Search

CN-122002321-A - Communication service quality intelligent optimization method based on deep learning

CN122002321ACN 122002321 ACN122002321 ACN 122002321ACN-122002321-A

Abstract

The invention discloses a communication service quality intelligent optimization method based on deep learning, which comprises the following steps of S1, collecting data and preprocessing, constructing a service quality sequence, S2, executing feature reconstruction operation on the service quality sequence, extracting multi-layer feature construction quality feature sets, S3, generating a quality state sequence based on the quality feature sets through a Mamba model, S4, carrying out separated mapping on the quality state sequence, constructing a quality prediction sequence, S5, screening a quality abnormal point set based on the quality prediction sequence, S6, constructing an optimization objective function according to the quality abnormal point set, adopting an A3C algorithm to generate an adjustment strategy set, S7, mapping the adjustment strategy set into an instruction sequence, executing, and updating Mamba model and A3C algorithm parameters. The invention integrates ResNet D network, mamba model and A3C algorithm, and has the advantages of high identification accuracy, high response speed and stable optimization effect.

Inventors

  • HU FENGYUAN
  • YANG LIFANG

Assignees

  • 郑州大学

Dates

Publication Date
20260508
Application Date
20260310

Claims (10)

  1. 1. The intelligent optimization method for the communication service quality based on the deep learning is characterized by comprising the following steps of: S1, collecting and preprocessing user terminal data and communication service data in a wireless local area network, and constructing a service quality sequence according to time steps; s2, performing feature reconstruction operation on the service quality sequence, dividing a reconstruction result into a plurality of time sequence fragments by adopting a sliding time window with a fixed length, and performing multi-layer feature extraction operation on the time sequence fragments through ResNet D network to construct a quality feature set; S3, performing state recursion and time sequence modeling operation on the quality feature set through a Mamba model, and generating a quality state sequence by adopting a selective gating mechanism; s4, constructing a prediction framework based on an MLP-Mixer structure, and respectively executing a separated nonlinear mapping operation on the quality state sequence along the time dimension and the characteristic dimension to construct a quality prediction sequence; S5, based on the quality prediction sequence, constructing a reconstruction error vector of each prediction time step in the variation self-encoder, calculating confidence deviation, and screening a quality abnormal point set according to the difference between the confidence deviation and a preset reference deviation; s6, constructing an optimization objective function according to the quality anomaly point set, and generating an adjustment strategy set by adopting an A3C algorithm and combining the existing historical scheduling data; And S7, mapping the adjustment strategy set into an instruction sequence and executing, and updating Mamba the parameters of the model and the A3C algorithm according to the execution feedback data.
  2. 2. The intelligent optimization method of communication service quality based on deep learning according to claim 1, wherein the communication service data includes channel utilization, average delay, packet loss rate, jitter value and throughput, the user terminal data represents equipment status information, connection behavior information and data usage related to the communication service, the preprocessing includes deletion filling, anomaly rejection, format standardization, numerical normalization and time alignment operations, the history scheduling data represents service quality adjustment record and parameter configuration results executed in the communication system, and the execution process of the Mamba model and the A3C algorithm is completed in a high-end router, and the high-end router has hardware acceleration capability and a local service quality status storage function.
  3. 3. The intelligent optimization method for communication service quality based on deep learning according to claim 1, wherein the step S2 specifically comprises: S21, performing feature reconstruction operation on the service quality sequence according to a fixed splicing template, performing splicing processing on communication service data and user terminal data of each time step in the service quality sequence, and performing length cutting or zero filling operation on a splicing result to construct a unified length combined feature sequence; S22, adopting a sliding time window with a fixed length, carrying out segmentation treatment on the combined characteristic sequence according to a set sliding step length, and intercepting the combined characteristic vectors of continuous time steps in each window to form a plurality of time sequence fragments; s23, performing tensor expansion and dimension arrangement operation on each time sequence segment, and constructing a two-dimensional tensor comprising a time dimension and a characteristic dimension; S24, inputting each two-dimensional tensor into ResNet D network, sequentially executing multi-round one-dimensional convolution, nonlinear activation and residual error connection operation, extracting characteristic representations at different depths, and taking the residual error connection result of each round as the characteristic tensor output by the round; S25, performing channel stitching and dimension compression operation on the feature tensors, stitching all the feature tensors of the two-dimensional tensors according to feature dimensions, and performing affine transformation operation on the stitching results by using the full-connection layer to generate quality feature vectors with uniform lengths; s26, arranging all the quality feature vectors according to the sequence of the time sequence segments to form a quality feature set.
  4. 4. The intelligent optimization method for communication service quality based on deep learning according to claim 3, wherein the step S24 specifically comprises: s241, taking the time dimension of each two-dimensional tensor as the unfolding direction of one-dimensional convolution processing, setting the convolution kernel size and the step length, and executing one-dimensional convolution operation of the round on the two-dimensional tensor to extract convolution response representing local time dependency relationship; S242, performing nonlinear activation processing on the convolution response, and performing element-by-element transformation on the convolution response by adopting a ReLU function; S243, performing residual connection operation, performing weighted superposition operation on the activation result and the feature tensor output in the previous round according to a preset weight set, generating the feature tensor of the current round, and taking a preset initial zero vector as the feature tensor output in the previous round in the first round; s244, continuing to execute multi-round one-dimensional convolution, nonlinear activation and residual error connection operation until the number of preset iteration rounds is reached; S245, performing unified dimension checking operation on all the feature tensors, screening abnormal tensors with dimensions inconsistent with the set standard dimension, and eliminating the abnormal tensors.
  5. 5. The intelligent optimization method for communication service quality based on deep learning according to claim 1, wherein the step S3 specifically comprises: S31, arranging the quality feature vectors in time sequence to construct a time-continuous quality feature sequence; s32, initializing a state recurrence structure in a Mamba model, and setting a state recurrence step length to be consistent with the time step of the quality feature sequence; S33, for each time step, extracting a quality characteristic vector corresponding to the current time step and a quality state vector of the previous time step, inputting an extraction result into a state recurrence structure to execute state update calculation, and generating an intermediate state representation of the current time step; s34, performing channel weight calculation operation on each characteristic channel in the intermediate state representation based on a preset gating weight calculation mode through a selective gating mechanism, and performing channel-by-channel weighted modulation operation on each channel value to generate a gating state representation of the current time step as a quality state vector; S35, repeatedly executing the state updating and weighted modulation operation until the state recursion operation of the time steps is completed, summarizing the quality state vectors corresponding to the time steps, and constructing a quality state sequence consistent with the time steps of the quality feature sequence.
  6. 6. The intelligent optimization method for communication service quality based on deep learning according to claim 5, wherein the step S33 specifically comprises: S331, extracting a quality feature vector of a current time step and a quality state vector of a last time step, respectively serving as feature representation and historical state information, performing feature dimension consistency verification, and if the feature dimension consistency verification is not passed, performing structure adjustment operation; s332, performing dynamic weight mapping operation on the extraction result subjected to consistency verification to generate a characteristic response vector and a state response vector, wherein the method specifically comprises the following steps: In the feature dimension, splicing the feature representation subjected to consistency verification with the history state information to generate a joint representation; Based on the joint representation, performing dynamic weight calculation operation in the current time step, and performing matrix operation and nonlinear transformation operation on the joint representation through a preset multi-layer mapping structure to generate a weight matrix and a bias vector of the current time step, wherein the multi-layer mapping structure comprises GELU function activation layers and a low-rank decomposition structure; according to the weight matrix and the bias vector of the current time step, matrix multiplication and bias superposition operations are respectively carried out on the characteristic representation and the historical state information, so that a characteristic response vector and a state response vector of the current time step are obtained; S333, performing element-by-element addition operation on the characteristic response vector and the state response vector according to dimensions to generate a fusion response vector; S334, performing nonlinear activation operation on the fusion response vector, and performing element-by-element calculation on all values in the fusion response vector by adopting a ReLU function to form an intermediate state representation of the current time step.
  7. 7. The intelligent optimization method for communication service quality based on deep learning according to claim 1, wherein the step S4 specifically comprises: s41, constructing a prediction framework based on an MLP-Mixer structure, taking a quality state sequence as input, and respectively constructing independent separation mapping paths along a time dimension and a feature dimension, wherein the separation mapping paths comprise a time mapping path and a feature mapping path; S42, constructing a cross-time-step aggregation structure based on state values of a plurality of time steps in the same characteristic dimension in a quality state sequence in a time mapping path, executing aggregation mapping operation on the aggregation structure based on a sliding time window, and performing linear transformation and ReLU function activation processing through a multi-layer perceptron to form a time mapping result reflecting the cross-time-step association relation; S43, in the feature mapping path, performing feature mixing operation on all feature values corresponding to the same time step in the quality state sequence, and performing channel-by-channel mapping and GELU function transformation operation on each feature value by adopting a multi-layer perceptron to form a feature mapping result reflecting the internal association relation of feature dimensions; s44, performing element-by-element fusion operation on the time mapping result and the feature mapping result at the corresponding time step position, and performing consistency correction processing on the fusion result to generate a prediction state vector; s45, arranging the prediction state vectors in time sequence, and constructing a quality prediction sequence consistent with the time length of the quality state sequence.
  8. 8. The intelligent optimization method for communication service quality based on deep learning according to claim 7, wherein the construction process of the prediction framework specifically comprises the following steps: Setting structural parameters for constructing a time mapping path and a feature mapping path, wherein the structural parameters comprise the mapping layer number, the feature dimension width of each layer, the activation function type and a linear transformation mode; In the time mapping path, arranging all quality state vectors according to the time step sequence of the quality state sequence to construct an aggregation structure, and setting the length and the step length of a sliding time window; In each time window, executing state value aggregation operation on a time dimension, distributing non-uniform weights to quality state vectors of different time steps in an aggregation structure, and executing weighted aggregation operation to form an aggregation result sequence with a time bias characteristic; Sequentially executing multi-layer perceptron processing operation on the aggregation result sequence, and sequentially completing linear transformation and nonlinear activation processing on each layer, wherein the nonlinear activation processing represents element-by-element calculation on linear transformation results by adopting a ReLU function; In the feature mapping path, dividing a quality state vector corresponding to each time step into a plurality of channel vectors according to feature dimensions, and setting channel mapping depth and parameter configuration of each layer of the multi-layer perceptron; Performing channel-by-channel multi-layer perceptron mapping operation on each channel vector, and performing nonlinear activation on a mapping result by adopting GELU functions, and simultaneously maintaining the same arrangement sequence as the quality state vector; setting a dimension alignment rule at the tail end of the separation mapping path, wherein the dimension alignment rule is used for adjusting the output results of the time mapping path and the feature mapping path into a consistent tensor structure; and after parameter setting, path construction, mapping processing and dimension alignment operation are completed, the construction of the prediction framework is completed.
  9. 9. The intelligent optimization method for communication service quality based on deep learning according to claim 1, wherein the step S5 specifically comprises: S51, expanding a quality prediction sequence according to time steps, executing a prediction state reconstruction operation in each time step, inputting a prediction state vector of a current time step into a decoder of a variable self-encoder, generating a corresponding reconstruction state vector, calculating a vector difference value between the current prediction state vector and the reconstruction state vector, and generating a reconstruction error vector of the current time step, wherein the generation process of the reconstruction state vector specifically comprises the following steps: Affine transformation is carried out on the prediction state vector through a set full-connection layer in the decoder to obtain an intermediate implicit vector; Performing element-by-element activation processing on the intermediate implicit vector by adopting GELU functions; Performing dimension expansion and structure reconstruction operation on the activation result to generate a reconstruction state vector consistent with the dimension of the prediction state vector; S52, performing confidence measure extraction operation on the reconstruction error vector of each time step, counting absolute differences of all feature dimensions in the reconstruction error vector, calculating an average value of all the absolute differences as the time step confidence deviation value, and arranging all the confidence deviation values according to time sequence to construct a confidence deviation sequence; And S53, calculating the difference value between the confidence deviation sequence and the set reference deviation sequence according to time steps, extracting time step indexes of which the difference value result exceeds a preset error threshold value to form an abnormal index set, and carrying out section merging processing on the abnormal indexes continuously appearing in adjacent time steps to generate a quality abnormal point set.
  10. 10. The intelligent optimization method for communication service quality based on deep learning according to claim 1, wherein the step S6 specifically comprises: s61, matching the quality abnormal point set with a quality prediction sequence, extracting prediction state vectors corresponding to different time steps, constructing an abnormal state set, and combining quality state vectors corresponding to a plurality of time steps before and after each abnormal time step to construct a context state window; s62, setting multidimensional target parameters according to the abnormal state set and the context state window, and constructing an optimized target function, wherein the multidimensional target parameters comprise service quality mutation amplitude, state fluctuation frequency, resource use efficiency and response time delay distribution characteristics; S63, setting a state space, an action space and a reward function based on an optimization objective function by adopting an A3C algorithm, and executing strategy learning operation to generate a plurality of candidate adjustment strategies; s64, scoring and screening operations are carried out on the candidate adjustment strategies by combining the historical scheduling data, and a strategy candidate pool is formed by screening a plurality of candidate adjustment strategies with scores larger than a preset scoring threshold; S65, executing multiple rounds of asynchronous updating operation in the strategy candidate pool, executing simulation execution and feedback evaluation operation on each candidate adjustment strategy, and iteratively updating parameters of the A3C algorithm according to feedback evaluation results until the feedback evaluation results reach preset expected conditions; s66, regenerating a strategy candidate pool through the updated A3C algorithm to form an adjustment strategy set.

Description

Communication service quality intelligent optimization method based on deep learning Technical Field The invention relates to the technical field of communication service quality, in particular to an intelligent optimization method for communication service quality based on deep learning. Background In the existing communication network system, evaluation and optimization of service quality are key tasks for guaranteeing user experience. The conventional method generally relies on manually set threshold rules or models based on simple statistical indexes, such as single or combined indexes of average delay, packet loss rate, channel utilization rate and the like, to judge service states, and performs parameter adjustment based on empirical strategies. However, due to a large number of nonlinear relations, dynamic change characteristics and multidimensional interference factors in a communication environment, the fixed rule method is difficult to effectively adapt to service quality fluctuation under different service scenes, the problems of response lag and inflexibility in parameter tuning often occur, and fine control and intelligent scheduling are difficult to realize. In recent years, some researches attempt to introduce a machine learning model to model communication data so as to improve the quality of service prediction and optimization capability. However, the method mostly adopts static feature input and a shallow network structure, lacks modeling capability of time-sequence-dependent features, and cannot fully mine the change rule of the communication state in the time dimension. Meanwhile, the existing method generally ignores the deep association between the state of the user terminal and the state of the network service in the communication network, and lacks a joint analysis path from the terminal side to the network side, so that the model generalization capability and the prediction accuracy are insufficient. In addition, a common control strategy in the existing communication optimization scheme is mainly based on a preset rule or a traditional Q learning structure in reinforcement learning, lacks robust modeling capability and a context sensing mechanism for complex abnormal states, and is difficult to effectively identify multidimensional anomalies and make self-adaptive optimization decisions. In terms of model feedback mechanisms, closed-loop optimization paths based on execution effects are also commonly lacking, resulting in difficulty in continuous adjustment and evolution of optimization strategies. Therefore, how to provide a communication service quality intelligent optimization method based on deep learning is a problem to be solved by those skilled in the art. Disclosure of Invention The invention provides a communication service quality intelligent optimization method based on deep learning, which integrates user terminal state data and multidimensional communication service indexes in a wireless local area network, combines ResNet D network to perform feature extraction, models service quality time sequence state by Mamba model, performs cross-dimension prediction by using MLP-Mixer structure, identifies abnormal quality fragments by a variation self-encoder, finally generates and optimizes an adjustment strategy based on A3C algorithm, realizes dynamic prediction, self-adaptive optimization and feedback closed loop execution of service quality, and has the advantages of high identification precision, high response speed and stable optimization effect. According to the embodiment of the invention, the intelligent optimization method for the communication service quality based on deep learning comprises the following steps: S1, collecting and preprocessing user terminal data and communication service data in a wireless local area network, and constructing a service quality sequence according to time steps; s2, performing feature reconstruction operation on the service quality sequence, dividing a reconstruction result into a plurality of time sequence fragments by adopting a sliding time window with a fixed length, and performing multi-layer feature extraction operation on the time sequence fragments through ResNet D network to construct a quality feature set; S3, performing state recursion and time sequence modeling operation on the quality feature set through a Mamba model, and generating a quality state sequence by adopting a selective gating mechanism; s4, constructing a prediction framework based on an MLP-Mixer structure, and respectively executing a separated nonlinear mapping operation on the quality state sequence along the time dimension and the characteristic dimension to construct a quality prediction sequence; S5, based on the quality prediction sequence, constructing a reconstruction error vector of each prediction time step in the variation self-encoder, calculating confidence deviation, and screening a quality abnormal point set according to the difference betwe