Real-time video QoE prediction method and system based on wireless network parameters

Abstract

The invention discloses a real-time video QoE prediction method and system based on wireless network parameters, and the method comprises the steps: connecting an OAI signal through terminal equipment, collecting a QoE parameter index in a video playing process from a video from WebRTC, and obtaining a corresponding wireless network state parameter index as original data; a training set is generated through preprocessing, and the training set is used for training a constructed two-stage QoE perception and prediction model based on a Transform model after data enhancement processing; and in the online stage, inputting real-time data into the pre-trained two-stage QoE perception and prediction model based on the Transform model to carry out QoE perception prediction. The adaptive capacity of the model to different network conditions and video content changes is effectively improved through a robust data enhancement technology, and the generalization and robustness of the model are enhanced. Real-time QoE prediction is completed in a short time interval through two-stage QoE perception and prediction based on a Transform model, and the requirements of a real-time XR video communication system are met.

Description

Technical Field

[0001] This invention relates to a technology in the field of wireless communication, specifically a real-time video QoE prediction method and system based on wireless network parameters. Background Technology

[0002] With the widespread adoption of interactive and immersive applications such as Extended Reality (XR), unprecedented demands are being placed on networks. Due to network security requirements, existing wireless networks cannot obtain the Quality of User Experience (QoE) status of XR services from the application layer. This means that existing network optimization methods can only be based on traditional Quality of Service (QoS) and cannot effectively match the actual user experience requirements of XR services. Summary of the Invention

[0003] This invention addresses the shortcomings of existing wireless networks, which cannot obtain the QoE status of XR services from the application layer and do not consider QoE prediction for real-time XR applications. It proposes a real-time video QoE prediction method and system based on wireless network parameters. Robust data augmentation techniques effectively improve the model's adaptability to different network conditions and video content variations, enhancing its generalization and robustness. Through a two-stage QoE perception and prediction based on a Transformer model, real-time QoE prediction is completed within a short time interval, meeting the requirements of real-time XR video communication systems.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a real-time video QoE prediction method based on wireless network parameters. The method involves connecting a terminal device to an OAI signal and collecting QoE parameter indices from video playback via WebRTC, along with corresponding wireless network state parameter indices as raw data. These raw data are preprocessed to generate a training set, which is then used to train a two-stage QoE perception and prediction model based on a Transformer model after data augmentation. During the online phase, real-time data is input into the pre-trained two-stage QoE perception and prediction model based on the Transformer model for QoE perception and prediction.

[0006] The two-stage QoE perception and prediction model based on the Transformer model includes an encoder and a decoder, wherein: the encoder performs QoE perception based on historical network information and extracts QoE feature information; the decoder performs inference and prediction based on historical network information and QoE feature information to generate QoE prediction results.

[0007] The data augmentation mentioned above refers to generating new training samples by mixing and / or superimposing random perturbation vectors onto the training samples to augment the data.

[0008] The aforementioned hybrid synthesis refers to taking two samples from the training set and mixing them to create a new synthetic sample, specifically: R i =λ(R) t1 )+(1-λ)R t1 ,in: The subscripts t1 and t2 represent two samples randomly drawn from the training data, where λ∈[0,1]. Training with mixed samples can prevent overfitting and reduce the impact of noisy or outlier samples in the dataset.

[0009] The aforementioned random perturbation refers to introducing random perturbations during training to enhance the robustness of the model, specifically: Where: η t It is a variance of The random perturbation Gaussian vector. Attached Figure Description

[0010] Figure 1 This is a flowchart of the present invention;

[0011] Figure 2 This is a schematic diagram of the system of the present invention;

[0012] Figure 3 This is a schematic diagram illustrating the effect of an example. Detailed Implementation

[0013] like Figure 2 As shown in this embodiment, a Radio Access Network Intelligent Controller (RIC) assisted XR video communication system implementing the above method is deployed in an edge server to provide real-time XR video services. The system encodes the real-time video and transmits it to users for playback via the Evolved Packet Core (EPC) and RAN. The system includes: a Radio Network Information Service (RNIS) unit, an AI / ML training service unit, and a QoE awareness service unit. Specifically, the RNIS unit periodically collects network status data through the RAN interface, and the AI / ML training service unit trains QoE awareness and prediction models using RAN information cached on the QoE awareness service and the QoE dataset. The QoE sensing service unit obtains RAN information from RNIS, maintains the RAN information matrix as input to the neural network, and utilizes the trained QoE sensing and prediction model. Perform QoE perception and prediction.

[0014] The aforementioned timed acquisition of network status refers to: the RNIS unit sampling RAN information r in each time slot t. t And store it on the cache server and consider a length of t hThe historical observation window is used to construct a RAN information matrix based on currently collected RAN information and historical RAN information. Specifically, the matrix consists of: RAN information. RAN Information Matrix in: Average Reference Signal Received Power (RSRP) For the average reference signal reception quality, For average channel quality indicators, For average modulation coding scheme, The number of physical resource blocks (PRBs) scheduled within the sampling period, The average Harq number.

[0015] The QoE perception and prediction model From RAN information matrix R t Predicting real-time video The future QoE is as follows: in: It is the predicted QoE vector, and the current QoE vector. q t For the QoE of XR video transmission within time slot t, t p θ represents the length of the prediction window, and θ represents the parameters of the QoE perception and prediction model.

[0016] The AI / ML training service unit utilizes the RAN information and QoE dataset cached on the QoE perception service, aiming to obtain the optimal parameters θ and using the mean absolute error (MAE) as the performance metric, to construct a multivariate regression problem. This enables the QoE perception and prediction model to accurately estimate the current and future QoE of XR videos using the RAN information. Specifically: in: E t for The output feature matrix, For the estimated historical QoE results, |·| denotes the l1 norm function.

[0017] The aforementioned multiple regression problem is solved in the following way:

[0018] Step ① Construct a representative function for the series connection. and And it is mapped to the encoder and decoder of the Transformer model.

[0019] Step ② Utilizing the encoder of the Transformer model The input sequence is encoded into a fixed-dimensional vector representation, while capturing semantic information and contextual relationships at different positions in the input sequence to generate a feature matrix;

[0020] Step ③ Utilizing the decoder of the Transformer model Based on the estimation results and the characteristic matrix E t Jointly predict QoE, which is the probability distribution on the target sequence.

[0021] The Transformer model includes an encoder and a decoder, wherein the encoder includes a multi-layer fully connected network, a multi-head attention layer, and a residual regularization layer; the decoder includes a multi-layer fully connected network, a multi-head attention layer, and a residual regularization layer.

[0022] This embodiment further conducts specific experiments on an OAI-based wireless transmission platform and a WebRTC-based real-time video transmission platform. The OAI-based wireless transmission platform consists of a software-defined EPC and RAN using an Intel i7-8559U processor, a USRP B210 RF front-end, and commercial user equipment (UEs). The real-time video server is built using PionWebRTC. WebRTC supports real-time audio and video transmission via a web browser and can be expanded into an XR transmission platform. Commercial terminals request and play real-time video through a browser.

[0023] RAN information is acquired in real time, and video transmission information is also collected in real time via WebRTC to obtain the video transmission QoE. Data is collected every 1 second. This invention collected a total of 30 video playback datasets, with a duration of 9366 seconds. Among them, 20 datasets were used as the training set, and 10 datasets were used as the test set.

[0024] Table 1 shows a comparison between the present invention and existing technologies in three types of QoE (QoE-OnRL, QoE-Loki, and QoE-R3Net), where the existing technologies are:

[0025] 1) RF: Random forest (RF) based schemes predict the QoE of real-time video by using ensemble learning of multiple decision trees.

[0026] 2) Support Vector Regression (SVR): The SVR-based approach predicts the QoE of real-time video using the support vector machine method.

[0027] 3) Deep Neural Networks: Deep neural networks use multi-layer fully connected networks to predict QoE.

[0028] 4) CNN-LSTM: An effective solution for bandwidth prediction based on a hybrid architecture of convolutional neural networks (CNN) and long short-term memory networks (LSTM).

[0029] Table 1. Performance Comparison of Different Algorithms

[0030] t in the table p =4,t h =6,

[0031] As shown in Table 1, considering MAE and median error E 50 90th percentile error E 90 The four evaluation metrics are: data augmentation, runtime, and other performance metrics. For QoE-OnRL, compared with RF, SVR, DNN, and CNN-LSTM algorithms, the present invention reduces MAE by 12.5%–44.0%, median error by 10.9%–56.6%, and 90th percentile error by 13.0%–34.8%. For QoE-Loki, the present invention reduces MAE by 12.5% ​​to 44.0%, median error by 10.9% to 56.63%, and 90th percentile error by 13.0% to 34.79%. For QoE-Loki and QoE-R3Net, the present invention reduces MAE by more than 23.8% and 2.6%, respectively. Compared with the present invention without data augmentation, data augmentation brings an additional 6.3%–15.9% performance improvement in MAE.

[0032] In summary, this invention can optimize the transmission and processing of video data, improve the quality and response speed of monitoring data, and is applicable to scenarios such as video transmission in in-vehicle entertainment systems, monitoring systems, and online education platforms.

[0033] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A real-time video QoE prediction method based on wireless network parameters, characterized in that, By connecting to the OAI signal through a terminal device and collecting QoE parameter indicators during video playback from WebRTC, and obtaining the corresponding wireless network status parameter indicators as raw data, a training set is generated after preprocessing and data augmentation processing, which is then used to train the constructed two-stage QoE perception and prediction model based on the Transformer model. In the online stage, real-time data is input into the pre-trained two-stage QoE perception and prediction model based on the Transformer model for QoE perception and prediction.

2. The real-time video QoE prediction method based on wireless network parameters according to claim 1, characterized in that, The two-stage QoE perception and prediction model based on the Transformer model includes an encoder and a decoder, wherein: the encoder performs QoE perception based on historical network information and extracts QoE feature information; the decoder performs inference and prediction based on historical network information and QoE feature information and generates QoE prediction results. The encoder includes: a multi-layer fully connected network, a multi-head attention layer, and a residual regularization layer; the decoder includes: a multi-layer fully connected network, a multi-head attention layer, and a residual regularization layer.

3. The real-time video QoE prediction method based on wireless network parameters according to claim 1, characterized in that, The aforementioned data augmentation refers to generating new training samples by mixing and / or superimposing random perturbation vectors onto the training samples to augment the data. The aforementioned hybrid synthesis refers to taking two samples from the training set and mixing them to create a new synthetic sample, specifically: R i =λ(R) t1 )+(1-λ)R t1 ,in: The subscripts t1 and t2 represent two samples randomly drawn from the training data, λ∈[0,1]. Training with mixed samples prevents overfitting and reduces the impact of noisy or outlier samples in the dataset. The aforementioned random perturbation refers to introducing random perturbations during training to enhance the robustness of the model, specifically: Where: η t It is a variance of The random perturbation Gaussian vector.

4. A wireless access network intelligent controller (RIC) assisted XR video communication system implementing the method of any one of claims 1-3, characterized in that, The system encodes real-time video and transmits it to users for playback via the Evolved Packet Core (EPC) and RAN. Specifically, it includes a Radio Network Information Service (RNIS) unit, an AI / ML training service unit, and a QoE awareness service unit. The RNIS unit periodically collects network status data via the RAN interface, and the AI / ML training service unit uses RAN information cached on the QoE awareness service and the QoE dataset to train QoE awareness and prediction models. The QoE sensing service unit obtains RAN information from RNIS, maintains the RAN information matrix as input to the neural network, and utilizes the trained QoE sensing and prediction model. Perform QoE perception and prediction.

5. The wireless access network intelligent controller (RIC) assisted XR video communication system according to claim 4, characterized in that, The aforementioned timed acquisition of network status refers to: the RNIS unit sampling RAN information r in each time slot t. t And store it on the cache server and consider a length of t h The historical observation window is used to construct a RAN information matrix based on currently collected RAN information and historical RAN information. Specifically, the matrix consists of: RAN information. RAN Information Matrix in: Average Reference Signal Received Power (RSRP) For the average reference signal reception quality, For average channel quality indicators, For average modulation coding scheme, The number of physical resource blocks (PRBs) scheduled within the sampling period, The average Harq number.

6. The wireless access network intelligent controller (RIC) assisted XR video communication system according to claim 4, characterized in that, The QoE perception and prediction model From RAN information matrix R t Predicting real-time video The future QoE is as follows: in: It is the predicted QoE vector, and the current QoE vector. q t For the QoE of XR video transmission within time slot t, t p θ represents the length of the prediction window, and θ represents the parameters of the QoE perception and prediction model.

7. The wireless access network intelligent controller (RIC) assisted XR video communication system according to claim 4, characterized in that, The AI / ML training service unit utilizes the RAN information and QoE dataset cached on the QoE perception service, aiming to obtain the optimal parameters θ and using the mean absolute error (MAE) as the performance metric, to construct a multivariate regression problem. This enables the QoE perception and prediction model to accurately estimate the current and future QoE of XR videos using the RAN information. Specifically: in: E t for The output feature matrix, For the estimated historical QoE results, |·| denotes the l1 norm function.

8. The wireless access network intelligent controller (RIC) assisted XR video communication system according to claim 4, characterized in that, The aforementioned multiple regression problem is solved in the following way: Step ① Construct a representative function for the series connection. and And mapped to the encoder and decoder of the Transformer model; Step ② Utilizing the encoder of the Transformer model The input sequence is encoded into a fixed-dimensional vector representation, while capturing semantic information and contextual relationships at different positions in the input sequence to generate a feature matrix; Step ③ Utilizing the decoder of the Transformer model Based on the estimation results and the characteristic matrix E t Jointly predict QoE, which is the probability distribution on the target sequence.

Images