WO2023045886A1 - 模型训练方法，视频用户体验预测方法，装置及电子设备 - Google Patents

模型训练方法，视频用户体验预测方法，装置及电子设备 Download PDF

Info

Publication number
WO2023045886A1
WO2023045886A1 PCT/CN2022/119734 CN2022119734W WO2023045886A1 WO 2023045886 A1 WO2023045886 A1 WO 2023045886A1 CN 2022119734 W CN2022119734 W CN 2022119734W WO 2023045886 A1 WO2023045886 A1 WO 2023045886A1
Authority
WO
WIPO (PCT)
Prior art keywords
user experience
base station
data
video
station side
Prior art date
Application number
PCT/CN2022/119734
Other languages
English (en)
French (fr)
Inventor
廖金龙
许靖
汪波
芮华
吕星哉
Original Assignee
中兴通讯股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中兴通讯股份有限公司 filed Critical 中兴通讯股份有限公司
Publication of WO2023045886A1 publication Critical patent/WO2023045886A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
    • H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies

Definitions

  • the present application relates to the field of communication technology, and in particular to a model training method, a video user experience prediction method, a device and electronic equipment.
  • KPI Key Performance Indicator, key performance indicator
  • the purpose of the embodiments of the present application is to provide a model training method, video user experience prediction method, device and electronic equipment for establishing a model capable of predicting video user experience based on wireless performance data at the base station side.
  • the embodiments of the present application provide a model training method, including: collecting video user experience category data on the user equipment UE side, and collecting base station side wireless performance data generated when users watch videos on the base station side; according to the The base station side wireless performance data generated when the user watches the video generates a feature matrix; the convolutional neural network model is trained based on a preset algorithm, and the trained convolutional neural network model is used as a video user experience prediction model; wherein, the The training data of the convolutional neural network model includes the feature matrix and video user experience category data; wherein, the video user experience prediction model is used for real-time prediction of video user experience according to the wireless performance data at the base station side.
  • the embodiment of the present application also provides a method for video user experience prediction, including: collecting base station side wireless performance data in real time; generating a real-time feature matrix according to the real-time collected base station side wireless performance data; inputting the real-time feature matrix into the video
  • the user experience prediction model obtains the real-time video user experience category, wherein the video user experience prediction model is trained by the above-mentioned model training method.
  • the embodiment of the present application also provides a model training device, including: a collection module, which is used to collect video user experience category data on the user equipment UE side, and collect base station side wireless performance data generated when users watch videos on the base station side; generate The module is used to generate a feature matrix according to the base station side wireless performance data generated when the user watches the video; the training module is used to train the convolutional neural network model based on a preset algorithm, and the trained convolutional neural network
  • the model is used as a video user experience prediction model; wherein, the training data of the convolutional neural network model includes the feature matrix and video user experience category data; wherein, the video user experience prediction model is used according to the wireless performance of the base station side data to make real-time predictions about video user experience.
  • the embodiment of the present application also provides a device for predicting video user experience, including: a collection module for collecting wireless performance data on the base station side in real time; a generating module for generating real-time features based on the wireless performance data collected in real time on the base station side Matrix; an acquisition module, configured to input the real-time feature matrix into a video user experience prediction model to obtain real-time video user experience categories, wherein the video user experience prediction model is trained by the above-mentioned model training method.
  • the embodiment of the present application also provides an electronic device, including: at least one processor; and a memory connected in communication with the at least one processor; wherein, the memory stores information that can be executed by the at least one processor. Instructions, the instructions are executed by the at least one processor, so that the at least one processor can execute the above-mentioned model training method, or execute the above-mentioned video user experience prediction method.
  • Embodiments of the present application also provide a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, implements the above-mentioned model training method, or implements the above-mentioned video user experience prediction method.
  • FIG. 1 is a flowchart of a model training method according to an embodiment of the present application
  • FIG. 2 is a flow chart of a method for video user experience prediction according to an embodiment of the present application
  • Fig. 3 is a schematic structural diagram of a model training device according to an embodiment of the present application.
  • Fig. 4 is a schematic structural diagram of an apparatus for predicting video user experience according to an embodiment of the present application
  • Fig. 5 is a schematic structural diagram of an electronic device according to an embodiment of the present application.
  • the model training device collects video user experience category data on the user equipment UE side, and collects base station side wireless performance data generated when users watch videos on the base station side; according to the base station side wireless performance data generated when users watch videos data to generate a feature matrix; train a convolutional neural network model based on a preset algorithm, and use the trained convolutional neural network model as a video user experience prediction model; wherein, the training data of the convolutional neural network model includes the The feature matrix and video user experience category data; wherein, the video user experience prediction model is used to predict video user experience in real time according to the wireless performance data at the base station side.
  • Step 101 the model training apparatus collects video user experience category data on the user equipment side, and collects wireless performance data on the base station side.
  • the model training apparatus first collects video user experience category data at the user equipment side.
  • User experience category data refers to the indicators for evaluating video user experience. You can select a specific video user experience evaluation standard, or customize a video user experience evaluation standard, and evaluate the video according to this standard to obtain the video user experience.
  • Video UX category data refers to the obtained video user experience category data and corresponding base station side wireless performance data.
  • the video user experience category data may adopt the video subjective quality scoring standard defined by the P.910 protocol of the International Telecommunication Union. This standard divides video user experience into five categories, and the specific classification of video user experience categories is shown in Table 1.
  • the wireless performance data generated by users watching videos is collected on the base station side.
  • the collected wireless performance data may include Packet Data Convergence Protocol (Packet Data Convergence Protocol, referred to as "PDCP”) layer data and multiple access channel (Multiple Access Channel, referred to as "MAC”) layer data.
  • PDCP Packet Data Convergence Protocol
  • MAC Multiple Access Channel
  • PDCP layer data includes data packet timestamp, data packet delay, data packet transmission direction)/downlink, data packet size, data packet serial number (Serial Number, referred to as "SN") value, etc.;
  • MAC layer data includes scheduling resource blocks (Resource Block, referred to as "RB") number, scheduling times, modulation and coding strategy (Modulation and Coding Scheme, referred to as "MCS”), confirmation character (Acknowledge character, referred to as "Ack”) and negative response signal (Negative Acknowledgment, referred to as “Nack”) feedback, etc.
  • RB Resource Block
  • MCS Modulation and Coding Scheme
  • Ack Acknowledge character
  • Nack negative response signal
  • SNR Signal-Noise Ratio
  • RSRP Reference Signal Receiving Power
  • CQI UE feedback information Quality Indication
  • the wireless performance data collected on the base station side can not only select PDCP layer data and MAC layer data, but also select from the high layer or physical layer of PDCP that can reflect PDCP layer data and MAC layer data. other wireless performance data. It is also possible to add relevant redundant features on the basis of existing features, or to combine or compress existing features.
  • the base station side wireless performance data after collecting the video user experience category data on the UE side, and collecting the base station side wireless performance data generated when the user watches the video on the base station side, it further includes: deleting missing values and abnormal values in the base station side wireless performance data;
  • the following formula is used to normalize the numerical non-discrete data in the wireless performance data at the base station side:
  • x i is the i-th original value of the wireless performance data column
  • x min is the minimum value of the wireless performance data column
  • x max is the maximum value of the wireless performance data column
  • x i new is the normalized value
  • the numerical or non-numeric discrete data in the wireless performance data of the base station side shall be effectively coded by one bit It can be specifically: one-bit effective encoding of numerical or non-numerical discrete data into numerical data and a state sparse matrix.
  • 0 is used for uplink and 1 for downlink; the SN value of the PDCP layer data packet is used to reflect the packet loss situation, and 0 indicates packet loss, and 1 indicates no packet loss; ACK and NACK of MAC layer data In feedback, 0 is used to represent NACK, 1 is used to represent ACK, etc.; MCS uses a 29-dimensional state sparse matrix to encode values from 0 to 28; CQI uses a 16-dimensional state sparse matrix to encode values from 0 to 15, etc.
  • the one-bit effective encoding of the video user experience category data can be specifically: for the case where the video user experience categories are 1 to 5, use a 5-dimensional state sparse matrix to encode them as [1, 0, 0, 0, 0], [0, 1, 0, 0, 0], [0, 0, 1, 0, 0], [0, 0, 0, 1, 0], [0, 0, 0, 0, 1] .
  • the collected base station wireless performance data and user experience category data are processed to clean the collected data, making the data suitable for building feature matrices, and then suitable for training convolutional neural network models as training data.
  • Step 102 the model training device generates a feature matrix according to the wireless performance data at the base station side.
  • the model training device processes and converts the collected wireless performance data at the base station side into a feature matrix, so that it is suitable for inputting the convolutional neural network model as training data.
  • generating the feature matrix according to the base station side wireless performance data generated when the user watches the video may specifically include: setting the column content of the feature matrix to the value of the same feature of different data packets, and setting the The row contents of the feature matrix are set to the values of different features of the same data packet; wherein, the features include wireless performance features extracted according to the collected wireless performance data of the base station side, and the extracted wireless performance features are combined to obtain wireless performance characteristics.
  • the specific setting of the elements in the feature matrix can be realized.
  • the combination of the extracted wireless performance characteristics involved in this example to obtain the wireless performance characteristics includes: reflecting whether the packet is lost with the PDCP layer packet timestamp and the packet SN value, combining and calculating the packet loss rate; Combining stamp and MAC layer data Ack and Nack feedback to calculate block error rate; combining data packet timestamp and data packet size to calculate throughput rate, etc.
  • after setting the rows of the feature matrix to the values of different features of the same data packet it also includes: when it is detected that the number of rows or columns of the feature matrix is less than a preset value, increasing The rows or columns of the feature matrix until the number of rows and columns of the feature matrix are equal to the preset value, the element value in the added row or column is set to 0; when the number of rows and columns is equal to the preset value Add one column to the left side of the start column and the right side of the end column of the feature matrix of the value, and set the value of the element in the added column to 0.
  • the feature matrix with insufficient dimensions is filled with zeros, which can make the row and column dimensions of the feature matrix of different video samples consistent, so that the convolution kernel of the convolutional neural network can extract features from the feature matrix.
  • a column is added before the start column and after the end column, and filled with zeros, so as to extract edge information on the column dimension of the feature matrix.
  • Step 103 the model training device trains the convolutional neural network model based on a preset algorithm.
  • training the convolutional neural network model based on a preset algorithm may specifically include: randomly dividing the data set into a training set and a prediction set; wherein, the data set is based on the video user experience category data and the feature The matrix is obtained; the parameters of the convolutional neural network model are trained according to the training set; and the accuracy of the trained convolutional neural network model is verified according to the prediction set.
  • 5,000 samples are obtained to train the model, specifically, 1,000 video samples of video user experience categories 1 to 5 are obtained.
  • the constructed data set will be randomly divided into 70% of the data sets, a total of 3500 as a training set, and the other 30% of a total of 1500 data sets will be used as a prediction set.
  • training the parameters of the convolutional neural network model according to the training set may specifically include: using the training set to train a convolutional neural network classifier, wherein the convolutional neural network structure includes an input layer, a convolutional layer , pooling layer, full connection layer, softmax layer.
  • the input layer is the feature matrix
  • the convolution layer using multiple convolution kernels is used to automatically extract the local features of the feature matrix
  • the maximum pooling method is used as the pooling layer to reduce Training parameters
  • ReLU Rectified Linear Unit
  • the full connection layer to collect local features to extract global features
  • use the dropout method to randomly discard some neurons to reduce the probability of overfitting
  • the output layer is the video user experience label vector corresponding to the feature matrix.
  • Input the training set into the convolutional neural network use the multi-classification loss function cross entropy as the loss function, and finally use the stochastic gradient descent method to iteratively optimize the loss function until the number of iterations reaches the set number of iterations or the value of the loss function reaches the set value
  • the convergence threshold is when the training is complete.
  • a convolutional neural network algorithm is used to construct the classifier, and those skilled in the art can understand that other algorithms with local or global feature extraction can also be used to construct the classifier.
  • the Kappa coefficient is calculated as a parameter tuning feedback amount of a convolutional neural network classifier.
  • This convolutional neural network classifier is a multivariate classifier.
  • the prediction set is input into the trained convolutional neural network classifier to verify the prediction, and the Kappa coefficient is calculated to evaluate the performance of the multi-classifier, and the Kappa coefficient is used as the convolutional neural network classification.
  • the parameter tuning feedback of the device is as follows:
  • P 0 is the overall classification accuracy rate, that is, the ratio of the number of correctly classified samples to the number of samples in the prediction set;
  • the calculation method of P e is: the number of real samples of each class is a 1 ... a m , and the predicted The number of samples of each class is b 1 ...b m respectively, and the total number of samples is n, then:
  • the video user experience category is divided into 5 categories, and the number of samples in the prediction set is 1500, the number of real samples in each category is a1, a2, a3, a4, a5, and each predicted category
  • the trained convolutional neural network model is verified with a prediction set, and the obtained prediction results are shown in the table below.
  • the overall classification accuracy P0 is 0.929
  • the calculated value of Pe is 0.186
  • the Kappa value is 0.912.
  • the coefficient is low, adjust the parameters of the convolutional neural network and repeat the training until the Kappa coefficient calculated on the prediction set meets the requirements, that is, the video user experience prediction model is constructed.
  • video user experience is related to factors such as delay, throughput rate, block error rate, and packet loss rate, and due to different wireless resource scheduling and wireless environments at different times, the video quality presented to users is also different, which eventually leads to The difference in video user experience. Therefore, the video user experience during video viewing is affected by the combined effects of PDCP layer data, MAC layer data, and wireless performance data collected during video viewing. Therefore, local and global features of feature quantities in the time dimension need to be extracted for modeling.
  • the convolutional neural network model can effectively and adaptively extract local features and global features for modeling.
  • Step 104 the model training device acquires a video user experience prediction model according to the trained convolutional neural network model.
  • the video user experience prediction model is used to predict the video user experience in real time according to the wireless performance data at the base station side.
  • a feature matrix is generated according to wireless performance data collected at the base station side and generated by users watching videos, and used as training data for a convolutional neural network model. Then, the experience category data representing the video user experience collected on the user equipment side and the corresponding feature matrix generated based on the wireless performance data on the base station side are used as training data, and then the convolutional neural network model is trained with a preset algorithm. The trained convolutional neural network model is used as a video user experience prediction model.
  • the video user experience prediction model can output corresponding video user experience category data reflecting user viewing experience according to the wireless performance data collected in real time on the base station side, and realize accurate real-time prediction of video user viewing experience based on the base station side data.
  • An embodiment of the present application relates to a method for video user experience prediction, and the specific process is shown in FIG. 2 .
  • the device for predicting video user experience first collects base station side wireless performance data in real time; then generates a real-time feature matrix based on the real-time collected base station side wireless performance data; inputs the real-time feature matrix into a video user experience prediction model, Get the live video UX category.
  • step 201 the video user experience prediction device collects wireless performance data at the base station side in real time.
  • the wireless performance data is collected in real time on the base station side, which is used as the input data of the video user experience prediction model, and then the user's video viewing experience category is obtained. Data to realize the prediction of user viewing experience.
  • step 202 the prediction device generates a real-time characteristic matrix according to the real-time collected wireless performance data of the base station side.
  • the predicting device After collecting the real-time wireless performance data on the base station side, in order to input the video user experience prediction model to realize the prediction of user viewing experience, the collected wireless performance data needs to be processed into a form suitable for inputting the video user experience prediction model. Therefore, the predicting device generates a real-time feature matrix according to the wireless performance data collected in real time at the base station side.
  • Step 203 the prediction device inputs the real-time feature matrix into the video user experience prediction model to obtain the real-time video user experience category.
  • the video user experience prediction device inputs the real-time feature matrix generated in the previous step into the video user experience prediction model, and the video user experience prediction model can output real-time video user experience category data to represent the viewing experience of video users.
  • the real-time video user experience category is obtained to realize real-time prediction of the user's video viewing experience, so that operators and the like can optimize video playback according to the user experience, so that users can improve the video viewing experience.
  • the embodiment of the present application can collect the base station side wireless performance data generated by the user watching the video in real time. Then, according to the wireless performance data collected in real time, a real-time feature matrix is generated as the input data of the video user experience prediction model. Then the real-time feature matrix is input into the video user experience prediction model, and the output video user experience category data that can represent the user viewing experience is obtained, that is, the real-time prediction of the video user viewing experience is realized.
  • An embodiment of the present application relates to a model training device, as shown in FIG. 3 , including the following modules.
  • the collection module 301 is configured to collect video user experience category data on the user equipment UE side, and collect base station side wireless performance data generated when users watch videos on the base station side.
  • the generation module 302 is configured to generate a feature matrix according to the base station side wireless performance data generated when the user watches the video.
  • the training module 303 is used to train a convolutional neural network model based on a preset algorithm, and use the trained convolutional neural network model as a video user experience prediction model; wherein, the training data of the convolutional neural network model includes the The feature matrix and video user experience category data; wherein, the video user experience prediction model is used to predict video user experience in real time according to the wireless performance data at the base station side.
  • the model training device may further include: a data processing module (not shown in the figure), which is used to collect video user experience category data on the user equipment UE side, and collect base station side radio data generated when the user watches a video on the base station side. After the performance data, before generating the feature matrix according to the base station side wireless performance data generated when the user watches the video, delete the missing values and abnormal values in the base station side wireless performance data; The numerical non-discrete data in the side wireless performance data is normalized:
  • x i is the i-th original value of the wireless performance data column
  • x min is the minimum value of the wireless performance data column
  • x max is the maximum value of the wireless performance data column
  • x i new is the normalized value
  • the generation module 302 can also be configured to set the column content of the feature matrix to the value of the same feature of different data packets, and set the row content of the feature matrix to the value of different features of the same data packet value; wherein, the feature includes a radio performance feature extracted according to the collected radio performance data at the base station side, and a radio performance feature obtained by combining the extracted radio performance features.
  • the generation module 302 may also be configured to detect that the number of rows or columns of the feature matrix is less than the preset value, increase the rows or columns of the feature matrix until the number of rows and columns of the feature matrix are equal to the preset value, and set the element value in the increased row or column to 0; in the row number Add one column to the left side of the start column and the right side of the end column of the feature matrix with the number of sum columns equal to the preset value, and set the value of the element in the added column to 0.
  • the model training device provided in the embodiment of the present application can generate a feature matrix according to the wireless performance data collected at the base station side and generated by the user watching the video, and use it as the training data of the convolutional neural network model. Furthermore, the experience category data representing the video user experience collected on the user equipment side and the corresponding feature matrix generated based on the wireless performance data on the base station side are used as training data, and then the convolutional neural network model is trained with a preset algorithm. The trained convolutional neural network model is used as a video user experience prediction model.
  • the video user experience prediction model can output corresponding video user experience category data reflecting user viewing experience according to the wireless performance data collected in real time on the base station side, and realize accurate real-time prediction of video user viewing experience based on the base station side data.
  • An embodiment of the present application relates to a device for predicting video user experience, as shown in FIG. 4 , including the following modules.
  • the collection module 401 is configured to collect wireless performance data at the base station side in real time.
  • the generation module 402 is configured to generate a real-time characteristic matrix according to the wireless performance data collected in real time at the base station side.
  • the obtaining module 403 is configured to input the real-time feature matrix into the video user experience prediction model to obtain the real-time video user experience category.
  • the device for predicting video user experience provided in the embodiments of the present application can collect in real time base station side wireless performance data generated by users watching videos. Then, according to the wireless performance data collected in real time, a real-time feature matrix is generated as the input data of the video user experience prediction model. Then the real-time feature matrix is input into the video user experience prediction model, and the output video user experience category data that can represent the user viewing experience is obtained, that is, the real-time prediction of the video user viewing experience is realized.
  • modules involved in the above embodiments of the present application are logical modules.
  • a logical unit can be a physical unit, or a part of a physical unit, and can also Combination of physical units.
  • units that are not closely related to solving the technical problems proposed in the present application are not introduced in this embodiment, but this does not mean that there are no other units in this embodiment.
  • An embodiment of the present application also provides an electronic device, as shown in FIG. 5 , including at least one processor 501; and a memory 502 communicatively connected to the at least one processor 501; Instructions executed by one processor 501, the instructions are executed by at least one processor 501, so that at least one processor 501 can execute the above-mentioned model training method, or execute the above-mentioned video user experience prediction method.
  • the memory 502 and the processor 501 are connected by a bus, and the bus may include any number of interconnected buses and bridges, and the bus connects one or more processors 501 and various circuits of the memory 502 together.
  • the bus may also connect together various other circuits such as peripherals, voltage regulators, and power management circuits, all of which are well known in the art and therefore will not be further described herein.
  • the bus interface provides an interface between the bus and the transceivers.
  • a transceiver may be a single element or multiple elements, such as multiple receivers and transmitters, providing a means for communicating with various other devices over a transmission medium.
  • the data processed by the processor 501 is transmitted on the wireless medium through the antenna, and further, the antenna also receives the data and transmits the data to the processor 501 .
  • Processor 501 is responsible for managing the bus and general processing, and may also provide various functions including timing, peripheral interface, voltage regulation, power management and other control functions. And the memory 502 may be used to store data used by the processor 501 when performing operations.
  • Embodiments of the present application also provide a computer-readable storage medium storing a computer program.
  • the computer program is executed by the processor, the above-mentioned model training method is implemented, or the above-mentioned video user experience prediction method is executed.
  • a device which can be A single chip microcomputer, a chip, etc.
  • a processor processor
  • the aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc, etc., which can store program codes. .

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Signal Processing (AREA)
  • Multimedia (AREA)
  • Testing, Inspecting, Measuring Of Stereoscopic Televisions And Televisions (AREA)

Abstract

本申请涉及通信技术领域，公开了一种模型训练方法，视频用户体验预测方法，装置及电子设备。该模型训练方法包括：在用户设备UE侧采集视频用户体验类别数据，在基站侧采集用户观看视频时产生的基站侧无线性能数据；根据用户观看视频时产生的基站侧无线性能数据，生成特征矩阵；基于预设算法训练卷积神经网络模型，并将训练完成的卷积神经网络模型作为视频用户体验预测模型；其中，模型训练数据包括特征矩阵和视频用户体验类别数据；其中，视频用户体验预测模型用于根据基站侧无线性能数据，对视频用户体验进行实时预测。

Description

模型训练方法，视频用户体验预测方法，装置及电子设备
相关申请的交叉引用
本申请要求在2021年09月26日提交的中国专利申请第202111131645.5号的优先权，该中国专利申请的全部内容通过引用包含于此。
技术领域
本申请涉及通信技术领域，尤其涉及一种模型训练方法，视频用户体验预测方法，装置及电子设备。
背景技术
随着通信技术和移动互联网的快速发展，短视频等视频业务类型的多媒体应用得到了迅猛发展，这为以覆盖和容量为中心的传统网络建设模式提出了极大挑战。在新的需求下，网络KPI(Key Performance Indicator,关键绩效指标)得到了不断提升，但是用户体验却并未得到保障。如用户观看视频过程中的视频卡顿、掉线等均会对视频用户体验造成较大的影响。为了保证更好的服务质量，如今运营商已经渐渐地将网络运维关注的重点，从KPI指标转向用户的体验质量。
然而，传统的在用户侧获知用户观看视频体验的方式，对用户体验的获取较为滞后，也就无法及时对用户体验进行优化。
发明内容
本申请实施方式的目的在于提供一种模型训练方法，视频用户体验预测方法，装置及电子设备，用以建立一个能够根据基站侧的无线性能数据，实现对视频用户体验进行预测的模型。
为了解决上述问题，本申请的实施方式提供了一种模型训练方法，包括：在用户设备UE侧采集视频用户体验类别数据，在基站侧采集用户观看视频 时产生的基站侧无线性能数据；根据所述用户观看视频时产生的基站侧无线性能数据，生成特征矩阵；基于预设算法训练卷积神经网络模型，并将训练完成的所述卷积神经网络模型作为视频用户体验预测模型；其中，所述卷积神经网络模型的训练数据包括所述特征矩阵和视频用户体验类别数据；其中，所述视频用户体验预测模型用于根据所述基站侧无线性能数据，对视频用户体验进行实时预测。
本申请的实施方式还提供了一种视频用户体验预测的方法，包括：实时采集基站侧无线性能数据；根据实时采集的基站侧无线性能数据，生成实时特征矩阵；将所述实时特征矩阵输入视频用户体验预测模型，获取实时视频用户体验类别，其中，所述视频用户体验预测模型采用上述的模型训练方法训练得到。
本申请的实施方式还提供了一种模型训练装置，包括：采集模块，用于在用户设备UE侧采集视频用户体验类别数据，在基站侧采集用户观看视频时产生的基站侧无线性能数据；生成模块，用于根据所述用户观看视频时产生的基站侧无线性能数据，生成特征矩阵；训练模块，用于基于预设算法训练卷积神经网络模型，并将训练完成的所述卷积神经网络模型作为视频用户体验预测模型；其中，所述卷积神经网络模型的训练数据包括所述特征矩阵和视频用户体验类别数据；其中，所述视频用户体验预测模型用于根据所述基站侧无线性能数据，对视频用户体验进行实时预测。
本申请的实施方式还提供了一种视频用户体验的预测装置，包括：采集模块，用于实时采集基站侧无线性能数据；生成模块，用于根据实时采集的基站侧无线性能数据，生成实时特征矩阵；获取模块，用于将所述实时特征矩阵输入视频用户体验预测模型，获取实时视频用户体验类别，其中，所述视频用户体验预测模型采用上述的模型训练方法训练得到。
本申请的实施方式还提供了一种电子设备，包括：至少一个处理器；以及，与所述至少一个处理器通信连接的存储器；其中，所述存储器存储有可 被所述至少一个处理器执行的指令，所述指令被所述至少一个处理器执行，以使所述至少一个处理器能够执行上述的模型训练方法，或执行上述视频用户体验预测的方法。
本申请的实施方式还提供了一种计算机可读存储介质，存储有计算机程序，所述计算机程序被处理器执行时实现上述的模型训练方法，或实现上述视频用户体验预测的方法。
附图说明
一个或多个实施方式通过与之对应的附图中的图片进行示例性说明，这些示例性说明并不构成对实施方式的限定，附图中具有相同参考数字标号的元件表示为类似的元件，除非有特别申明，附图中的图不构成比例限制。
图1是根据本申请一实施方式中的模型训练方法的流程图；
图2是根据本申请一实施方式中的视频用户体验预测的方法的流程图；
图3是根据本申请一实施方式中的模型训练装置的结构示意图；
图4是根据本申请一实施方式中的视频用户体验的预测装置的结构示意图
图5是根据本申请一实施方式中的电子设备的结构示意图。
具体实施方式
为使本申请实施例的目的、技术方案和优点更加清楚，下面将结合附图对本申请的各实施方式进行详细的阐述。然而，本领域的普通技术人员可以理解，在本申请各实施方式中，为了使读者更好地理解本申请而提出了许多技术细节。但是，即使没有这些技术细节和基于以下各实施方式的种种变化和修改，也可以实现本申请所要求保护的技术方案。
本申请的一实施方式涉及一种模型训练方法，具体流程如图1所示。在本实施方式中，模型训练装置在用户设备UE侧采集视频用户体验类别数据， 在基站侧采集用户观看视频时产生的基站侧无线性能数据；根据所述用户观看视频时产生的基站侧无线性能数据，生成特征矩阵；基于预设算法训练卷积神经网络模型，并将训练完成的所述卷积神经网络模型作为视频用户体验预测模型；其中，所述卷积神经网络模型的训练数据包括所述特征矩阵和视频用户体验类别数据；其中，所述视频用户体验预测模型用于根据所述基站侧无线性能数据，对视频用户体验进行实时预测。
下面对本实施例中的模型训练方法的实现细节进行具体的说明，以下内容仅为方便理解本方案的实现细节，并非实施本方案的必须。具体流程如图1所示，可包括如下步骤。
步骤101，模型训练装置在用户设备侧采集视频用户体验类别数据，在基站侧采集无线性能数据。
具体地说，模型训练装置首先在用户设备侧采集视频用户体验类别数据。用户体验类别数据，指评价视频用户体验的指标，可选取某种特定的视频用户体验评价标准，也可自定义某种视频用户体验评价标准，按该标准对视频进行评价，以获得该视频的视频用户体验类别数据。在具体实施时，可以将获得的视频用户体验类别数据及其对应的基站侧无线性能数据构成一个样本。
在一个例子中，视频用户体验类别数据可以采用国际电信联盟P.910协议定义的视频主观质量评分标准。该标准将视频用户体验分成5类，具体视频用户体验类别划分如表1所示。
表1视频用户体验类别划分
Figure PCTCN2022119734-appb-000001
值得一提的是，为了防止样本不均衡导致模型泛化能力差，应对视频用户体验的五种类别获取数量接近的视频样本。在一个例子中，需要获取5000个视频样本对模型进行训练，则优选获取视频用户体验类别为1至5的视频样本各1000个。
在基站侧采集用户观看视频产生的无线性能数据，采集的无线性能数据可以包括分组数据汇聚协议(Packet Data Convergence Protocol，简称“PDCP”)层数据和多址接入信道(Multiple Access Channel，简称“MAC”)层数据。PDCP层数据包括数据包时间戳、数据包时延、数据包传输方向上)/下行、数据包大小、数据包序列号(Serial Number，简称“SN”)值等；MAC层数据包括调度资源块(Resource Block，简称“RB”)数、调度次数、调制与编码策略(Modulation and Coding Scheme，简称“MCS”)、确认字符(Acknowledge character，简称“Ack”)和否定应答信号(Negative Acknowledgement，简称“Nack”)反馈等。通过测量装置、测量报告或者UE的反馈信息获得无线性能数据信噪比(Signal-Noise Ratio，简称“SNR”)、时延、参考信号接收功率(Reference Signal Receiving Power，简称“RSRP”)、信道质量指示(Channel Quality Indication，简称“CQI”)等。
需要说明的是，本领域技术人员可以理解，在基站侧采集的无线性能数据不仅可选取PDCP层数据、MAC层数据，也可从PDCP的高层或者物理层选取能反映PDCP层数据、MAC层数据的其它无线性能数据。也可在已有特征的基础上增加相关的冗余特征，或者将已有的特征进行组合或压缩。
在一个例子中，在UE侧采集视频用户体验类别数据，在基站侧采集用户观看视频时产生的基站侧无线性能数据之后还包括：删除所述基站侧无线性能数据中的缺失值和异常值；采用以下公式对所述基站侧无线性能数据中的数值型非离散数据进行归一化处理：
Figure PCTCN2022119734-appb-000002
其中，x i为无线性能数据列的第i个原始值，x min为所述无线性能数据列的最小值，x max为所述无线性能数据列的最大值，x i new为归一化后的值；将所 述基站侧无线性能数据中的数值型或非数值型类别离散数据，以及视频用户体验类别数据，进行一位有效编码。
在本例中，若在上述步骤中根据采集的基站侧无线性能数据能够构建起15维的特征，则对基站侧无线性能数据中的数值型或非数值型类别离散数据，进行一位有效编码可以具体为：将数值型或非数值型类别离散数据进行一位有效编码成数值型数据和状态稀疏矩阵。PDCP层的上下行方向，上行采用0表示，下行采用1表示；PDCP层数据包的SN值用于反映丢包情况，用0表示丢包，1表示未丢包；MAC层数据的ACK和NACK反馈中，用0表示NACK，用1表示ACK等等；MCS用29维状态稀疏矩阵将0～28的取值进行编码；CQI用16维状态稀疏矩阵将0～15的取值进行编码等。
在本例中，对视频用户体验类别数据进行一位有效编码可以具体为：对于视频用户体验类别为1至5的情况，采用5维状态稀疏矩阵分别编码为[1，0，0，0，0]、[0，1，0，0，0]、[0，0，1，0，0]、[0，0，0，1，0]、[0，0，0，0，1]。
此处需要说明的是，对所述基站侧无线性能数据中的数值型或非数值型类别离散数据，以及视频用户体验类别数据的一位有效编码，也可以采用其它的编码算法，如贝叶斯目标编码、留一法编码等。本例对采集的基站侧无线性能数据和用户体验类别数据进行处理，以对采集的数据进行清洗，使得数据适用于建立特征矩阵，进而适用于作为训练数据对卷积神经网络模型进行训练。
步骤102，模型训练装置根据基站侧无线性能数据，生成特征矩阵。
具体地说，为了对卷积神经网络模型进行训练，模型训练装置将采集的基站侧无线性能数据处理转化成特征矩阵，使得其适用于作为训练数据输入卷积神经网络模型。
在一个例子中，根据所述用户观看视频时产生的基站侧无线性能数据，生成特征矩阵，可以具体包括：将所述特征矩阵的列内容设置为不同数据包 的同一个特征的值，将所述特征矩阵的行内容设置为同一数据包的不同特征的值；其中，所述特征包括根据采集的所述基站侧无线性能数据提取的无线性能特征，以及将所述提取的无线性能特征组合得到的无线性能特征。在本例中可以实现对特征矩阵中元素的具体设置。
本例中涉及的将所述提取的无线性能特征组合得到无线性能特征，包括：将PDCP层数据包时间戳和数据包SN值反映是否丢包，组合并计算出丢包率；将数据包时间戳和MAC层数据Ack和Nack反馈组合并计算出误块率；将数据包时间戳和数据包大小组合并计算出吞吐率等。
需要说明的是，本例采用了特征组合的方式构建了多维组合特征，在具体实施中，也可采用特征降维的方式如主成分分析(Principal Component Analysis，简称“PCA”)方法，进行特征组合。
在一个例子中，在所述将所述特征矩阵的行设置为同一数据包的不同特征的值后，还包括：在检测到特征矩阵的行数或列数小于预设值的情况下，增加所述特征矩阵的行或列直至所述特征矩阵的行数和列数均等于预设值，将增加的行或列中的元素值设置为0；在所述行数和列数等于预设值的特征矩阵的开始列左侧和结束列右侧各增加一列，将增加列中的元素值设置为0。
本例对于维度不足的特征矩阵采用补零方式填充，能够使得不同视频样本的特征矩阵行列维度保持一致，以便于卷积神经网络的卷积核从特征矩阵中提取特征。另外，在开始列之前和结束列之后各增加一列，并采用补零方式填充，以便于提取特征矩阵列维度上的边缘信息。
步骤103，模型训练装置基于预设算法训练卷积神经网络模型。
在一个例子中，基于预设算法训练卷积神经网络模型，可以具体包括：将数据集随机划分为训练集和预测集；其中，所述数据集根据所述视频用户体验类别数据和所述特征矩阵得到；根据所述训练集对所述卷积神经网络模型的参数进行训练；根据所述预测集对训练后的所述卷积神经网络模型的准确性进行验证。
在一个具体实例中，获取5000个样本对模型进行训练，具体获取视频用户体验类别为1至5的视频样本各1000个。将构建好的数据集，随机划分70％数据集共计3500个作为训练集，另外30％数据集共计1500个作为预测集。
在一个例子中，根据所述训练集对所述卷积神经网络模型的参数进行训练可以具体包括：使用训练集训练卷积神经网络分类器，其中卷积神经网络结构包括输入层、卷积层、池化层、全链接层、softmax层。对卷积神经网络的具体描述如下：输入层为特征矩阵；使用多个卷积核的卷积层，用于自动提取特征矩阵的局部特征；使用最大池化方式作为池化层，用于减少训练参数；采用线性整流函数(Rectified Linear Unit,简称“ReLU”)作为激活函数；采用全链接层将局部特征归集提取全局特征；采用Dropout方式随机丢弃部分神经元以降低过拟合几率；最后在全链接层后的输出层之前加入Softmax层用于分类；输出层为特征矩阵对应的视频用户体验标签向量。将训练集输入至该卷积神经网络中，使用多分类损失函数交叉熵作为损失函数，最后采用随机梯度下降法迭代优化该损失函数，直到迭代次数达到设定迭代次数或者损失函数值达到设定收敛阈值即训练完成。
需要说明的是，本例采用卷积神经网络算法构建分类器，本领域技术人员可以理解，也可采用其它具有局部或者全局特征提取的算法构建分类器。
在一个例子中，计算Kappa系数作为卷积神经网络分类器的调参反馈量。此卷积神经网络分类器为多元分类器，将预测集输入训练得到的卷积神经网络分类器中验证预测，计算Kappa系数来评价多分类器的性能，并将Kappa系数作为卷积神经网络分类器的调参反馈量。Kappa系数计算公式如下：
Figure PCTCN2022119734-appb-000003
其中，P 0为总体分类正确率，即分类正确的样本数与预测集样本数的比值；P e的计算方法为：每一类的真实样本个数分别为a 1...a m，预测出来的每一类的样本个数分别为b 1...b m，总样本个数为n，则有：
Figure PCTCN2022119734-appb-000004
在一个将视频用户体验类别分为5类，且预测集样本数为1500的具体实例中，每一类的真实样本个数分别为a1、a2、a3、a4、a5，预测出来的每一类的样本个数分别为b1、b2、b3、b4、b5，总样本个数为n＝1500，则有：
Figure PCTCN2022119734-appb-000005
以预测集对训练后的所述卷积神经网络模型进行验证，得到的预测结果如下表所示。
表2视频用户体验预测结果
Figure PCTCN2022119734-appb-000006
可根据上述计算得到，总体分类正确率P 0为0.929，Pe的计算值为0.186，Kappa值为0.912。
如果系数低则调整卷积神经网络参数重复训练，直到预测集上计算的Kappa系数满足要求，即构建出视频用户体验预测模型。
由于视频用户体验与时延、吞吐率、误块率、丢包率等因素有关，且在不同时刻由于无线资源调度的不同、无线环境的不同，导致呈现给用户的视频质量也不同，最终导致视频用户体验的差别。所以视频观看期间的视频用户体验，受视频观看期间所采集的PDCP层数据、MAC层数据以及无线性能数据的综合影响，因此需提取特征量在时间维度上的局部特征和全局特征进行建模。采用卷积神经网络模型可以有效的自适应提取局部特征和全局特征进行建模。
步骤104，模型训练装置根据训练完成的卷积神经网络模型，获取视频用户体验预测模型。其中，所述视频用户体验预测模型用于根据所述基站侧无线性能数据，对视频用户体验进行实时预测。
本申请实施方式根据在基站侧采集的用户观看视频产生的无线性能数据生成特征矩阵，用以作为卷积神经网络模型的训练数据。进而将在用户设备 侧采集的表征视频用户体验的体验类别数据，以及对应的根据基站侧无线性能数据生成的特征矩阵作为训练数据，再以预设算法对卷积神经网络模型进行训练。将训练完成的卷积神经网络模型作为视频用户体验预测模型。该视频用户体验预测模型能够根据实时采集的基站侧无线性能数据，输出对应的反映用户观看体验的视频用户体验类别数据，实现根据基站侧数据对视频用户观看体验进行准确的实时预测。
本申请的一实施方式涉及一种视频用户体验预测的方法，具体流程如图2所示。在本实施方式中，视频用户体验的预测装置首先实时采集基站侧无线性能数据；进而根据实时采集的基站侧无线性能数据，生成实时特征矩阵；将所述实时特征矩阵输入视频用户体验预测模型，获取实时视频用户体验类别。
下面对本实施例中的视频用户体验预测的方法的实现细节进行具体的说明，以下内容仅为方便理解本方案的实现细节，并非实施本方案的必须。具体流程如图2所示，可包括如下步骤。
步骤201，视频用户体验预测装置实时采集基站侧无线性能数据。
在视频用户体验预测模型训练完成后，为对用户的视频观看体验进行实时预测，在基站侧实时采集无线性能数据，用以作为视频用户体验预测模型的输入数据，进而获取用户的视频观看体验类别数据，实现对用户观看体验的预测。
步骤202，预测装置根据实时采集的基站侧无线性能数据生成实时特征矩阵。
在采集到基站侧的实时无线性能数据后，为了使得能够输入视频用户体验预测模型以实现对用户观看体验的预测，需要将采集的无线性能数据处理为适用于输入视频用户体验预测模型的形式。因此，预测装置根据实时采集的基站侧无线性能数据生成实时特征矩阵。
步骤203，预测装置将实时特征矩阵输入视频用户体验预测模型，获取 实时视频用户体验类别。
具体地说，视频用户体验预测装置将在上一步骤中生成的实时特征矩阵输入视频用户体验预测模型，视频用户体验预测模型能够输出实时视频用户体验类别数据，用以表征视频用户的观看体验。获取实时视频用户体验类别，实现对用户视频观看体验的实时预测，进而以便于运营商等根据用户体验对视频播放进行优化，以使得用户提高视频观看的体验。
本申请实施方式能够对用户观看视频产生的基站侧无线性能数据进行实时采集。进而根据实时采集的无线性能数据生成实时特征矩阵作为视频用户体验预测模型的输入数据。再将实时特征矩阵输入视频用户体验预测模型，获取输出的能够表征用户观看体验的视频用户体验类别数据，即实现了对视频用户观看体验的实时预测。
本申请的一实施方式涉及一种模型训练装置，如图3所示，包括以下模块。
采集模块301，用于在用户设备UE侧采集视频用户体验类别数据，在基站侧采集用户观看视频时产生的基站侧无线性能数据。
生成模块302，用于根据所述用户观看视频时产生的基站侧无线性能数据，生成特征矩阵。
训练模块303，用于基于预设算法训练卷积神经网络模型，并将训练完成的所述卷积神经网络模型作为视频用户体验预测模型；其中，所述卷积神经网络模型的训练数据包括所述特征矩阵和视频用户体验类别数据；其中，所述视频用户体验预测模型用于根据所述基站侧无线性能数据，对视频用户体验进行实时预测。
在一个例子中，模型训练装置还可以包括：数据处理模块(图中未使出)，用于在用户设备UE侧采集视频用户体验类别数据，在基站侧采集用户观看视频时产生的基站侧无线性能数据之后，在所述根据所述用户观看视频时产生的基站侧无线性能数据，生成特征矩阵前，删除所述基站侧无线性能数据 中的缺失值和异常值；采用以下公式对所述基站侧无线性能数据中的数值型非离散数据进行归一化处理：
Figure PCTCN2022119734-appb-000007
其中，x i为无线性能数据列的第i个原始值，x min为所述无线性能数据列的最小值，x max为所述无线性能数据列的最大值，x i new为归一化后的值；将所述基站侧无线性能数据中的数值型或非数值型类别离散数据，以及视频用户体验类别数据，进行一位有效编码。
在一个例子中，生成模块302，还可以用于将所述特征矩阵的列内容设置为不同数据包的同一个特征的值，将所述特征矩阵的行内容设置为同一数据包的不同特征的值；其中，所述特征包括根据采集的所述基站侧无线性能数据提取的无线性能特征，以及将所述提取的无线性能特征组合得到的无线性能特征。
在另一个例子中，生成模块302，还可以用于在所述将所述特征矩阵的行设置为同一数据包的不同特征的值后，在检测到特征矩阵的行数或列数小于预设值的情况下，增加所述特征矩阵的行或列直至所述特征矩阵的行数和列数均等于预设值，将增加的行或列中的元素值设置为0；在所述行数和列数等于预设值的特征矩阵的开始列左侧和结束列右侧各增加一列，将增加列中的元素值设置为0。
本申请实施方式提供的模型训练装置能够根据在基站侧采集的用户观看视频产生的无线性能数据生成特征矩阵，用以作为卷积神经网络模型的训练数据。进而将在用户设备侧采集的表征视频用户体验的体验类别数据，以及对应的根据基站侧无线性能数据生成的特征矩阵作为训练数据，再以预设算法对卷积神经网络模型进行训练。将训练完成的卷积神经网络模型作为视频用户体验预测模型。该视频用户体验预测模型能够根据实时采集的基站侧无线性能数据，输出对应的反映用户观看体验的视频用户体验类别数据，实现根据基站侧数据对视频用户观看体验进行准确的实时预测。
本申请的一实施方式涉及一种视频用户体验的预测装置，如图4所示， 包括以下模块。
采集模块401，用于实时采集基站侧无线性能数据。
生成模块402，用于根据实时采集的基站侧无线性能数据，生成实时特征矩阵。
获取模块403，用于将所述实时特征矩阵输入视频用户体验预测模型，获取实时视频用户体验类别。
本申请实施方式提供的视频用户体验的预测装置能够对用户观看视频产生的基站侧无线性能数据进行实时采集。进而根据实时采集的无线性能数据生成实时特征矩阵作为视频用户体验预测模型的输入数据。再将实时特征矩阵输入视频用户体验预测模型，获取输出的能够表征用户观看体验的视频用户体验类别数据，即实现了对视频用户观看体验的实时预测。
值得一提的是，本申请上述实施方式中所涉及到的各模块均为逻辑模块，在实际应用中，一个逻辑单元可以是一个物理单元，也可以是一个物理单元的一部分，还可以以多个物理单元的组合实现。此外，为了突出本申请的创新部分，本实施方式中并没有将与解决本申请所提出的技术问题关系不太密切的单元引入，但这并不表明本实施方式中不存在其它的单元。
本申请的实施例还提供一种电子设备，如图5所示，包括至少一个处理器501；以及，与所述至少一个处理器501通信连接的存储器502；其中，存储器502存储有可被至少一个处理器501执行的指令，指令被至少一个处理器501执行，以使至少一个处理器501能够执行上述模型训练方法，或执行上述视频用户体验预测的方法。
其中，存储器502和处理器501采用总线方式连接，总线可以包括任意数量的互联的总线和桥，总线将一个或多个处理器501和存储器502的各种电路连接在一起。总线还可以将诸如外围设备、稳压器和功率管理电路等之类的各种其他电路连接在一起，这些都是本领域所公知的，因此，本文不再对其进行进一步描述。总线接口在总线和收发机之间提供接口。收发机可以 是一个元件，也可以是多个元件，比如多个接收器和发送器，提供用于在传输介质上与各种其他装置通信的单元。经处理器501处理的数据通过天线在无线介质上进行传输，进一步，天线还接收数据并将数据传送给处理器501。
处理器501负责管理总线和通常的处理，还可以提供各种功能，包括定时，外围接口，电压调节、电源管理以及其他控制功能。而存储器502可以被用于存储处理器501在执行操作时所使用的数据。
上述产品可执行本申请实施例所提供的方法，具备执行方法相应的功能模块和有益效果，未在本实施例中详尽描述的技术细节，可参见本申请实施例所提供的方法。
本申请的实施例还提供一种计算机可读存储介质，存储有计算机程序。计算机程序被处理器执行时实现上述模型训练方法，或执行上述视频用户体验预测的方法。
本领域技术人员可以理解，实现上述实施例方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成，该程序存储在一个存储介质中，包括若干指令用以使得一个设备(可以是单片机，芯片等)或处理器(processor)执行本申请各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、磁碟或者光盘等各种可以存储程序代码的介质。
上述实施例是提供给本领域普通技术人员来实现和使用本申请的，本领域普通技术人员可以在不脱离本申请的申请思想的情况下，对上述实施例做出种种修改或变化，因而本申请的保护范围并不被上述实施例所限，而应该符合权利要求书所提到的创新性特征的最大范围。

Claims (11)

  1. 一种模型训练方法，包括：
    在用户设备UE侧采集视频用户体验类别数据，在基站侧采集用户观看视频时产生的基站侧无线性能数据；
    根据所述用户观看视频时产生的基站侧无线性能数据，生成特征矩阵；
    基于预设算法训练卷积神经网络模型，并将训练完成的所述卷积神经网络模型作为视频用户体验预测模型；其中，所述卷积神经网络模型的训练数据包括所述特征矩阵和视频用户体验类别数据；
    其中，所述视频用户体验预测模型用于根据所述基站侧无线性能数据，对视频用户体验进行实时预测。
  2. 根据权利要求1所述的模型训练方法，其中，所述在用户设备UE侧采集视频用户体验类别数据，在基站侧采集用户观看视频时产生的基站侧无线性能数据之后，在所述根据所述用户观看视频时产生的基站侧无线性能数据，生成特征矩阵前，还包括：
    删除所述基站侧无线性能数据中的缺失值和异常值；
    采用以下公式对所述基站侧无线性能数据中的数值型非离散数据进行归一化处理：
    Figure PCTCN2022119734-appb-100001
    其中，x i为无线性能数据列的第i个原始值，x min为所述无线性能数据列的最小值，x max为所述无线性能数据列的最大值，x i new为归一化后的值；
    将所述基站侧无线性能数据中的数值型或非数值型类别离散数据，以及视频用户体验类别数据，进行一位有效编码。
  3. 根据权利要求1所述的模型训练方法，其中，所述根据所述用户观看视频时产生的基站侧无线性能数据，生成特征矩阵，包括：
    将所述特征矩阵的列内容设置为不同数据包的同一个特征的值，将所述 特征矩阵的行内容设置为同一数据包的不同特征的值；
    其中，所述特征包括根据采集的所述基站侧无线性能数据提取的无线性能特征，以及将所述提取的无线性能特征组合得到的无线性能特征。
  4. 根据权利要求3所述的模型训练方法，其中，在所述将所述特征矩阵的行设置为同一数据包的不同特征的值后，还包括：
    在检测到特征矩阵的行数或列数小于预设值的情况下，增加所述特征矩阵的行或列直至所述特征矩阵的行数和列数均等于预设值，将增加的行或列中的元素值设置为0；
    在所述行数和列数等于预设值的特征矩阵的开始列左侧和结束列右侧各增加一列，将增加列中的元素值设置为0。
  5. 根据权利要求3所述的模型训练方法，其中，所述将所述提取的无线性能特征组合得到的无线性能特征包括以下之一或其任意组合：吞吐率、丢包率、误块率。
  6. 根据权利要求1所述的模型训练方法，其中，所述基于预设算法训练卷积神经网络模型，包括：
    将数据集随机划分为训练集和预测集；其中，所述数据集根据所述视频用户体验类别数据和所述特征矩阵得到；
    根据所述训练集对所述卷积神经网络模型的参数进行训练；
    根据所述预测集对训练后的所述卷积神经网络模型的准确性进行验证。
  7. 一种视频用户体验预测的方法，包括：
    实时采集基站侧无线性能数据；
    根据实时采集的基站侧无线性能数据，生成实时特征矩阵；
    将所述实时特征矩阵输入视频用户体验预测模型，获取实时视频用户体验类别；其中，所述视频用户体验预测模型采用如权利要求1至6中任一项所述的模型训练方法训练得到。
  8. 一种模型训练装置，包括：
    采集模块，用于在用户设备UE侧采集视频用户体验类别数据，在基站侧采集用户观看视频时产生的基站侧无线性能数据；
    生成模块，用于根据所述用户观看视频时产生的基站侧无线性能数据，生成特征矩阵；
    训练模块，用于基于预设算法训练卷积神经网络模型，并将训练完成的所述卷积神经网络模型作为视频用户体验预测模型；其中，所述卷积神经网络模型的训练数据包括所述特征矩阵和视频用户体验类别数据；其中，所述视频用户体验预测模型用于根据所述基站侧无线性能数据，对视频用户体验进行实时预测。
  9. 一种视频用户体验的预测装置，包括：
    采集模块，用于实时采集基站侧无线性能数据；
    生成模块，用于根据实时采集的基站侧无线性能数据，生成实时特征矩阵；
    获取模块，用于将所述实时特征矩阵输入视频用户体验预测模型，获取实时视频用户体验类别；其中，所述视频用户体验预测模型采用如权利要求1至6中任一项所述的模型训练方法训练得到。
  10. 一种电子设备，包括：
    至少一个处理器；以及，
    与所述至少一个处理器通信连接的存储器；其中，
    所述存储器存储有可被所述至少一个处理器执行的指令，所述指令被所述至少一个处理器执行，以使所述至少一个处理器能够执行如权利要求1至6中任一项所述的模型训练方法，或执行如权利要求7所述的视频用户体验预测的方法。
  11. 一种计算机可读存储介质，存储有计算机程序，其中，所述计算机程序被处理器执行时实现如权利要求1至6中任一项所述的模型训练方法，或实现如权利要求7所述的视频用户体验预测的方法。
PCT/CN2022/119734 2021-09-26 2022-09-19 模型训练方法，视频用户体验预测方法，装置及电子设备 WO2023045886A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111131645.5A CN115883927A (zh) 2021-09-26 2021-09-26 模型训练方法，视频用户体验预测方法，装置及电子设备
CN202111131645.5 2021-09-26

Publications (1)

Publication Number Publication Date
WO2023045886A1 true WO2023045886A1 (zh) 2023-03-30

Family

ID=85720058

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/119734 WO2023045886A1 (zh) 2021-09-26 2022-09-19 模型训练方法，视频用户体验预测方法，装置及电子设备

Country Status (2)

Country Link
CN (1) CN115883927A (zh)
WO (1) WO2023045886A1 (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102685790A (zh) * 2012-05-22 2012-09-19 北京东方文骏软件科技有限责任公司 模拟用户行为的移动流媒体业务感知体验QoE的测评方法
CN102769551A (zh) * 2012-07-02 2012-11-07 深信服网络科技（深圳）有限公司 网络质量评测与网络优化的方法及系统
CN106604290A (zh) * 2016-12-19 2017-04-26 南京华苏科技有限公司 基于网页浏览的用户感知测评无线网络性能方法
US20190287031A1 (en) * 2018-03-17 2019-09-19 Wipro Limited Method and system for generating synchronized labelled training dataset for building a learning model
US20200053591A1 (en) * 2018-08-10 2020-02-13 Verizon Patent And Licensing Inc. Systems and methods for wireless low latency traffic scheduler
CN111401637A (zh) * 2020-03-16 2020-07-10 湖南大学 融合用户行为和表情数据的用户体验质量预测方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102685790A (zh) * 2012-05-22 2012-09-19 北京东方文骏软件科技有限责任公司 模拟用户行为的移动流媒体业务感知体验QoE的测评方法
CN102769551A (zh) * 2012-07-02 2012-11-07 深信服网络科技（深圳）有限公司 网络质量评测与网络优化的方法及系统
CN106604290A (zh) * 2016-12-19 2017-04-26 南京华苏科技有限公司 基于网页浏览的用户感知测评无线网络性能方法
US20190287031A1 (en) * 2018-03-17 2019-09-19 Wipro Limited Method and system for generating synchronized labelled training dataset for building a learning model
US20200053591A1 (en) * 2018-08-10 2020-02-13 Verizon Patent And Licensing Inc. Systems and methods for wireless low latency traffic scheduler
CN111401637A (zh) * 2020-03-16 2020-07-10 湖南大学 融合用户行为和表情数据的用户体验质量预测方法

Also Published As

Publication number Publication date
CN115883927A (zh) 2023-03-31

Similar Documents

Publication Publication Date Title
CN111447083B (zh) 动态带宽和不可靠网络下的联邦学习架构及其压缩算法
US11700518B2 (en) Methods and systems for relaying feature-driven communications
CN102802089B (zh) 一种基于体验质量预测的移动视频码率调整方法
CN112584417B (zh) 无线业务质量确定方法、装置、计算机设备和存储介质
WO2023051318A1 (zh) 模型训练方法、无线资源调度方法及其装置及电子设备
US20210241441A1 (en) Methods, Apparatus and Computer-Readable Mediums Relating to Detection of Cell Conditions in a Wireless Cellular Network
US11888703B1 (en) Machine learning algorithms for quality of service assurance in network traffic
CN105577836B (zh) 一种可穿戴设备的数据处理方法及可穿戴设备
CN106537959B (zh) 电信网络中编码和解码帧的方法
CN105812834A (zh) 基于聚类信息的视频推荐服务器、推荐方法和预缓存方法
Edwards et al. Quality of information-aware mobile applications
CN115037416A (zh) 数据前向纠错处理方法、装置、电子设备和存储介质
WO2018192414A1 (zh) QoE的确定方法、装置、存储介质及处理器
CN117749800B (zh) 新能源发电侧实现边缘数据存储与传输的方法和相关装置
US10785681B1 (en) Methods and apparatuses for feature-driven machine-to-machine communications
CN106817582B (zh) 在视频传输过程中动态优化视频质量的系统和方法
WO2023045886A1 (zh) 模型训练方法，视频用户体验预测方法，装置及电子设备
CN114064394A (zh) 基于边缘计算的安全监控方法、装置及终端设备
WO2024041117A1 (zh) 一种计算任务的分割方法及相关装置
CN117221894A (zh) 一种基于大数据的5g通信传输方法
EP3035759B1 (en) System and method for adaptive downlink scheduling for wireless networks
CN118540232B (zh) 基于工控设备的数据交互方法
EP4345699A1 (en) Training method for gan, machine learning system and communication apparatus
WO2023207920A1 (zh) 信道信息反馈方法、终端及网络侧设备
CN116017092B (zh) 一种QoE量化方法以及基于此的传输系统调参方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22871932

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22871932

Country of ref document: EP

Kind code of ref document: A1