provide grouping basis for channel allocation. The channel bandwidth ratio refers to a channel resource allocation weight dynamically calculated according to the number of devices in a group and the service type, which can be implemented by using a proportional allocation algorithm, and is used to ensure that the resource demands of different groups are differentiated to meet. The terminal scheduling priority sequence refers to a device communication order list generated by sorting according to a service urgency degree, which can be implemented by using a priority queue data structure, and is used to prioritize the transmission demand of high-urgency services when the resources are limited.

[0055] In this embodiment, after the terminal device uploads the collected geographic position coordinates to the edge aggregation node, it first generates an Euclidean distance matrix through coordinate difference calculation, which reflects the spatial distribution density between devices. Further combining device service type data, such as real-time video transmission, Internet of Things sensor data, etc., the service type is encoded into a feature vector and weightedly fused with the distance matrix to form a space-time correlation matrix. Through clustering algorithm analysis on the space-time correlation matrix, devices with similar spatial distribution and service demand are divided into the same communication group. Based on the number of devices in each group and the data throughput demand of the service type, the channel bandwidth proportion allocated to each group is dynamically calculated. At the same time, according to the emergency parameter defined in the device service type, such as medical monitoring devices with the highest priority, a terminal scheduling sequence is generated. Finally, the channel allocation proportion and the scheduling sequence are fused to form a network resource strategy that takes into account the spatial resource utilization efficiency and service timeliness.

[0056] It can be understood that in the prior art, resource allocation is generally based on signal strength or service type alone, without fully considering the influence of device spatial distribution on channel interference. For example, the prior art may only generate a scheduling sequence according to the service emergency of the device, but does not solve the problem of transmission efficiency decline caused by channel competition of devices with similar geographic positions. However, the present scheme jointly analyzes the physical location and service characteristics of the device by constructing a space-time correlation matrix, so that the channel allocation can not only avoid signal interference of spatially adjacent devices, but also match the resource demand differences of different service types, thereby effectively reducing the channel conflict probability caused by spatially adjacent devices, and improving the rationality of channel resource allocation. At the same time, by combining the service emergency to generate a scheduling sequence, high-priority services can be ensured to be transmitted in time, and the overall network service quality is optimized in a resource-limited scenario. In addition, the channel allocation strategy based on clustering grouping can dynamically adapt to changes in device distribution, enhancing the robustness of the resource management scheme.

[0057] In a feasible implementation manner, referring to Figure 9 , the method further includes steps S910-S930, wherein: Step S910, calculating the cosine similarity between the network resource strategies of the consecutive preset second number of decision periods; Step S920, when the cosine similarity is less than a preset threshold, generating a sparsity adjustment instruction according to the cosine similarity and sending it to the terminal device; Step S930, the terminal device updates the sliding window size in the feature importance analysis according to the sparsity adjustment instruction.

[0058] In this embodiment, the cosine similarity refers to measuring the similarity between different periodic network resource policies through a vector space model, which can be implemented by calculating the ratio of the inner product of vectors to the product of their lengths, and is used to detect the degree of change in the policy. The preset second number refers to the number of consecutive periods for policy similarity comparison, which can be dynamically adjusted according to the network environment, for example, set to 3-5 periods. The sparsity adjustment instruction refers to a parameter update instruction triggered according to the change in policy similarity, which can generate a binary trigger signal through a preset threshold comparison, and is used to control the sensitivity of feature importance analysis. The size of the sliding window refers to the length of the time window used for smoothing in feature importance analysis, which can use a ring buffer to store historical data, and adjust the window length to change the stability of feature importance evaluation.

[0059] In this embodiment, in a plurality of consecutive decision periods, the vector representation of the network resource policy is extracted and similarity calculation is performed. When the policy similarity is lower than the preset threshold, it indicates that the network environment or device state has changed significantly, and an adjustment instruction is sent to the terminal device at this time. After receiving the instruction, the terminal device dynamically increases or decreases the size of the sliding window in the feature importance analysis, thereby changing the smoothing degree of feature evaluation. For example, when the similarity is low, the window size is increased to enhance the stability, and when the similarity is high, the window size is decreased to improve the response speed.

[0060] In some specific embodiments, the cosine similarity calculation can use a sliding window mechanism, which retains the policy vectors of the last N periods each time. The generation of the sparsity adjustment instruction can be combined with the similarity change gradient, and the adjustment is triggered when the gradient exceeds the set range. The update of the sliding window size can set an incremental step, for example, increase or decrease the window length by 10% each time, while setting maximum and minimum value constraints.

[0061] Compared with the prior art, the traditional method uses a fixed parameter feature analysis window, which cannot adapt to dynamic network environment changes. In the prior art, the size of the sliding window is usually set to an empirical value, which can cause the feature importance evaluation to be inaccurate when the network state suddenly changes. The present scheme realizes the automatic adaptation of the feature analysis parameters through the policy similarity dynamic feedback mechanism, solves the problem of mismatch between static parameters and dynamic environment, and enables the present application to automatically optimize the feature analysis parameters according to the change trend of the network resource policy, while ensuring the stability of the feature importance evaluation and improving the response speed to the network environment changes. This dynamic adjustment mechanism effectively avoids the problem of resource allocation policy lag caused by environmental mutation, improves the real-time performance and accuracy of wireless network resource optimization, and reduces the additional communication overhead caused by parameter mismatch.

[0062] In the embodiments of the present application, the wireless network resource joint optimization method based on federal feature sparsification can reduce the amount of transmission data through feature sparsification processing, protect privacy combined with noise disturbance, and dynamically generate resource allocation strategies based on global feature analysis, which can improve resource allocation efficiency and reduce communication overhead, while improving privacy protection capability.

[0063] It should be noted that the above examples are only for understanding the present application and do not constitute a limitation on the wireless network resource joint optimization method based on federal feature sparsification of the present application. More forms of simple transformation based on this technical concept are within the protection scope of the present application.

[0064] The present application also provides a wireless network resource joint optimization system based on federal feature sparsification, which refers to Figure 10 , the system comprises a memory 10, a processor 20, and a wireless network resource joint optimization program based on federal feature sparsification stored on the memory 10 and executable on the processor 20, which is configured to implement the steps of the wireless network resource joint optimization method based on federal feature sparsification.

[0065] The wireless network resource joint optimization system based on federal feature sparsification provided by the present application adopts the wireless network resource joint optimization method based on federal feature sparsification in the above embodiments, which can improve resource allocation efficiency and reduce communication overhead. Compared with the prior art, the wireless network resource joint optimization system based on federal feature sparsification provided by the present application has the same beneficial effects as the wireless network resource joint optimization method based on federal feature sparsification provided by the above embodiments, and other technical features in the wireless network resource joint optimization system based on federal feature sparsification are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0066] It should be understood that parts of the present application can be realized by hardware, software, firmware or their combination. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0067] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any equivalent structural transformation made by using the contents of the present application specification and drawings, or direct / indirect application in other related technical fields is included in the patent protection scope of the present application.

Claims

1. A joint optimization method for wireless network resources based on federated feature sparsity, characterized in that, The method includes: The terminal device collects the raw dataset of local network status and performs feature extraction and standardization on the raw dataset to generate a standardized feature vector; The terminal device performs a feature importance analysis operation based on the standardized feature vector to generate a feature importance vector; The terminal device receives channel capacity indication information sent by the base station, and performs a sparse mask generation operation in combination with the feature importance vector to generate a binary feature mask. The terminal device extracts a subset of key features from the standardized feature vector based on the binary feature mask, adds noise perturbation to the subset of key features to generate privacy-preserving feature data, and uploads the privacy-preserving feature data to the edge aggregation node. The edge aggregation node receives the privacy protection feature data uploaded by multiple terminals, performs cross-device feature alignment operation on the privacy protection feature data, and generates an aligned feature dataset; A weighted fusion process is performed on the aligned feature dataset to generate a global feature space matrix; Based on the global feature space matrix analysis of the device spatiotemporal correlation, a network resource strategy including channel allocation parameters and terminal scheduling sequence is generated. The base station controller executes the network resource policy and broadcasts the updated channel state parameters to the terminal devices.

2. The joint optimization method for wireless network resources based on federated feature sparsity as described in claim 1, characterized in that, The steps of the terminal device collecting the original dataset of local network status and performing feature extraction and standardization on the original dataset to generate standardized feature vectors include: Terminal devices collect raw data on local network signal strength, transmission delay, and device power consumption. The original data is segmented using a preset time window, and time-domain statistical features are calculated within the time window; Perform Fourier transform on the data within the time window to extract the energy features of the principal components in the frequency domain; The time-domain statistical features and the frequency-domain principal component energy features are respectively subjected to minimum-maximum normalization, and the normalized time-domain statistical features and frequency-domain principal component energy features are concatenated to generate the standardized feature vector.

3. The joint optimization method for wireless network resources based on federated feature sparsity as described in claim 1, characterized in that, The steps by which the terminal device performs feature importance analysis based on the standardized feature vector to generate the feature importance vector include: The standardized feature vector is copied to generate a test vector copy, and the specified feature dimension value in the test vector copy is set to zero to generate a masked feature vector; Based on a preset network performance prediction function, calculate the predicted network throughput value corresponding to the shielding feature vector; Calculate the deviation between the baseline network throughput prediction value of the standardized feature vector and the network throughput prediction value of the masked feature vector; An initial sequence of feature importance is generated based on the deviation, and a sliding window smoothing process is performed on the initial sequence of feature importance to generate the feature importance vector.

4. The joint optimization method for wireless network resources based on federated feature sparsity as described in claim 1, characterized in that, The terminal device receives channel capacity indication information from the base station and performs a sparse mask generation operation in conjunction with the feature importance vector. The steps for generating a binary feature mask include: The available transmission bits are parsed from the channel capacity indication information, and the average number of coded bits per feature dimension is calculated based on the locally stored historical transmission records. The maximum number of transmittable dimensions is calculated based on the available number of transmittable bits and the average number of encoded bits per single feature dimension. The feature dimensions are sorted in descending order according to the feature importance vector, and the top K dimensions are selected as the set of retained feature dimensions, where K is equal to the maximum number of transmittable dimensions. Construct a binary mask matrix by setting the corresponding positions of the reserved feature dimension set to 1 and the rest to 0 to generate the binary feature mask.

5. The joint optimization method for wireless network resources based on federated feature sparsity as described in claim 1, characterized in that, The steps for generating privacy-preserving feature data by adding noise perturbation to the subset of key features include: Based on the feature type of the key feature subset, query the preset noise intensity coefficient mapping table to obtain the noise intensity coefficient of the current feature, and generate a Laplace distribution random number based on the noise intensity coefficient. The random number is superimposed onto the numerical portion of the key feature subset; Boundary truncation is performed on the superimposed feature values ​​to generate the privacy-preserving feature data.

6. The joint optimization method for wireless network resources based on federated feature sparsity as described in claim 1, characterized in that, The edge aggregation node receives the privacy-preserving feature data uploaded by multiple terminals, and performs cross-device feature alignment on the privacy-preserving feature data to generate an aligned feature dataset. The steps include: Parse the dimensional description information of the privacy protection feature data uploaded by each terminal, and extract the correspondence between feature names and dimensional indices; perform cross-terminal dimensional matching based on feature name similarity to generate a terminal feature dimension mapping table; Adjust the dimensional order of privacy protection feature data for each terminal according to the terminal feature dimension mapping relationship table; The missing dimension locations in the adjusted data are detected, and the missing dimension locations are filled with the historical average of global features from the most recent preset first number of historical data collection periods to generate the aligned feature dataset.

7. The joint optimization method for wireless network resources based on federated feature sparsity as described in claim 1, characterized in that, The steps of performing weighted fusion processing on the aligned feature dataset to generate a global feature space matrix include: The number of local data samples for each terminal is counted, and the aggregation weight coefficient for each terminal is determined based on the number of local data samples and the total number of samples. Perform a weighted calculation on the feature values ​​of each feature dimension in the aligned feature dataset; The weighted calculation results are normalized by standard deviation, and the normalized values ​​of all dimensions are integrated to generate the global feature space matrix.

8. The joint optimization method for wireless network resources based on federated feature sparsity as described in claim 1, characterized in that, The steps for generating a network resource strategy that includes channel allocation parameters and terminal scheduling sequences based on the analysis of the spatiotemporal correlation of devices using the global feature space matrix include: Obtain the geographic location coordinate data reported by the terminal device, and calculate the Euclidean distance matrix between the terminal devices based on the geographic location coordinate data; Construct a spatiotemporal correlation matrix based on the Euclidean distance matrix and the device service type; Cluster analysis of the spatiotemporal correlation matrix is ​​performed to identify communication device groups, and a channel bandwidth ratio is allocated to each group based on the communication device groups; Generate a terminal scheduling priority sequence based on the urgency of the service; The network resource strategy is generated by integrating the channel bandwidth allocation ratio and the terminal scheduling priority sequence.

9. The joint optimization method for wireless network resources based on federated feature sparsity as described in claim 3, characterized in that, The method further includes: Calculate the cosine similarity between the network resource strategies for a second consecutive predetermined number of decision periods; When the cosine similarity is less than a preset threshold, a sparsity adjustment instruction is generated based on the cosine similarity and sent to the terminal device; The terminal device updates the sliding window size in the feature importance analysis according to the sparsity adjustment instruction.

10. A joint optimization system for wireless network resources based on federated feature sparsity, characterized in that, The system includes: a memory, a processor, and a federated feature sparsity-based wireless network resource joint optimization program stored in the memory and executable on the processor, the federated feature sparsity-based wireless network resource joint optimization program being configured to implement the steps of the federated feature sparsity-based wireless network resource joint optimization method as described in any one of claims 1 to 9.

Images