Energy saving method for playing short video streaming media in mobile terminal

Abstract

The present invention belongs to the field of software system technology and relates to an energy-saving method for playing short video streaming media on a mobile terminal. The method comprises the following steps: Step 1, using the E-Stream system to train a neural network; Step 2, deploying the trained neural network on a server; Step 3, collecting data from the environment to form an environmental state, and inputting the state into the neural network; Step 4, the neural network determines the output action, namely, the download order and bit rate of the video; Step 5, the server transmits the corresponding video block to the client for the user to play. The method provided by the present invention can significantly reduce the energy consumption during short video playback, extend the average video playback time, and ensure high quality of experience (QoE).

Description

Technical Field

[0001] The present invention belongs to the technical field of software systems and relates to an energy saving method for playing short video streaming media in a mobile terminal. Background Art

[0002] In recent years, with the rapid development of high-speed wireless networks, browsing the internet via mobile devices has become a common practice. Among mobile applications, short video streaming boasts the strongest user engagement. According to global statistics, today's mobile users spend an average of over 40% of their daily internet time watching short videos. Short video apps such as TikTok and Kuaishou have been ranked as the most popular mobile apps for many consecutive years. A common complaint among users of short videos is the rapid power consumption of their mobile devices. Measurements show that an average smartphone supports only six hours of continuous short video playback, with some models offering less than three hours. This energy consumption disrupts user immersion in video viewing, significantly reducing the quality of service experience and user engagement with short video streaming.

[0003] Measurement studies have shown that energy consumption is positively correlated with video resolution, so reducing the video bitrate can reduce energy consumption. However, this approach introduces a new problem: low bitrates limit video quality, significantly reducing Quality of Experience (QoE). Furthermore, the trade-off between QoE and energy consumption varies across different network environments, further complicating the issue. Summary of the Invention

[0004] In view of the problems existing in the prior art, the purpose of the present invention is to explore how to optimize the short video streaming algorithm at the software level to save energy and extend viewing time.

[0005] To achieve the above-mentioned object, the technical solution adopted by the present invention is: an energy saving method for playing short video streaming media in a mobile terminal, comprising the following steps:

[0006] Step 1: Use the E-Stream system to train the neural network;

[0007] Step 2: Deploy the trained neural network on the server;

[0008] Step 3, collecting data from the environment, forming an environmental state, and inputting it into the neural network;

[0009] Step 4: The neural network determines the output action, which is the download order and bit rate of the videos.

[0010] Step 5: The server transmits the corresponding video block to the client for the user to play.

[0011] Preferably, using the E-Stream system to train a neural network specifically includes:

[0012] (1) Divide the network throughput data of all videos into several sets using the time-series K-mean algorithm;

[0013] (2) Randomly select the network throughput data of several videos from each set and put them into two new sets, which are used as training datasets and validation datasets respectively;

[0014] (3) Use the training dataset to train the neural network through deep reinforcement learning algorithm.

[0015] Preferably, after each round of deep reinforcement learning training, the performance of the current neural network is evaluated using a validation dataset; the normalized performance of each set is calculated, and the set with the worst normalized performance is found;

[0016] The network throughput data of several videos are randomly selected from the set with the worst normalized performance and appended to the training dataset for the next round of deep reinforcement learning.

[0017] Preferably, the environmental state input by the neural network is used to describe network conditions and energy consumption information, including the throughput of downloading historical video blocks, the energy consumption of multiple video blocks in the past, the viewing probability of multiple videos pre-downloaded in the sliding window, the cache duration of multiple videos in the sliding window, and the image quality of the last downloaded video block of multiple videos in the sliding window.

[0018] Preferably, the output action of the neural network is defined as a combination of the download order and bit rate level of the pre-downloaded video; each output port of the neural network corresponds to an output probability value of one of the combinations, and the probability values ​​of all combinations form the output action probability distribution for video block k:

[0019] Ω k =Π k ·Δ k ;

[0020] Among them, Δ k Represents a shielding vector, whose element value is 0 or 1, 0 means ignoring the output probability of the corresponding port, and 1 means that the output probability of the corresponding port will be considered; Π k represents the output action probability distribution of the neural network for video block k; Ω k Represents the final output action probability distribution of the neural network.

[0021] Preferably, the element value decision rule of the shielding vector is a: if the cache length of the currently playing video is less than 1 second, the video must be downloaded first; rule b: the current video can only be downloaded when the previous video has started downloading; rule a has a higher priority than rule b.

[0022] The present invention provides an energy-saving method for playing short video streaming media in a mobile terminal, which can significantly reduce energy consumption during short video playback, extend the average video playback time, and ensure high QoE. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A schematic diagram of the structure of a neural network in an embodiment of the present invention;

[0024] Figure 2 Schematic diagram of a neural network training process in an embodiment of the present invention;

[0025] Figure 3 The following are the results of the performance evaluation of the neural network in the embodiment of the present invention; (a) the energy consumption of video playback on five models of smartphones; (b) the energy consumption of video downloading under three network conditions;

[0026] Figure 4 The QoE performance and energy consumption comparison of the neural network in the embodiment of the present invention on five types of mobile phones are as follows: (a) XCover; (b) Galaxy; (c) Nexus; (d) G5; (e) V20;

[0027] Figure 5 Comparison of QoE and energy consumption of the neural network in the embodiment of the present invention under three network conditions; (a) QoE; (b) energy consumption;

[0028] Figure 6 This is the QoE indicator evaluation of the neural network in the embodiment of the present invention; where (a) video quality; (b) playback freeze; (c) download energy consumption; (d) playback energy consumption;

[0029] Figure 7 Schematic diagram of the ablation experiment results of the method of the present invention. DETAILED DESCRIPTION

[0030] To facilitate understanding of the present invention, the present invention will be described in more detail below with reference to the accompanying drawings and specific embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. On the contrary, the purpose of providing these embodiments is to make the understanding of the present invention more thorough and comprehensive.

[0031] The present invention provides an energy-saving method for playing short video streaming media in a mobile terminal, which adopts a neural network to determine the video download order and bit rate of the short videos. The neural network is trained by a novel deep reinforcement learning method based on curriculum learning.

[0032] 1. Construction of Neural Network

[0033] The neural network structure used in the present invention includes: input layer, middle layer, and output layer. The input layer receives the environment state by a one-dimensional convolutional neural network (CNN). Then, all CNN outputs are centrally processed by the middle layer fully connected network (FCN). Finally, the output layer outputs the action decision. The neural network structure is as follows Figure 1 The neural network used in the present invention outputs download parameters of pre-downloaded videos by inputting environmental status.

[0034] 1. Environmental status

[0035] In a short video streaming system, the video client maintains a sliding window containing K subsequent videos, and the video chunks in these videos (video chunks are different segments that the video is cut into before transmission, for example, a 100s video is cut into many 2s video chunks) will be cross-downloaded. In this way, even if the viewer quickly switches the video, the buffer data will not be completely exhausted in a short time. At input, the neural network receives the environment state s k (for video block k) to capture network conditions, energy consumption information, etc. Figure 1 As shown, it contains five independent states, each of which is constructed as a vector: State 0: The throughput of downloading historical video blocks. It contains the past n video blocks, denoted as C k = <c k-1 , c k-2 ,...,c k-n >. Each element records the average throughput of downloading each video chunk (from chunk kn to chunk k-1), calculated by the chunk size and download duration. This state records the current network status.

[0036] State 1: Energy consumption of the past n video blocks, denoted as E k = <e k-1 , e k-2 ,...,e k-n >. Each element represents the sum of the download energy consumption and the playback energy consumption of a video chunk (from video chunk kn to video chunk k-1).

[0037] State 2: The viewing probability of the K pre-downloaded videos in the sliding window, denoted as P k = <p k-1 , p k-2,...,p k，K >. This state is obtained from large-scale user video viewing data, where each element corresponds to a video (i.e., video 1 to video K). This state reflects the viewer's video switching behavior.

[0038] State 3: The cache duration of K videos in the sliding window, denoted as B k = k，1 , b k，2 ,...,b k，K >. Each element corresponds to a video (i.e., video 1 to video K). This status is a warning of potential playback freezes.

[0039] State 4: The quality of the last downloaded video chunk of the K videos in the sliding window, denoted as V k = <v k，1 , v k，2 ,...,v k，K Each element corresponds to the VMAF score of a video block in a different video (i.e., video 1 to video K). This state is used to quantify the degree of video quality fluctuation between adjacent video blocks.

[0040] 2. Output action

[0041] like Figure 2 As shown, in order to realize the cross-downloading of short videos in the sliding window, the output action of the neural network is defined as the download order of the pre-downloaded videos ({τ i |i=1,…,K}) and bitrate levels ({σ r Each output port of the neural network outputs a probability value of one of the combinations, and all probability values ​​in the combination form a probability distribution for video block k:

[0042]

[0043] Experiments show that cross-downloading can indeed reduce the impact of video switching on performance, but it will also lead to download gap problems. For example, video i has been completely downloaded, but video i-1 has not yet started downloading. Since videos are played sequentially, this situation will cause playback to freeze. In order to solve this problem, the present invention proposes an action masking strategy based on the original action decision of the neural network, that is, formula (1). By introducing a masking vector (also called masking vector, represented by Δ k denoted) to change the probability distribution in formula (1) (for video block k). The masking vector contains elements of different dimensions, for example: Δ k =[z0, z1, ..., z n ​], where each dimension (element) corresponds to an output port of the neural network. Each element in the masking vector takes a value of 0 or 1, 0 means ignoring the output probability of the corresponding port, and 1 means that the output probability of the corresponding port will be considered. The present invention sets the following two rules to determine the element value of the masking vector, rule a: if the cache length of the currently playing video is less than 1 second, the video must be downloaded first; rule b: video i can only be downloaded when video i-1 has started downloading; rule a has a higher priority than rule b. The value of the masking vector is dynamically adjusted for the output probability of each video block, and then compared with the output probability distribution ∏ of the neural network k Perform dot product to form the final output action probability distribution:

[0044] Ω k =∏ k ·Δ k (2).

[0045] The final output action of the neural network will be based on Ω k Randomly selected.

[0046] 2. Neural Network Training

[0047] Studies have shown that neural networks trained through deep reinforcement learning outperform heuristic models in most cases, but perform poorly under certain specific network conditions. This performance imbalance significantly reduces the overall results. This problem mainly stems from the limitations of deep reinforcement learning training: on the one hand, if the parameter range of the training environment is large, for example, the average throughput or throughput fluctuation varies greatly in different time periods, it is difficult for the trained neural network to achieve the same good performance under all environmental conditions; on the other hand, if the parameter range of the training environment is small, the trained neural network may not be able to make reasonable decisions when encountering environmental conditions that have never been seen before, or even exhibit wrong behavior. Since the optimal trade-off between energy consumption and QoE is completely different in different network environments, how to train the neural network to balance its decision-making performance in different network environments is the key to the present invention.

[0048] This paper proposes a new training method based on course learning. Unlike traditional deep reinforcement learning, which has a fixed data distribution training environment, the training method of this paper can dynamically adjust the data distribution according to the performance of the neural network during training. The whole training process is as follows: Figure 2Specifically, to accelerate training, the E-Stream system uses virtual streaming to simulate the training environment of the neural network, in which network conditions are simulated by capturing throughput tracking data of real networks. In short video streaming, mobile network conditions can vary significantly over time even for a single viewer, not to mention the large number of viewers served by various mobile operators in constantly changing geographical locations. Therefore, the E-Stream system first uses the time series K-mean algorithm (denoted as TS-KMeans) to collect the network throughput data (denoted as ), divided into M sets {C m |m=0, 1, ..., M-1}:

[0049]

[0050] In each set, the network bandwidth data has similar time series variation characteristics, so it can be defined as having the same network conditions. Then from each set C m In the E-Stream system, the network throughput data of S videos are randomly selected (by t s denoted), and put them into a new set D m To construct the training dataset:

[0051] The training dataset is represented as

[0052] Use the training dataset and train the neural network (expressed as function T) through the most advanced deep reinforcement learning algorithm PPO PPO (.)) to optimize neuron weights:

[0053]

[0054] Among them, η neural Represents the trained neural network. After each round of deep reinforcement learning training, the E-Stream system will evaluate the performance of the current neural network, which is quantified by an energy efficiency function that combines streaming performance and energy consumption:

[0055]

[0056] Where, j represents the count of video chunks played in video i, and k represents the count of video chunks downloaded in video i (due to the swipe-to-switch operation of short video apps, the number of video chunks played and downloaded in a video session is usually not equal); R i，j 、S i,j 、P i，j represent the video quality, image quality fluctuation and playback energy consumption of the played video block j respectively; Zi,k 、W i,k They represent the freeze duration and download energy consumption of the downloaded video block k respectively; α1, α2, α3, α4, and α5 correspond to the weight coefficients of the above five indicators respectively.

[0057] In addition, the evaluation was performed on the validation dataset The construction of the validation dataset is similar to that of the training dataset:

[0058]

[0059] Where m=0, 1, ..., M-1 represents the set index, t x Represents the set C m Network throughput data for x randomly selected videos in .

[0060] After training, the neural network η neural In each validation dataset V m However, the energy efficiency function (6) can only get an absolute performance. Due to the difference in throughput of different sets, the absolute performance cannot reflect the pros and cons of different sets. To solve this problem, a standard is needed to normalize the absolute performance. Based on previous work, this paper designs a strategy to calculate the offline optimal performance, using η optimal The strategy is based on a dynamic programming algorithm, which fully knows the changes in future network throughput. Therefore, the performance upper limit can be calculated through this strategy. Based on this, the normalized performance g of each set is calculated by the function F(·) m :

[0061]

[0062] The E-Stream system collects the normalized performance of all sets, expressed as {g m |m=0,1,...,M-1}, and find the set index value m with the worst normalized performance * :

[0063]

[0064] In order to balance the performance of different sets in the next round of training, the E-Stream system will adjust the data distribution in the training dataset to enhance the set m * From the set with the worst normalized performance (C m* ) randomly selects network throughput data of y videos They are then appended (indicated by a "+") to the training dataset for the next round of deep reinforcement learning middle:

[0065]

[0066] By increasing the set m * The neural network's decision making in this set can be strengthened in the next round of training. Ultimately, after h rounds of training, until the total number of training rounds H is reached, the neural network will eventually show more balanced performance across all sets.

[0067] 3. Deployment of Neural Networks

[0068] The trained neural network is deployed on the server, collecting data from the environment to form an environment state. This environment state is fed into the neural network, which then outputs actions, namely the download order and bitrate of the videos. The server then transmits the corresponding video chunks to the client, which the user's player plays.

[0069] 1. Parameter settings:

[0070] E-Stream parameters: For the environment state of the neural network, it is defined that state 0 and state 1 contain 8 video block samples, and the other three states contain 5 video block samples. Each state is received by a one-dimensional convolutional neural network with 128 neurons, which outputs to two hidden layers, each with 128 neurons. The Adam optimizer is used for training, and the learning rate is set to 10 -4 Throughput data clustering uses the time-series K-means algorithm, with a total of 10 clusters. The video block length is set to 2 seconds, and the encoding bit rate ranges from 0.2 to 12.0 Mbps.

[0071] This paper implements four state-of-the-art streaming algorithms for performance comparison: two algorithms that are suitable for short video streaming but do not consider energy consumption: Dashlet and TikTok; two energy-aware algorithms but only suitable for long video streaming: GreenABR and RED.

[0072] 2. Experimental Dataset: In the experiment, three types of data are used, namely network throughput data, video data, and energy consumption data:

[0073] (1) Network throughput data is used to simulate network conditions. It covers 3G, 4G / LTE, 5G, and Wi-Fi networks and is collected in multiple scenarios (e.g., campus, subway, supermarket, etc.) and networks served by multiple mobile operators.

[0074] (2) Video data records video information, such as the physical duration of the video, the video switching time points in the viewer access log, etc. This video data is captured in the commercial server of a real short video streaming provider, containing three months of logs, covering 3 million visits from 120,000 viewers;

[0075] (3) Energy consumption data comes from the open source dataset of existing research work, namely, the energy consumption of video downloading and video playing under different mobile phone and network conditions, such as Figure 3 In this evaluation, 60% of the dataset is used for neural network training and 40% of the dataset is used for testing.

[0076] 3. Performance evaluation indicators:

[0077] (1) QoE performance is quantified by the following formula:

[0078] In video i, j and k are used to count the video chunks played and downloaded respectively; R j 、S j Respectively represent the video quality and image quality fluctuation of the playing video block j; Z k represents the playback freeze duration during the download of video chunk k, with a weight coefficient of σ = 6.0.

[0079] (2) Energy consumption is quantified by the following formula:

[0080] Among them, P i，j Indicates playback energy consumption, W i，k Indicates download energy consumption.

[0081] 4. Performance evaluation results:

[0082] Figure 4 The energy consumption and QoE performance of different algorithms on five smartphones were compared. The result values ​​were normalized to the range of 0 to 1. Specifically, the method of the present invention (E-Stream) performed best in all cases, achieving the highest QoE and the least energy consumption. It can reduce energy consumption by 27.1% to 48.2% without QoE loss, and accordingly, the streaming media playback time is extended by an average of 3.29 hours. In contrast, although the short video algorithms Dashlet and TikTok can achieve relatively high QoE, they do not consider energy-saving design, so the energy consumption is relatively high. Although GreenABR and RED take energy consumption issues into consideration and can achieve lower energy consumption, they are only aimed at long video streaming, so in the scenario of short video streaming, frequent video switching leads to their low QoE.

[0083] Figure 5 The energy consumption and QoE performance under three network conditions: poor, moderate, and good are described. Figure 5 Regarding the QoE performance in (a), we can see that the QoE of all algorithms improves significantly as the network conditions change from poor to good. For a deeper analysis, we evaluate the most important QoE indicators: video quality and playback freezes. The results are as follows: Figure 6 As shown in (a) and (b), the video quality improves as the network condition changes from poor to good due to the increase in average throughput, and the video freezes are reduced due to the more stable throughput.

[0084] Although the QoE of the method of the present invention (E-Stream) is better than all competing algorithms, the improvement over the QoE of Dashlet is very limited. This is because the method of the present invention achieves energy consumption reduction by sacrificing a certain amount of QoE. Figure 6 This is verified in Figure (a), which shows that the video quality of the proposed method (E-Stream) is consistently lower than that of the Dashlet algorithm because the lower bitrate reduces energy consumption. However, also due to the lower bitrate, the proposed method achieves less playback pauses. In contrast, the two long video algorithms, GreenABR and RED, experience more playback pauses (due to video switching), significantly reducing QoE. Figure 5 (b) compares the energy consumption. It can be seen that the energy consumption under poor and medium network conditions is similar, while the energy consumption under good network conditions is significantly higher. Since the energy consumption comes from video downloading and playing, Figure 6 Figures (c) and (d) compare the energy consumption for downloading and playback, respectively. In both cases, the energy consumption under good network conditions is significantly higher than under the other two network conditions. Among the five algorithms, our method consistently achieves the lowest energy consumption in all cases. This is due to its energy-aware design, which achieves the optimal trade-off between energy consumption and QoE under different network conditions.

[0085] To quantify the performance contribution of the main technical features of the present invention, an ablation study was conducted, i.e., removing each technical feature one by one and then evaluating the results. The three main technical features of interest were: (a) the action masking strategy (removing the masking vectors in the neural network); (b) curriculum learning (removing the curriculum learning portion of the training process); and (c) the reinforcement learning algorithm (using the A3C algorithm instead of the PPO algorithm). In addition, the full version of the present invention's method (E-Stream full version) was evaluated for performance comparison. Figure 7Their QoE, video quality, and stuttering were compared. First, "no curriculum learning" has the lowest QoE, which shows that curriculum learning plays a key role in training neural networks. It dynamically adjusts the distribution of training data according to the performance of the neural network during training, so that the neural network performs balanced in different environments and avoids incorrect decision-making behaviors in some cases. Secondly, the action masking strategy also shows significant effects because it effectively avoids the problem of video download gaps in cross-downloading, thereby effectively reducing stuttering. Third, the QoE achieved by training with the A3C algorithm is closest to the full version, but there is still a 5% gap. Neural networks trained by the A3C algorithm usually behave too aggressively, so more stuttering is produced compared to using the PPO algorithm. Overall, the QoE performance of these three variants is significantly lower than that of the method of the present invention.

Claims

1. A method for saving energy consumption when playing short video streaming media in a mobile terminal, characterized in that: The steps include: Step 1: Use the E-Stream system to train the neural network; specifically, (1) Divide the network throughput data of all videos into several sets using the time-series K-mean algorithm; (2) Randomly select the network throughput data of several videos from each set and put them into two new sets, which are used as training sets and validation sets respectively; (3) Using the training set, train the neural network using a deep reinforcement learning algorithm; Among them, after each round of deep reinforcement learning training, the performance of the current neural network is evaluated using the validation set; the normalized performance of each set is calculated, and the set with the worst normalized performance is found; the network throughput data of several videos are randomly selected from the set with the worst normalized performance and appended to the training set of the next round of deep reinforcement learning; the performance of the current neural network is determined by the energy efficiency function Quantification: ； in, Indicated in the video The count of video chunks played in Indicated in the video The number of video chunks downloaded in ; 、 and They are the video blocks played video quality, picture quality fluctuation and playback energy consumption; and Download video blocks The duration of freezes and download energy consumption; 、 、 、 、 is the weight coefficient; Step 2: Deploy the trained neural network on the server; Step 3, collecting data from the environment, forming an environmental state, and inputting it into the neural network; Step 4: The neural network determines the output action, i.e., the video download order and bit rate; Step 5: The server transmits the corresponding video block to the client for playback by the user's player.

2. The energy saving method for playing short video streaming media in a mobile terminal according to claim 1, characterized in that: The environmental state input to the neural network is used to describe network conditions and energy consumption information, including the throughput of downloading historical video blocks, the energy consumption of multiple video blocks in the past, the viewing probability of multiple videos pre-downloaded in the sliding window, the cache duration of multiple videos in the sliding window, and the image quality of the last downloaded video block of multiple videos in the sliding window.

3. The energy saving method for playing short video streaming media in a mobile terminal according to claim 1 or 2, characterized in that: The output action of the neural network is defined as a combination of the download order and bit rate level of the pre-downloaded video; each output port of the neural network corresponds to the output probability value of one of the combinations, and the probability values ​​of all combinations form the output action probability distribution for the video chunk of the pre-downloaded video ; The action masking strategy is used to adjust the output action probability distribution of each video block: ； in, Represents a mask vector, whose element value is 0 or 1. 0 means ignoring the output probability of the corresponding port, and 1 means the output probability of the corresponding port will be considered; represents the final output action probability distribution.

4. The energy saving method for playing short video streaming media in a mobile terminal according to claim 3, characterized in that: The element value decision rule of the shielding vector is a: if the cache length of the currently playing video is less than 1 second, the video must be downloaded first; rule b: the current video can only be downloaded if the previous video has started downloading; rule a has a higher priority than rule b.

Images