A method and system for intelligent performance perception and prediction for Wi-Fi

Abstract

This invention discloses an intelligent performance perception and prediction method and system for Wi-Fi, belonging to the field of communication technology. The method constructs a high-order tensor based on collected user signal data, with user, performance index, and time as dimensions, and further constructs a low-order matrix. Combining a deep reinforcement learning algorithm, the high-order tensor and low-order matrix are decomposed and predicted separately, constructing a tensor prediction model and a matrix prediction model. By adjusting the fusion weights through error feedback, the prediction results of the two models are weighted and fused to generate the final prediction result. This invention can effectively capture the nonlinear and sudden changes in Wi-Fi signals, improve the accuracy and generalization ability of network performance perception and prediction, provide optimization basis for frequency band switching, load balancing, and bandwidth scheduling, and improve the overall Wi-Fi service quality and user experience.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, and more specifically, relates to an intelligent performance perception and prediction method and system for Wi-Fi. Background Technology

[0002] With the development of wireless network technology and the widespread adoption of smart devices, Wi-Fi has become an indispensable part of people's daily lives. However, Wi-Fi signals are affected by various factors, such as physical obstacles, electronic interference, and user density, leading to fluctuating performance. Therefore, sensing and predicting future trends in Wi-Fi signals can help network administrators more efficiently identify and resolve potential Wi-Fi service problems, effectively improving the overall Wi-Fi user experience. This technology has attracted widespread attention from academia and industry and has been applied in fields such as smart homes, smart cities, and the Industrial Internet of Things (IIoT).

[0003] Current research on Wi-Fi signal sensing and prediction mainly focuses on propagation model design, link budget analysis, and measured data correction. Propagation model design considers the impact of factors such as distance, frequency, antenna height, and surrounding environment on signal strength. Link budget analysis allows researchers to calculate the maximum path loss that communication systems can withstand based on given signal data. Measured data correction relies on actual signal strength data obtained from field measurements to adjust theoretical models and improve their accuracy. However, these methods are mostly based on traditional communication theory and mathematical statistics, resulting in low efficiency and practicality in real-world applications. Therefore, there is an urgent need to develop more efficient methods and systems to more accurately sense and predict changes in Wi-Fi signals.

[0004] Against this backdrop, artificial intelligence (AI) technology, with its advantages in complex pattern recognition, big data processing, and adaptive learning, has brought new opportunities for breakthroughs in Wi-Fi signal perception and prediction. However, most existing solutions focus on using long and short time series forecasting and deep learning methods such as recurrent neural networks for Wi-Fi signal prediction. Although these methods have shown strong capabilities in processing time series data, they still have limitations in dealing with the sudden changes and nonlinear characteristics of Wi-Fi signals. In addition, these methods struggle to effectively uncover deep correlations between data, resulting in low accuracy in Wi-Fi signal perception and prediction. Summary of the Invention

[0005] To address the shortcomings of related technologies, the present invention aims to provide an intelligent performance perception and prediction method and system for Wi-Fi, which addresses the limitations of existing methods in handling sudden changes, nonlinear characteristics, and cross-dimensional index correlations of Wi-Fi signals, resulting in low accuracy in Wi-Fi performance perception and prediction.

[0006] To achieve the above objectives, in a first aspect, the present invention provides an intelligent performance perception and prediction method for Wi-Fi, comprising:

[0007] S100. Collect raw signal data of Wi-Fi users, obtain corresponding feature values ​​according to the network performance indicators of each user, and construct high-order tensor data in the user dimension based on the feature values; wherein, the feature values ​​are in matrix form, the rows of the matrix represent different network performance indicators, and the columns represent different sampling time slots;

[0008] S200. Based on the high-order tensor data, normalize the various network performance indicators, calculate the cross-time slot weights of the various network performance indicators, and integrate the multi-dimensional network performance indicators composed of multiple network performance indicators corresponding to multiple sampling time slots into one-dimensional standardized data and form a low-order matrix.

[0009] S300: Input the high-order tensor data into a pre-trained tensor prediction model to obtain a high-order prediction tensor, and input the low-order matrix data into a pre-trained matrix prediction model to obtain a low-order prediction matrix; wherein, the tensor prediction model and the matrix prediction model are trained based on a deep reinforcement learning algorithm.

[0010] S400. Obtain the high-order prediction error based on the high-order prediction tensor and the high-order tensor data, obtain the low-order prediction error based on the low-order prediction matrix and the low-order matrix data, dynamically adjust the fusion weights based on the high-order prediction error and the low-order prediction error, and fuse the prediction results of the tensor prediction model and the matrix prediction model by weighted averaging according to the adjusted fusion weights to obtain the fused prediction result.

[0011] Optionally, step S100 includes:

[0012] S110. Capture signal interaction information between user equipment and access point in a Wi-Fi network environment, and collect raw signal data of Wi-Fi users during different time periods of network usage;

[0013] S120. For each independent Wi-Fi user, using network performance indicators as rows and sampling time slots as columns, convert the different network performance indicator data in the original signal data into a matrix, which serves as the feature value matrix of the network performance indicator for the corresponding user.

[0014] S130. Treat each user's feature value matrix as an independent substructure, and stack and combine them along the user dimension to form high-order tensor data. .

[0015] Optionally, step S200 includes:

[0016] S210, Transfer the high-order tensor data The elements of each eigenvalue matrix are normalized, and the normalized result is... Represented as:

[0017]

[0018] in, This represents the column vector in the eigenvalue matrix, where the column vector is... , Represents column vectors The One element, column vector The maximum value in;

[0019] S220. Construct a vector based on the normalized values ​​of each network performance metric across all sampling time slots. Calculate the weight of each network performance metric across all sampling time slots. ;in, This represents the total number of network performance metrics. This represents the total number of sampling time slots. Indicates the first The network performance metric in the first Normalized values ​​over each sampling time slot;

[0020] S230. Integrate multi-dimensional network performance metrics into one-dimensional standardized data to construct a low-order matrix. Among them, the low-order matrix Corresponding element The expression is: , Indicates the first The weights of each network performance metric Indicates the first The network performance metric in the first Normalized values ​​over each sampling time slot.

[0021] Optionally, step S300 includes:

[0022] S310. Tensor decomposition based on deep reinforcement learning algorithm is used to process the high-order tensor data. The model parameters of the initial tensor prediction model are constructed by decomposing the data into higher-order tensor data. The optimization objective is to minimize the prediction error between the predicted tensor and the predicted tensor of the initial tensor prediction model. The decomposition method of tensor decomposition is continuously adjusted through strategy iteration, and the model parameters of the initial tensor prediction model are updated. The initial tensor prediction model is then trained based on the updated model parameters until the network converges or reaches the preset number of iterations, thus obtaining the tensor prediction model.

[0023] S320. The low-order matrix is ​​decomposed using matrix factorization based on a deep reinforcement learning algorithm to construct the model parameters of the initial matrix prediction model, using the low-order matrix... Minimizing the error between the initial matrix prediction model and the predicted matrix is ​​the optimization objective. The matrix decomposition method is continuously adjusted through strategy iteration to update the model parameters of the initial matrix prediction model. The initial matrix prediction model is then trained based on the updated model parameters until the network converges or reaches the preset number of iterations to obtain the matrix prediction model.

[0024] S330. Input the high-order tensor data into a pre-trained tensor prediction model to obtain a high-order prediction tensor, and input the low-order matrix data into a pre-trained matrix prediction model to obtain a low-order prediction matrix.

[0025] Optionally, step S310 includes:

[0026] S311. Initialize the experience replay pool, randomly initialize the main network, and set the parameters of the main network... Copy the data to the target network; input the higher-order tensor data into the initialization tensor prediction model to obtain the higher-order prediction state set. ,make Get the state Determine the set of optional actions ,in, Indicates at time To implement the selected action for tensor decomposition;

[0027] S312, with probability Randomly select an action, to The probability is used to select the optimal action based on the main network's valuation. ;in, ;

[0028] S313, The optimal action obtained from S312 Receive the corresponding reward value To reach the next state , will vector Place them into the experience replay pool respectively; if the experience replay pool is full, proceed to step S314; if the experience replay pool is not full, then... Then proceed to step S312;

[0029] S314. Randomly select a batch of vectors from the experience replay pool, and input each vector into the main network. Find the action corresponding to the largest Q value in the main network and input it into the target network along with the above vectors to obtain the target network Q value under the new mode. Calculate the loss function using the obtained main network Q value and target network Q value, and update the parameters of the tensor prediction model using gradient descent. probability;

[0030] S315. Repeat step S314, and change the parameter every time interval T. The algorithm is copied to the target network of the tensor prediction model until the network converges or reaches the preset number of iterations, thus obtaining the optimal tensor decomposition strategy and tensor prediction model.

[0031] Optionally, step S320 includes:

[0032] S321. Initialize the experience replay pool, randomly initialize the main network, and set the parameters of the main network. Copy to the target network; input the low-order matrix into the initialization matrix prediction model to obtain the low-order prediction state set. ,make Get the state Determine the set of optional actions ,in, Indicates at time The selected action to achieve matrix decomposition;

[0033] S322, with probability Randomly select an action, to The probability is used to select the optimal action based on the main network's valuation. ;in, ;

[0034] S323, The optimal action obtained from S322 Receive the corresponding reward value To reach the next state , will vector Place them into the experience replay pool respectively; if the experience replay pool is full, proceed to step S324; if the experience replay pool is not full, then... And proceed to step S322;

[0035] S324. Randomly select a batch of vectors from the experience replay pool, and input each vector into the main network. Find the action corresponding to the largest Q value in the main network and input it into the target network along with the above vectors to obtain the target network Q value under the new model. Calculate the loss function using the obtained main network Q value and target network Q value, and update the parameters of the matrix prediction model using gradient descent. ;

[0036] S325. Repeat step S324, and change the parameter every time interval T. The algorithm is copied to the target network of the matrix prediction model until the network converges or reaches the preset number of iterations, thus obtaining the optimal matrix decomposition strategy and matrix prediction model.

[0037] Optionally, step S400 includes:

[0038] S410. Perform a restoration operation on the prediction result of the low-order prediction matrix based on the original dimension information index and cross-time slot weight of the low-order matrix to obtain a restored tensor with the same structure as the high-order tensor.

[0039] S420. Based on the errors of the matrix prediction model and the tensor prediction model, determine the preliminary fusion weights of the matrix prediction model and the tensor prediction model. The preliminary fusion weights are expressed as follows:

[0040]

[0041] in, and The matrix prediction model and the tensor prediction model are respectively at time 10:00. The error, It is a smoothing coefficient used to avoid division by zero and to smooth out errors;

[0042] S430. The preliminary fusion weights are normalized to calculate the dynamic fusion weight coefficients:

[0043]

[0044] in, The influence coefficient determines the relative influence between the errors of the two models; when When the tensor prediction model and the matrix prediction model have smaller errors, the fusion weights are set higher than those for the model with larger errors; when At that time, the fusion weights for the models with larger errors in the tensor prediction model and matrix prediction model are set to be greater than those for the models with smaller errors; the fusion weight coefficients satisfy the following: ;

[0045] S440. The higher-order prediction tensor is weighted and fused with the restored tensor obtained from the lower-order matrix to obtain the fused prediction result as the prediction output tensor; the weighted fusion formula is as follows:

[0046]

[0047] in, For the tensor prediction model at time 1000 The prediction tensor, Based on time The reconstructed tensor obtained from the low-order matrix prediction results This is the result of the fusion prediction.

[0048] Secondly, the present invention provides an intelligent performance sensing and prediction system for Wi-Fi, comprising:

[0049] The collection module is used to collect raw signal data from Wi-Fi users, obtain corresponding feature values ​​based on the network performance indicators of each user, and construct high-order tensor data in the user dimension based on the feature values; wherein, the feature values ​​are in matrix form, the rows of the matrix represent different network performance indicators, and the columns represent different sampling time slots;

[0050] The processing module is used to normalize various network performance indicators based on the high-order tensor data, calculate the cross-time slot weights of various network performance indicators, and integrate the multi-dimensional network performance indicators composed of multiple network performance indicators corresponding to multiple sampling time slots into one-dimensional standardized data and form a low-order matrix.

[0051] The prediction module is used to input the high-order tensor data into a pre-trained tensor prediction model to obtain a high-order prediction tensor, and to input the low-order matrix data into a pre-trained matrix prediction model to obtain a low-order prediction matrix; wherein, the tensor prediction model and the matrix prediction model are trained based on a deep reinforcement learning algorithm;

[0052] The integration module is used to obtain high-order prediction error based on high-order prediction tensor and the high-order tensor data, obtain low-order prediction error based on low-order prediction matrix and low-order matrix data, dynamically adjust the fusion weight based on the high-order prediction error and low-order prediction error, and fuse the prediction results of tensor prediction model and matrix prediction model by weighted averaging according to the adjusted fusion weight to obtain fused prediction result.

[0053] Thirdly, the present invention provides an electronic device, including a memory and a processor; the memory is used to store computer-executed instructions, and the processor is connected to the memory via a bus;

[0054] When the Wi-Fi-oriented intelligent performance perception and prediction system is running, the processor executes the computer execution instructions stored in the memory to cause the Wi-Fi-oriented intelligent performance perception and prediction system to perform the Wi-Fi-oriented intelligent performance perception and prediction method as described in any one of the first aspects.

[0055] Fourthly, the present invention provides a computer-readable storage medium comprising computer-executable instructions that, when executed on a computer, cause the computer to perform the intelligent performance sensing and prediction method for Wi-Fi as described in any one of the first aspects.

[0056] Compared with the prior art, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0057] 1. This invention provides an intelligent performance perception and prediction method for Wi-Fi. By constructing a three-dimensional high-order tensor and a normalized low-order matrix with user, performance index, and time series dimensions, it fully preserves the spatiotemporal characteristics and cross-user behavioral differences of Wi-Fi signals. Secondly, it employs a prediction strategy guided by deep reinforcement learning to replace the static model training mechanism, enabling it to adapt to rapid fluctuations in channel conditions. Combining tensor decomposition and matrix decomposition techniques, the decomposition method is continuously adjusted through strategy iteration to learn the optimal decomposition strategy. This allows the model to deeply explore the structural correlations between performance indicators through an interactive strategy optimization process during training, thereby improving prediction accuracy and generalization ability. This solves the technical problem of low accuracy in Wi-Fi signal perception and prediction due to limitations in dealing with sudden changes and nonlinear characteristics of Wi-Fi signals, effectively improving the prediction accuracy of network performance and the efficiency of resource allocation, thus better meeting users' stringent requirements for network service quality.

[0058] 2. This invention provides an intelligent performance perception and prediction method for Wi-Fi. It employs a dynamic adjustment mechanism to flexibly adjust the weights of various factors in prediction and decision-making based on different network scenarios and user needs. By dynamically adjusting the fusion weights based on prediction error feedback, it fuses high-order prediction tensors with low-order prediction matrices to achieve efficient and accurate prediction of Wi-Fi network performance indicators and service quality. This enables the system to more accurately perceive and predict performance and make optimal decisions when facing complex and ever-changing network environments. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating an intelligent performance perception and prediction method for Wi-Fi provided in an embodiment of the present invention.

[0060] Figure 2 This is a high-order tensor prediction model architecture diagram of an intelligent performance perception and prediction method for Wi-Fi provided in an embodiment of the present invention.

[0061] Figure 3 This is a low-order matrix prediction model architecture diagram of an intelligent performance perception and prediction method for Wi-Fi provided in an embodiment of the present invention.

[0062] Figure 4 This is a block diagram of an intelligent performance perception and prediction method for Wi-Fi provided in an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0064] The following description, in conjunction with a preferred embodiment, illustrates the content involved in the above embodiments.

[0065] Example 1

[0066] like Figure 1 As shown, this invention provides an intelligent performance perception and prediction method for Wi-Fi, including:

[0067] S100. Collect raw signal data of Wi-Fi users, obtain corresponding feature values ​​according to the network performance indicators of each user, and construct high-order tensor data in the user dimension based on the feature values; wherein, the feature values ​​are in matrix form, the rows of the matrix represent different network performance indicators, and the columns represent different sampling time slots;

[0068] S200. Based on the high-order tensor data, normalize the various network performance indicators, calculate the cross-time slot weights of the various network performance indicators, and integrate the multi-dimensional network performance indicators composed of multiple network performance indicators corresponding to multiple sampling time slots into one-dimensional standardized data and form a low-order matrix.

[0069] S300: Input the high-order tensor data into a pre-trained tensor prediction model to obtain a high-order prediction tensor, and input the low-order matrix data into a pre-trained matrix prediction model to obtain a low-order prediction matrix; wherein, the tensor prediction model and the matrix prediction model are trained based on a deep reinforcement learning algorithm.

[0070] S400. Obtain the high-order prediction error based on the high-order prediction tensor and the high-order tensor data, obtain the low-order prediction error based on the low-order prediction matrix and the low-order matrix data, dynamically adjust the fusion weights based on the high-order prediction error and the low-order prediction error, and fuse the prediction results of the tensor prediction model and the matrix prediction model by weighted averaging according to the adjusted fusion weights to obtain the fused prediction result.

[0071] To address the limitations of traditional methods in handling sudden changes, nonlinear features, and mining deep correlations between data, this paper proposes an innovative system architecture that combines multidimensional modeling with intelligent strategies. In the intelligent performance perception and prediction method for Wi-Fi provided in this application, tensor prediction models and matrix prediction models are constructed in parallel, and deep reinforcement learning-based training mechanisms are introduced for each. The optimal decomposition strategy is autonomously learned for different data, and finally, a dynamic fusion strategy based on prediction error feedback is combined to achieve collaborative prediction. This improves the accuracy and generalization ability of the prediction model, enabling efficient and accurate prediction of Wi-Fi network performance indicators and service quality. This provides a basis for the rational allocation of Wi-Fi resources and can further enhance the overall user experience.

[0072] Optionally, step S100 includes:

[0073] S110. Capture signal interaction information between user equipment and access point in a Wi-Fi network environment, and collect raw signal data of Wi-Fi users during different time periods of network usage;

[0074] S120. For each independent Wi-Fi user, using network performance indicators as rows and sampling time slots as columns, convert the different network performance indicator data in the original signal data into a matrix, which serves as the feature value matrix of the network performance indicator for the corresponding user.

[0075] S130. Treat each user's feature value matrix as an independent substructure, and stack and combine them along the user dimension to form high-order tensor data. .

[0076] Data is collected using network monitoring equipment and software systems to comprehensively capture signal interaction information between user devices and access points in a Wi-Fi network environment.

[0077] In terms of time, a long-term continuous monitoring approach is adopted, with sampling at second-level intervals to ensure coverage of different time periods of network usage and comprehensively capture data characteristics under varying network load conditions. In terms of users, by identifying and classifying network access devices, a wide range of common Wi-Fi terminal devices are covered, such as smartphones, tablets, laptops, and smart IoT devices, thereby obtaining diverse user data samples and enhancing the representativeness and universality of the data.

[0078] After data collection is complete, for each individual user, the collected network performance metrics data are transformed into a matrix structure with "metrics as rows and time as columns." Specifically, different network performance metrics serve as the row elements of the matrix; for example, signal strength data is the first row, transmission rate data as the second row, latency data as the third row, packet loss rate data as the fourth row, and so on, ensuring that each network performance metric has a unique corresponding row in the matrix. The columns of the matrix are determined by different sampling time slots. According to a pre-set fixed time interval sampling strategy, the sampled data, arranged in chronological order, is sequentially filled into the matrix columns.

[0079] After constructing the feature value matrix for each user, the matrices of multiple users are integrated to form a higher-order tensor. In higher-order tensors, each user's matrix is ​​treated as an independent substructure and combined in new dimensions, thereby enabling efficient organization and management of large-scale Wi-Fi user data.

[0080] Optionally, step S200 includes:

[0081] S210, Transfer the high-order tensor data The elements of each eigenvalue matrix are normalized, and the normalized result is... Represented as:

[0082]

[0083] in, This represents the column vector in the eigenvalue matrix, where the column vector is... , Represents column vectors The One element, column vector The maximum value in;

[0084] S220. Construct a vector based on the normalized values ​​of each network performance metric across all sampling time slots. Calculate the weight of each network performance metric across all sampling time slots. ;in, This represents the total number of network performance metrics. This represents the total number of sampling time slots. Indicates the first The network performance metric in the first Normalized values ​​over each sampling time slot;

[0085] S230. Integrate multi-dimensional network performance metrics into one-dimensional standardized data to construct a low-order matrix. Among them, the low-order matrix The corresponding element expression is: , Indicates the first The weights of each network performance metric Indicates the first The network performance metric in the first Normalized values ​​over each sampling time slot.

[0086] After acquiring high-order tensor data, various network performance metrics are normalized. Combined with a weighted scoring mechanism, a comprehensive performance score for each sampling time slot is extracted. Each sampling time slot corresponds to multiple performance metrics (such as signal strength, rate, and latency), and the performance metric data from multiple time slots constitute a multi-dimensional vector. The normalized average of each performance metric across all sampling time slots is calculated as the cross-time slot weight for that metric. The normalized values ​​of each metric are then weighted and summed in each time slot according to their corresponding weights to obtain a one-dimensional rating sequence for a user. Stacking all user rating sequences along the user dimension forms a low-order matrix with users as rows and time as columns, simplifying the data structure and facilitating subsequent analysis.

[0087] By normalizing the elements of the aforementioned eigenvalue matrix, the data in this network performance metric dimension is mapped to an interval. This ensures the comparability of data of different scales and dimensions, avoiding the impact of large differences in data magnitude on the accuracy of subsequent analysis. The above normalization operation is performed on the column vectors of all network performance indicator dimensions in the higher-order tensor, ensuring that the entire data is on a uniform standard scale.

[0088] After data normalization, the influence weights of each network performance indicator are determined based on its average performance over the entire time series, thus generating a low-order matrix. The multi-dimensional network performance indicators are then integrated into one-dimensional standardized data to construct the low-order matrix. This makes the data structure more compact and can more efficiently reflect the comprehensive characteristics of network performance.

[0089] Optionally, step S300 includes:

[0090] S310. Tensor decomposition based on deep reinforcement learning algorithm is used to process the high-order tensor data. The model parameters of the initial tensor prediction model are constructed by decomposing the data into higher-order tensor data. The optimization objective is to minimize the prediction error between the predicted tensor and the predicted tensor of the initial tensor prediction model. The decomposition method of tensor decomposition is continuously adjusted through strategy iteration, and the model parameters of the initial tensor prediction model are updated. The initial tensor prediction model is then trained based on the updated model parameters until the network converges or reaches the preset number of iterations, thus obtaining the tensor prediction model.

[0091] S320. The low-order matrix is ​​decomposed using matrix factorization based on a deep reinforcement learning algorithm to construct the model parameters of the initial matrix prediction model, using the low-order matrix... Minimizing the error between the initial matrix prediction model and the predicted matrix is ​​the optimization objective. The matrix decomposition method is continuously adjusted through strategy iteration to update the model parameters of the initial matrix prediction model. The initial matrix prediction model is then trained based on the updated model parameters until the network converges or reaches the preset number of iterations to obtain the matrix prediction model.

[0092] S330. Input the high-order tensor data into a pre-trained tensor prediction model to obtain a high-order prediction tensor, and input the low-order matrix data into a pre-trained matrix prediction model to obtain a low-order prediction matrix.

[0093] This scheme employs the Deep Q-Network (DQN) framework, combining deep learning neural networks with Q-learning in reinforcement learning. It seeks the optimal decomposition strategy to achieve accurate predictions by maximizing cumulative reward. This method requires defining the basic triples: state, action, and reward. It utilizes the DQN algorithm from deep reinforcement learning, combined with tensor decomposition and matrix decomposition techniques. The optimization objective is to minimize the error between the original tensor or matrix and its corresponding prediction result. The decomposition method is continuously adjusted through policy iteration to learn the optimal strategy. Since both high-order tensors and low-order matrices are implemented using deep reinforcement learning algorithms, the only difference lies in the triple settings. High-order tensors employ CP-WOPT, CP-NLS, and Tucker decomposition rules, while low-order matrices employ Adagrad, RMSProp, and Adam decomposition rules.

[0094] Step S310 includes:

[0095] S311. Initialize the experience replay pool, randomly initialize the main network, and copy the parameters of the main network to the target network; input the higher-order tensor data into the initialized tensor prediction model to obtain the higher-order prediction state set. ,make Get the state Determine the set of optional actions ,in, Indicates at time To implement the selected action for tensor decomposition;

[0096] S312, with probability Randomly select an action, to The probability is used to select the optimal action based on the main network's valuation. ;in, ;

[0097] S313, The optimal action obtained from S312 Receive the corresponding reward value To reach the next state , will vector Place them into the experience replay pool respectively; if the experience replay pool is full, proceed to step S314; if the experience replay pool is not full, then... Then proceed to step S312;

[0098] S314. Randomly select a batch of vectors from the experience replay pool, and input each vector into the main network. Find the action corresponding to the largest Q value in the main network and input it into the target network along with the above vectors to obtain the target network Q value under the new mode. Calculate the loss function using the obtained main network Q value and target network Q value, and update the parameters of the tensor prediction model using gradient descent. ;

[0099] S315. Repeat step S314, and change the parameter every time interval T. The algorithm is copied to the target network of the tensor prediction model until the network converges or reaches the preset number of iterations, thus obtaining the optimal tensor decomposition strategy and tensor prediction model.

[0100] Specifically, for high-order tensor data :

[0101] state: state set at time ;

[0102] In the formula, For the original tensor, To predict tensors, , and For one-dimensional coefficient tensors of different sizes, different tensor decomposition rules are applied from... From the middle, hour , and Randomly generated.

[0103] action: Moment Action Set ;

[0104] In the formula, Indicates at time The selected actions are implemented according to specific rules for tensor decomposition. During tensor decomposition, a tensor can be represented as a combination of multiple factor matrices and weights. The tensor decomposition rules referred to are CP-WOPT, CP-NLS, and Tucker, respectively.

[0105] Specifically,

[0106]

[0107] in, Indicates the outer product. , and The dimension factor matrix corresponding to each tensor and The collective term for column vectors, specifically they can be... , , wait, For rank.

[0108] Factor matrix and The update is as follows:

[0109]

[0110] in, and These represent the Khatri-Rao product and the mode-n product, respectively. This represents the tensor to be decomposed.

[0111]

[0112] Factor matrix and The update is as follows:

[0113]

[0114] in, It is the loss function (usually the sum of squared residuals). It is the learning rate.

[0115]

[0116] in, It is the core tensor. The factor matrix is ​​updated as follows:

[0117]

[0118] in, This represents the tensor to be decomposed. This represents the Moore-Penrose generalized inverse.

[0119] award: Momentary Rewards In the current state Next action The set of immediate rewards obtained afterward. The aim is to minimize... and The global model is optimized with the prediction error between the two sides as the target. Represented as:

[0120]

[0121] Right now The weighted cumulative residual between and.

[0122] Step S320 includes:

[0123] S321. Initialize the experience replay pool, randomly initialize the main network, and set the parameters of the main network. Copy the low-order matrix to the target network; input the low-order matrix into the initialization matrix prediction model to obtain the high-order prediction state set. ,make Get the state Determine the set of optional actions ,in, Indicates at time The selected action to achieve matrix decomposition;

[0124] S322, with probability Randomly select an action, to The probability is used to select the optimal action based on the main network's valuation. ;in, ;

[0125] S323, The optimal action obtained from S322 Receive the corresponding reward value To reach the next state , will vector Place them into the experience replay pool respectively; if the experience replay pool is full, proceed to step S324; if the experience replay pool is not full, then... And proceed to step S322;

[0126] S324. Randomly select a batch of vectors from the experience replay pool, and input each vector into the main network. Find the action corresponding to the largest Q value in the main network and input it into the target network along with the above vectors to obtain the target network Q value under the new model. Calculate the loss function using the obtained main network Q value and target network Q value, and update the parameters of the matrix prediction model using gradient descent. ;

[0127] S325. Repeat step S324, and change the parameter every time interval T. The algorithm is copied to the target network of the matrix prediction model until the network converges or reaches the preset number of iterations, thus obtaining the optimal matrix decomposition strategy and matrix prediction model.

[0128] Specifically, for low-order matrix data :

[0129] state: state set at time ;

[0130] In the formula, For the original matrix, For the prediction matrix, Different matrix decomposition rules are used to decompose coefficient matrices of different sizes. get, hour Randomly generated. For decomposition rules. This can be further expressed as:

[0131]

[0132] in, This represents an optional matrix factorization rule, including Adagrad, RMSProp, and Adam.

[0133] Specifically,

[0134]

[0135]

[0136]

[0137] in, It is a matrix elements, It is a matrix elements, and They are matrices and Total offset, , and These are three constants. , , and Representing elements respectively , and First-moment estimate after bias correction , , and It is the second-order moment estimate after correction for the corresponding bias. , , and These are the results of the next decomposition. This represents the objective function.

[0138] action: Moment Action Set ;

[0139] In the formula, Indicates at time The action of matrix decomposition follows the rules of matrix decomposition. .

[0140] award: Momentary Rewards In the current state Next action The set of immediate rewards obtained afterward, designed to minimize and The global model is optimized with the prediction error between the two sides as the target.

[0141] award Represented as:

[0142]

[0143] in, This is the weight matrix. For coefficients, and The number of rows, the number of columns, and the number of latent feature dimensions, respectively. For matrix The average of all non-zero elements in the matrix.

[0144] Once the experience replay pool contains a sufficient number of vectors, a batch of vectors can be randomly selected from the pool. One of these vectors... For example, and Inputting the main neural network yields the Q estimate. , These represent the weights of the main neural network. Find the action corresponding to the largest Q-value in the current Q-network:

[0145]

[0146] Will as well as Input the target Q network, plus Obtain the target Q value:

[0147]

[0148] As a discount factor, The weights of the target Q-network are represented by the value of the target Q-network. The loss function is obtained by subtracting the Q-value of the main network from the Q-value of the target network.

[0149]

[0150] Using loss function Loss Update the main network weights using gradient descent. Weights of the target Q-network The updates depend on the training results of the main network. To reduce the impact of data fluctuations on model stability, updates are performed at intervals of [time value missing]. T The weights of the main network Weights copied to the target network This process keeps the target network constant for a period of time, reducing model volatility. This process is repeated iteratively until the loss function converges to a sufficiently small range. At this point, the main neural network has converged, and the optimal strategy can be obtained by deploying the system based on the Q-evaluation.

[0151] like Figure 2 and Figure 3 As shown, the specific steps for generating higher-order prediction tensors and lower-order prediction matrices are as follows:

[0152] (1) Initialize the experience replay pool, randomly initialize the main Q network, and copy the parameters to the target Q network, i.e. .

[0153] (2) Initialize the higher-order prediction state set With low-order prediction state set And let Get the state and Determine the set of optional actions. and .

[0154] (3) with Randomly select an action with a probability of [missing information]. The probability is used to select the optimal action based on the main network's valuation. and .

[0155] (4) Perform the finally selected action. and Each received a reward value. and To reach the next state and , will vector and The data are then placed into the experience replay pools for each model training step; if the experience replay pool is full, proceed to step (5); if the experience replay pool is not full, then... Then proceed to step (3).

[0156] (5) Randomly select a small batch of vector sets from the experience replay pool.

[0157] (6) Take out the vectors one by one. and The vector is then input into the main network. The action corresponding to the maximum Q value in the main network is found and combined with the above vector and input into the target network to obtain the target network Q value under the new mode. The loss function is then calculated using the obtained main network Q value and the target network Q value.

[0158]

[0159] Update parameters using gradient descent with the loss function. .

[0160] (7) Repeat steps (5)-(6) every time length T Copy the weights of the main network to the target network. Until the network converges.

[0161] Optionally, step S400 includes:

[0162] S410. Based on the original dimension information index of the low-order matrix and the cross-time slot weight, the prediction result of the low-order prediction matrix is ​​restored to obtain a restored tensor with the same structure as the high-order tensor, which is used for subsequent unified fusion processing.

[0163] Since the structure of low-order matrices only retains partial dimensional information (such as weighted integration of indicator dimensions), their prediction results cannot be directly fused with the data structure output by high-order tensor prediction models. To achieve structure alignment and fusion calculation, the matrix prediction results are reverse-engineered based on the cross-slot weights previously used for compression operations, reconstructing a three-dimensional structural tensor consistent with the structure of the high-order tensor. This allows for weighted fusion with the high-order tensor prediction results under a unified dimension, improving fusion accuracy and adaptability.

[0164] S420. Based on the errors of the matrix prediction model and the tensor prediction model, determine the preliminary fusion weights of the matrix prediction model and the tensor prediction model. The preliminary fusion weights are expressed as follows:

[0165]

[0166] in, and The matrix prediction model and the tensor prediction model are respectively at time 10:00. The error, It is a smoothing coefficient used to avoid division by zero and to smooth out errors;

[0167] S430. The preliminary fusion weights are normalized to calculate the dynamic fusion weight coefficients:

[0168]

[0169] in, The influence coefficient determines the relative influence between the errors of the two models; when When the tensor prediction model and the matrix prediction model have smaller errors, the fusion weights are set higher than those for the model with larger errors; when At that time, the fusion weights for the models with larger errors in the tensor prediction model and matrix prediction model are set to be greater than those for the models with smaller errors; the fusion weight coefficients satisfy the following: ;

[0170] S440. The higher-order prediction tensor is weighted and fused with the restored tensor obtained from the lower-order matrix to obtain the fused prediction result as the prediction output tensor; the weighted fusion formula is as follows:

[0171]

[0172] in, For the tensor prediction model at time 1000 The prediction tensor, Based on time The reconstructed tensor obtained from the low-order matrix prediction results This is the result of the fusion prediction.

[0173] The fusion prediction results obtained in the above scheme Based on a thorough study of the evolution trends of user performance indicators across multiple historical sampling time slots, the time series prediction output generated through a deep reinforcement learning-driven model inference process can accurately characterize the evolution trend of network performance indicators over a future period. In essence, it reflects the time The invention then predicts key performance parameters such as signal strength, transmission delay, throughput, and packet loss rate for one or more time slots. Based on these predictions, the network system can identify potential performance bottlenecks in advance and adaptively adjust dynamic resource allocation strategies, such as switching frequency bands in advance, load balancing, and adjusting link priorities, thereby effectively alleviating network congestion and preventing signal degradation. Furthermore, this invention can provide decision-making support for user roaming strategies, bandwidth scheduling, and QoS guarantees in high-density access environments, further improving overall network operating efficiency and the continuity, stability, and satisfaction of user experience.

[0174] In this embodiment of the invention, by constructing a three-dimensional high-order tensor and a normalized low-order matrix with user, performance index, and time series dimensions, the spatiotemporal characteristics and cross-user behavioral differences of Wi-Fi signals are fully preserved. Secondly, a prediction strategy guided by deep reinforcement learning replaces the static model training mechanism, enabling adaptation to rapid fluctuations in channel conditions. Combining tensor decomposition and matrix decomposition techniques, the decomposition method is continuously adjusted through strategy iteration to learn the optimal decomposition strategy. This allows the model to deeply explore the structural correlations between performance indicators through an interactive strategy optimization process during training, thereby improving prediction accuracy and generalization ability. This solves the technical problem of low accuracy in Wi-Fi signal perception and prediction due to limitations in dealing with sudden changes and nonlinear characteristics of Wi-Fi signals, effectively improving the prediction accuracy of network performance and the efficiency of resource allocation, thus better meeting users' stringent requirements for network service quality.

[0175] Example 2

[0176] like Figure 4 As shown, this invention provides an intelligent performance sensing and prediction system for Wi-Fi, comprising:

[0177] The collection module is used to collect raw signal data from Wi-Fi users, obtain corresponding feature values ​​based on the network performance indicators of each user, and construct high-order tensor data in the user dimension based on the feature values; wherein, the feature values ​​are in matrix form, the rows of the matrix represent different network performance indicators, and the columns represent different sampling time slots;

[0178] The processing module is used to normalize various network performance indicators based on the high-order tensor data, calculate the cross-time slot weights of various network performance indicators, and integrate the multi-dimensional network performance indicators composed of multiple network performance indicators corresponding to multiple sampling time slots into one-dimensional standardized data and form a low-order matrix.

[0179] The prediction module is used to input the high-order tensor data into a pre-trained tensor prediction model to obtain a high-order prediction tensor, and to input the low-order matrix data into a pre-trained matrix prediction model to obtain a low-order prediction matrix; wherein, the tensor prediction model and the matrix prediction model are trained based on a deep reinforcement learning algorithm;

[0180] The integration module is used to obtain high-order prediction error based on high-order prediction tensor and the high-order tensor data, obtain low-order prediction error based on low-order prediction matrix and low-order matrix data, dynamically adjust the fusion weight based on the high-order prediction error and low-order prediction error, and fuse the prediction results of tensor prediction model and matrix prediction model by weighted averaging according to the adjusted fusion weight to obtain fused prediction result.

[0181] The intelligent performance perception and prediction system for Wi-Fi provided in this embodiment of the invention is used to perform the above-mentioned tasks. Figure 1 The illustrated embodiment presents an intelligent performance perception and prediction method for Wi-Fi. For details not covered in this embodiment, please refer to the foregoing. Figure 1 The intelligent performance perception and prediction method for Wi-Fi shown in the embodiment will not be described in detail here.

[0182] Example 3

[0183] This invention provides an electronic device, including a memory and a processor; the memory is used to store computer-executed instructions, and the processor is connected to the memory via a bus;

[0184] When the Wi-Fi-oriented intelligent performance perception and prediction system is running, the processor executes the computer execution instructions stored in the memory to cause the Wi-Fi-oriented intelligent performance perception and prediction system to perform the Wi-Fi-oriented intelligent performance perception and prediction method as described in any one of Embodiment 1.

[0185] Example 4

[0186] The present invention provides a computer-readable storage medium comprising computer-executable instructions, which, when executed on a computer, cause the computer to perform the intelligent performance perception and prediction method for Wi-Fi as described in any one of the embodiments.

[0187] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent performance perception and prediction for Wi-Fi, characterized in that, include: S100. Collect raw signal data of Wi-Fi users, obtain corresponding feature values ​​according to the network performance indicators of each user, and construct high-order tensor data in the user dimension based on the feature values; wherein, the feature values ​​are in matrix form, the rows of the matrix represent different network performance indicators, and the columns represent different sampling time slots; S200. Based on the high-order tensor data, normalize the various network performance indicators, calculate the cross-time slot weights of the various network performance indicators, and integrate the multi-dimensional network performance indicators composed of multiple network performance indicators corresponding to multiple sampling time slots into one-dimensional standardized data and form a low-order matrix. S300: Input the high-order tensor data into a pre-trained tensor prediction model to obtain a high-order prediction tensor, and input the low-order matrix data into a pre-trained matrix prediction model to obtain a low-order prediction matrix; wherein, the tensor prediction model and the matrix prediction model are trained based on a deep reinforcement learning algorithm. S400. Obtain the high-order prediction error based on the high-order prediction tensor and the high-order tensor data, obtain the low-order prediction error based on the low-order prediction matrix and the low-order matrix data, dynamically adjust the fusion weights based on the high-order prediction error and the low-order prediction error, and fuse the prediction results of the tensor prediction model and the matrix prediction model by weighted averaging according to the adjusted fusion weights to obtain the fused prediction result.

2. The method as described in claim 1, characterized in that, Step S100 includes: S110. Capture signal interaction information between user equipment and access point in a Wi-Fi network environment, and collect raw signal data of Wi-Fi users during different time periods of network usage; S120. For each independent Wi-Fi user, using network performance indicators as rows and sampling time slots as columns, convert the different network performance indicator data in the original signal data into a matrix, which serves as the feature value matrix of the network performance indicator for the corresponding user. S130. Treat each user's feature value matrix as an independent substructure, and stack and combine them along the user dimension to form high-order tensor data. .

3. The method as described in claim 1, characterized in that, Step S200 includes: S210, Transfer the high-order tensor data The elements of each eigenvalue matrix are normalized, and the normalized result is... Represented as: in, This represents the column vector in the eigenvalue matrix, where the column vector is... , Represents column vectors The One element, column vector The maximum value in; S220. Construct a vector based on the normalized values ​​of each network performance metric across all sampling time slots. Calculate the weight of each network performance metric across all sampling time slots. ;in, This represents the total number of network performance metrics. This represents the total number of sampling time slots. Indicates the first The network performance metric in the first Normalized values ​​over each sampling time slot; S230. Integrate multi-dimensional network performance metrics into one-dimensional standardized data to construct a low-order matrix. Among them, the low-order matrix Corresponding element The expression is: , Indicates the first The weights of each network performance metric Indicates the first The network performance metric in the first Normalized values ​​over each sampling time slot.

4. The method as described in claim 1, characterized in that, Step S300 includes: S310. Tensor decomposition based on deep reinforcement learning algorithm is used to process the high-order tensor data. The model parameters of the initial tensor prediction model are constructed by decomposing the data using high-order tensor data. The optimization objective is to minimize the prediction error between the predicted tensor and the predicted tensor of the initial tensor prediction model. The decomposition method of tensor decomposition is continuously adjusted through strategy iteration, and the model parameters of the initial tensor prediction model are updated. The initial tensor prediction model is then trained based on the updated model parameters until the network converges or reaches the preset number of iterations, thus obtaining the tensor prediction model. S320. The low-order matrix is ​​decomposed using matrix factorization based on a deep reinforcement learning algorithm to construct the model parameters of the initial matrix prediction model, using the low-order matrix... Minimizing the error between the initial matrix prediction model and the predicted matrix is ​​the optimization objective. The matrix decomposition method is continuously adjusted through strategy iteration to update the model parameters of the initial matrix prediction model. The initial matrix prediction model is then trained based on the updated model parameters until the network converges or reaches the preset number of iterations to obtain the matrix prediction model. S330. Input the high-order tensor data into a pre-trained tensor prediction model to obtain a high-order prediction tensor, and input the low-order matrix data into a pre-trained matrix prediction model to obtain a low-order prediction matrix.

5. The method as described in claim 4, characterized in that, Step S310 includes: S311. Initialize the experience replay pool, randomly initialize the main network, and set the parameters of the main network... Copy the data to the target network; input the higher-order tensor data into the initialization tensor prediction model to obtain the higher-order prediction state set. ,make Get the state Determine the set of optional actions ,in, Indicates at time To implement the selected action for tensor decomposition. To predict tensors, , and For one-dimensional coefficient tensors of different sizes; S312. Randomly select an action with probability φ, and select the optimal action based on the main network's estimate with probability 1-φ. ;in, ; S313, The optimal action obtained from S312 Receive the corresponding reward value To reach the next state , will vector Place them into the experience replay pool respectively; if the experience replay pool is full, proceed to step S314; if the experience replay pool is not full, then... And proceed to step S312; wherein, This is the current state; S314. Randomly select a batch of vectors from the experience replay pool, and input each vector into the main network. Find the action corresponding to the largest Q value in the main network and input it into the target network along with the above vectors to obtain the target network Q value under the new mode. Calculate the loss function using the obtained main network Q value and target network Q value, and update the parameters of the tensor prediction model using gradient descent. ; S315. Repeat step S314, and change the parameter every time interval T. The algorithm is copied to the target network of the tensor prediction model until the network converges or reaches the preset number of iterations, thus obtaining the optimal tensor decomposition strategy and tensor prediction model.

6. The method as described in claim 4, characterized in that, Step S320 includes: S321. Initialize the experience replay pool, randomly initialize the main network, and set the parameters of the main network. Copy to the target network; input the low-order matrix into the initialization matrix prediction model to obtain the low-order prediction state set. ,make Get the state Determine the set of optional actions ,in, Indicates at time The selected action to achieve matrix decomposition. For the prediction matrix, Different matrix decomposition rules are used to decompose coefficient matrices of different sizes. get, For decomposition rules; S322. Randomly select an action with probability φ, and select the optimal action based on the main network's estimate with probability 1-φ. ;in, ; S323, The optimal action obtained from S322 Receive the corresponding reward value To reach the next state , will vector Place them into the experience replay pool respectively; if the experience replay pool is full, proceed to step S324; if the experience replay pool is not full, then... And proceed to step S322; wherein, This is the current state; S324. Randomly select a batch of vectors from the experience replay pool, and input each vector into the main network. Find the action corresponding to the largest Q value in the main network and input it into the target network along with the above vectors to obtain the target network Q value under the new model. Calculate the loss function using the obtained main network Q value and target network Q value, and update the parameters of the matrix prediction model using gradient descent. ; S325. Repeat step S324, and change the parameter every time interval T. The algorithm is copied to the target network of the matrix prediction model until the network converges or reaches the preset number of iterations, thus obtaining the optimal matrix decomposition strategy and matrix prediction model.

7. The method as described in claim 1, characterized in that, Step S400 includes: S410. Perform a restoration operation on the prediction result of the low-order prediction matrix based on the original dimension information index and cross-time slot weight of the low-order matrix to obtain a restored tensor with the same structure as the high-order tensor. S420. Based on the errors of the matrix prediction model and the tensor prediction model, determine the preliminary fusion weights of the matrix prediction model and the tensor prediction model. The preliminary fusion weights are expressed as follows: in, and The matrix prediction model and the tensor prediction model are respectively at time 10:

00. The error, It is a smoothing coefficient used to avoid division by zero and to smooth out errors; S430. The preliminary fusion weights are normalized to calculate the dynamic fusion weight coefficients: in, The influence coefficient determines the relative influence between the errors of the two models; when When the tensor prediction model and the matrix prediction model have smaller errors, the fusion weights are set higher than those for the model with larger errors; when At that time, the fusion weights for the models with larger errors in the tensor prediction model and matrix prediction model are set to be greater than those for the models with smaller errors; the fusion weight coefficients satisfy the following: ; S440. The higher-order prediction tensor is weighted and fused with the restored tensor obtained from the lower-order matrix to obtain the fused prediction result as the prediction output tensor; the weighted fusion formula is as follows: in, For the tensor prediction model at time 1000 The prediction tensor, Based on time The reconstructed tensor obtained from the low-order matrix prediction results This is the result of the fusion prediction.

8. An intelligent performance sensing and prediction system for Wi-Fi, characterized in that, include: The collection module is used to collect raw signal data from Wi-Fi users, obtain corresponding feature values ​​based on the network performance indicators of each user, and construct high-order tensor data in the user dimension based on the feature values; wherein, the feature values ​​are in matrix form, the rows of the matrix represent different network performance indicators, and the columns represent different sampling time slots; The processing module is used to normalize various network performance indicators based on the high-order tensor data, calculate the cross-time slot weights of various network performance indicators, and integrate the multi-dimensional network performance indicators composed of multiple network performance indicators corresponding to multiple sampling time slots into one-dimensional standardized data and form a low-order matrix. The prediction module is used to input the high-order tensor data into a pre-trained tensor prediction model to obtain a high-order prediction tensor, and to input the low-order matrix data into a pre-trained matrix prediction model to obtain a low-order prediction matrix; wherein, the tensor prediction model and the matrix prediction model are trained based on a deep reinforcement learning algorithm; The integration module is used to obtain high-order prediction error based on high-order prediction tensor and the high-order tensor data, obtain low-order prediction error based on low-order prediction matrix and low-order matrix data, dynamically adjust the fusion weight based on the high-order prediction error and low-order prediction error, and fuse the prediction results of tensor prediction model and matrix prediction model by weighted averaging according to the adjusted fusion weight to obtain fused prediction result.

9. An electronic device, characterized in that, It includes a memory and a processor; the memory is used to store computer-executed instructions, and the processor is connected to the memory via a bus; When the Wi-Fi-oriented intelligent performance perception and prediction system is running, the processor executes the computer execution instructions stored in the memory to cause the Wi-Fi-oriented intelligent performance perception and prediction system to perform the Wi-Fi-oriented intelligent performance perception and prediction method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes computer-executable instructions that, when executed on a computer, cause the computer to perform the intelligent performance sensing and prediction method for Wi-Fi as described in any one of claims 1-7.

Images