Video stream pushing method and system based on cloud mobile phone and storage medium

Abstract

The invention relates to a video stream pushing method and system based on a cloud mobile phone and a storage medium, and the method comprises the steps: firstly collecting cloud mobile phone task data, spatial network features and service labels in real time, and generating preprocessing data through a noise reduction auto-encoder, LSTM interpolation and normalization processing; secondly, a bidirectional gating circulation unit is combined with an attention mechanism to extract time features to construct a spatial relation graph model to excavate a spatial propagation rule, the time features, the spatial features and a service type coding vector are input into a gating fusion unit to generate a fusion feature vector, and then a network output resource demand prediction result is predicted; and finally, executing three-level resource allocation based on a prediction result, accessing external risk early warning data, switching resource suppliers by an evaluation system, and encrypting and monitoring full-link data. According to the scheme, the CPU and memory utilization rate is improved, the plug flow delay and the resource allocation response time are reduced, the limitation of traditional static allocation is broken through, and dynamic resource optimization is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of cloud computing, and in particular relates to a video streaming method, system and storage medium based on a cloud phone. Background Art

[0002] Cloud phone video streaming generates video sources in a virtual cloud phone environment, encodes them, and then streams them to user terminals. Existing technologies use a static resource allocation model that only collects basic cloud phone operational data. This model lacks real-time monitoring of spatial network characteristics (such as regional user density and network quality) and service tags, and resource management strategies are unable to adapt to the sudden nature of streaming tasks and the diversity of service types.

[0003] Traditional solutions have multi-dimensional technical bottlenecks. For example, at the data level, the collection dimension is single and the pre-processing capability is insufficient, and spatiotemporal characteristics and business types are not integrated; at the forecasting level, only time series analysis is relied upon, which cannot capture the spatiotemporal distribution patterns of resource demand; at the scheduling level, there is a lack of a flexible control architecture, and idle resources and over-occupancy coexist; at the supplier management level, a dynamic evaluation and emergency response mechanism has not been established, and the system has poor robustness.

[0004] Due to data limitations, single-source prediction models, and scheduling lags in existing technologies, cloud phone CPU resource utilization is low on average, resulting in high streaming latency and difficulty meeting dynamic business needs. There is an urgent need to optimize resource allocation through multi-source data fusion, spatiotemporal feature prediction, and a hierarchical scheduling system to improve streaming efficiency and system stability. Summary of the Invention

[0005] In response to the defects and problems of the existing technology, the present invention provides a video streaming method, system and storage medium based on cloud phones. Through multi-source data fusion, spatiotemporal feature prediction and three-level elastic scheduling, it solves the problems of low resource utilization and allocation lag in cloud phone video streaming, and improves streaming efficiency and quality.

[0006] The present invention solves the technical problem by: A video streaming method based on a cloud phone is adopted, which includes the following steps: S1. Multi-source heterogeneous data collection and preprocessing: real-time data collection from the cloud phone cluster management platform, CDN nodes, and user behavior analysis system: task creation time, duration, and CPU / memory / network resource consumption curves; regional longitude and latitude, operator network quality, and regional user online density; live broadcast type and cloud game category labels; data denoising through denoising autoencoders, using LSTM models for time series interpolation of missing values, and normalizing the data to the [0,1] interval to generate preprocessed data.

[0007] S2. Temporal and spatial feature fusion and resource demand prediction based on preprocessed data. Based on the preprocessed data generated in step S1, a bidirectional gated recurrent unit is used in combination with an attention mechanism to extract temporal features and capture the periodicity, trend, and burstiness of tasks. A spatial relationship graph model is constructed, with regions as nodes, and the spatial propagation patterns of resource demands between regions are mined through a graph convolutional network. The temporal feature, spatial feature, and business type encoding vectors are input into the gated fusion unit to generate a fused feature vector. The fused feature vector is input into a prediction network based on the Transformer architecture, and the resource demand prediction results are output through multi-head attention mechanism and Huber loss function optimization.

[0008] S3. Dynamic resource pre-allocation and elastic regulation based on prediction results. Based on the resource demand prediction results output by step S2, three-level resource allocation is performed: short-term regulation dynamically adjusts cloud phone resource quotas through container orchestration tools; medium-term regulation formulates cross-regional task migration plans; long-term regulation applies to cloud service providers for expanded computing resources; real-time access to network public opinion and policy change warning data, and analysis of external risks through NLP technology; switching resource suppliers based on the supplier dynamic evaluation system, and establishing resource sharing agreements with high-quality suppliers; implementing transmission and storage encryption for full-link data, and deploying intrusion detection systems to monitor traffic in real time.

[0009] Preferably, the preprocessing of spatial data in step S1 further includes: constructing a Voronoi diagram based on regional longitude and latitude to divide the service area, using regional user online density as graph node weights, and dynamically adjusting edge connection weights in combination with the operator's network quality.

[0010] Preferably, the spatiotemporal feature fusion in step S2 specifically includes: performing tensor splicing on the temporal feature vector output by Bi-GRU and the spatial feature vector output by GCN; inputting the spliced ​​features and the business type coding vector into the gated fusion unit, generating feature selection weights through the Sigmoid function, and generating a fused feature vector by weighted summation.

[0011] Preferably, the execution of the supplier dynamic evaluation system in step S3 includes: defining a supplier comprehensive scoring formula: ;in: Indicates the supplier's comprehensive rating; Indicates the resource supply quality score, calculated based on CPU performance stability and memory compatibility indicators; Indicates the on-time delivery score, which is calculated based on the resource allocation response time compliance rate; Indicates the technical support capability score, which is calculated based on the troubleshooting efficiency and service level agreement (SLA) compliance rate; is the preset weight coefficient and satisfies ;when When the supplier is triggered, is the preset scoring threshold; if the supplier Second-rate Or if a major supply accident occurs, a new supplier will be called from the alternative supplier database as a replacement.

[0012] Preferably, step S3 also includes an extreme event response mechanism: building a resource supply risk simulation system to simulate supply and demand changes under emergency situations based on historical failure data; when a large-scale network attack or computer room failure is detected, idle resources of multiple cloud service providers are called upon through a resource mutual aid alliance.

[0013] On the other hand, it also includes a video streaming system based on cloud mobile phones, including the following modules: Data acquisition module: configured to execute step S1; Prediction engine module: Integrates Bi-GRU, GCN, and gated fusion units, and is configured to execute step S2. Resource scheduling module: Contains elastic scaling controller, task migration planner, and resource request interface, and is configured to execute step S3. Supplier management module: built-in supplier rating model and resource mutual assistance alliance interface; Security protection module: Deploy AES-256 full-link encryption and real-time intrusion detection system.

[0014] Preferably, the prediction engine module further includes: a time series cross-validation unit, which is used to divide the training set and the test set, and optimize the Transformer prediction network through a weighted mean square error function.

[0015] In another aspect, the present invention further discloses a computer-readable storage medium having a computer program stored thereon, wherein the program implements the steps of the method when executed by a processor.

[0016] The beneficial effects of the present invention are as follows: A fusion collection and preprocessing mechanism for multi-source heterogeneous data is adopted. Cloud phone task data, spatial network data and business labels are collected in real time through distributed probes. Combined with denoising autoencoders, LSTM time series interpolation and Min-Max normalization, a standardized feature data set is constructed to solve the problems of single dimension and high noise interference in traditional data collection.

[0017] A spatiotemporal feature fusion and intelligent prediction model is constructed, using Bi-GRU combined with the attention mechanism to extract time series features, building a spatial relationship graph based on GCN, integrating business type encoding through gating units, and using the Transformer architecture to achieve resource demand prediction, breaking through the limitations of traditional methods that rely solely on time series analysis.

[0018] Using a three-level elastic resource control architecture, we designed a hierarchical control system that includes short-term elastic expansion (dynamically adjusting resources based on CPU utilization), mid-term task migration (optimizing load balancing using genetic algorithms), and long-term gradient expansion (combining predicted demand with actual usage) to achieve forward-looking resource allocation.

[0019] The dynamic supplier management and emergency dispatch mechanism evaluates suppliers through a comprehensive scoring model (covering resource quality, delivery punctuality, and technical support). It combines resource mutual aid alliances with Bayesian network fault simulation to improve system robustness and address the response lag problem of traditional static supplier management. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is the system module data stream of Example 1 of the present invention; Figure 2 is a software flow chart of the method of embodiment 2 of the present invention; Figure 3 This is the gated fusion unit structure of the method of Example 2 of the present invention; Figure 4 is an emergency resource scheduling path diagram of the method of embodiment 2 of the present invention; Figure 5 2 is a step diagram of the method of embodiment 2 of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be further described below with reference to the accompanying drawings and examples.

[0022] Example 1: A video streaming system based on cloud phones is designed mainly to address the problem that existing technologies rely on static resource allocation and are difficult to cope with the suddenness of tasks and business differences, resulting in idle or over-occupied resources. This embodiment constructs a resource management paradigm that adapts to dynamic business changes through multi-dimensional data fusion, intelligent prediction and dynamic scheduling systems, breaking through the technical bottlenecks of traditional methods in the collaborative processing of spatiotemporal features, multi-level resource regulation and dynamic management of suppliers, and providing a systematic solution for improving cloud phone resource utilization and streaming quality. The video streaming system adopts a distributed microservice architecture, realizes intelligent management of cloud phone resources through modular design, and runs the video streaming method of Example 2. The core of the system consists of a data acquisition module, a prediction engine module, a resource scheduling module, a supplier management module and a security protection module. Each module realizes data interaction through a message bus, and the bottom layer relies on a container orchestration platform to achieve elastic deployment. The system architecture is as follows Figure 1 As shown in the figure, this layered design not only ensures the independence of modules, but also achieves collaborative work through standardized interfaces, and can meet the resource management needs of large-scale cloud phone clusters.

[0023] The data collection module serves as the system's data entry point, utilizing a distributed collection architecture to acquire multi-source, heterogeneous data in real time. A lightweight probe program is embedded in the cloud phone cluster management platform, collecting time series data such as task creation time, runtime, and CPU / memory / network resource consumption curves on a 1-second basis. Spatial data such as regional latitude and longitude, carrier network quality (latency / packet loss rate), and regional user online density are periodically pulled through the CDN node API. Event tracking is deployed within the user behavior analysis system to capture data related to business tags, such as live broadcast type and cloud gaming categories.

[0024] The data preprocessing stage adopts a three-stage pipeline mechanism. First, sensor noise is removed through the denoising autoencoder (DAE). This autoencoder adopts a three-layer structure of "input layer-bottleneck layer-output layer". The number of neurons in the input layer is consistent with the feature dimension. The bottleneck layer uses a 16-dimensional compressed representation, and noisy data is identified and filtered through reconstruction error. Secondly, the LSTM model is used to interpolate missing values ​​in time series. Dedicated models are trained for different resource indicators (CPU / memory / network). The feature vectors of the previous 10 time steps are input and the predicted value of the current missing point is output. Finally, all features are mapped to the interval [0,1] through Min-Max normalization to ensure the stability of subsequent model training.

[0025] The prediction engine module is the intelligent core of the system, integrating spatiotemporal feature fusion with resource prediction algorithms. Temporal feature extraction utilizes a bidirectional gated recurrent unit (Bi-GRU) architecture combined with an attention mechanism. The Bi-GRU consists of two layers of GRU units, with the forward and reverse networks capturing the forward and backward dependencies of the time series, respectively, and outputting a 128-dimensional temporal feature vector. The attention mechanism calculates the importance weight of each time step using the formula: , highlighting the characteristics of key time points such as business peaks. Among them, Wh and Wx is the weight matrix, obtained through training, Wh and Wx They act on different inputs respectively; tanh: is an activation function used to introduce nonlinearity to the model. It maps the input value to between -1 and 1, which helps the model learn complex patterns; ht :Indicates the t The hidden state vector of the time step is generated by the Bidirectional Gated Recurrent Unit (Bi-GRU) in the t Feature representation of the output of each time step; b : It is a bias term and also a model parameter used to adjust the output of the model.

[0026] Spatial features are extracted and a spatial relationship graph is constructed based on a graph convolutional network. Nodes are defined by the service areas delineated by the Voronoi diagram. Node attributes include metrics such as user density and network quality. Edge weights are dynamically determined by geographic distance (e.g., calculated using a Gaussian kernel function) and service relevance. The GCN network consists of two convolutional layers. The first layer maps node features to 64 dimensions, and the second layer outputs a 32-dimensional spatial feature vector.

[0027] The feature fusion unit uses a gating mechanism to concatenate 128-dimensional temporal features and 32-dimensional spatial features into a 160-dimensional vector. The service type code (32-dimensional embedding vector) is input into the gating unit and the feature selection weight is generated using the Sigmoid function. The formula is: , the final fusion feature is the weighted sum of the splicing feature and the weight. g is the weight vector calculated by the Sigmoid function. This weight vector will be used to weight the concatenated feature vector. σ It is an activation function whose output range is between 0 and 1 and is suitable for mapping input values ​​to probability values; Wg is the weight matrix used in the feature selection gating mechanism, which is multiplied by the input vector. [ t ; s ; b ] is the concatenation vector, where: t Represents the time feature vector, with a dimension of 128. s Represents the spatial feature vector, the dimension is 32, b Indicates the business type code, with a dimension of 32, represented by an Embedding vector. bg is the bias term, and the weight matrix Wg The prediction network is based on the Transformer architecture and consists of a six-layer encoder. Each layer consists of an eight-head attention mechanism and a feedforward neural network. The loss function uses a combination of Huber loss and weighted mean squared error (weighting 0.7:0.3). Time series cross-validation is used to optimize the model's generalization ability.

[0028] The resource scheduling module implements the implementation of the prediction results to resource allocation, and adopts a three-level control architecture, namely short-term elastic control, medium-term task migration and long-term resource expansion. Among them, short-term elastic control: monitor the cloud phone resource utilization through the container orchestration tool. When the CPU utilization exceeds 80% for 5 consecutive minutes, the hot expansion mechanism is triggered, and 0.5 vCPU or 2GB of memory is dynamically added to the instance through the libvirtAPI. After the adjustment, wait for 5 minutes to observe the effect to avoid frequent adjustments. Medium-term task migration: The task migration planner adopts a genetic algorithm based on load balancing. The chromosome encoding is the task-cloud phone mapping relationship. The fitness function comprehensively considers the regional load balancing and migration bandwidth consumption. When the average load of a region exceeds 70% and lasts for 30 minutes, the migration process is started, and non-real-time tasks are migrated first. Long-term resource expansion: The resource application interface is connected to the cloud service provider's API, and a gradient expansion strategy is adopted (the first application is 20% of the predicted demand, and the subsequent application is based on the formula every hour). Adjustment), to avoid excessive resource application, among which, Qt +1 means in t +1 The amount of resources requested at the moment, Qt Indicates in t The amount of resources requested at any given time, α Indicates the expansion ratio coefficient, which is used to control the ratio of resource expansion. Dt Indicates in t The demand forecast at any moment, Ut Indicates in t The actual amount of resources used at any given moment. The difference between the predicted demand and the actual usage determines the capacity expansion of resources. α The value of can control the speed and extent of resource expansion to avoid over-application or insufficient resources.

[0029] In addition, the module accesses online public opinion and policy change data in real time, performs sentiment analysis and keyword extraction through NLP technology, and triggers the emergency dispatch process when a risk event is detected, reallocating resources to protect core business.

[0030] The supplier management module has a built-in dynamic evaluation model to implement the comprehensive scoring formula: (Weight 0.4, 0.3, 0.3 respectively). Among them: Indicates the supplier's comprehensive rating; Indicates the resource supply quality score, calculated based on CPU performance stability and memory compatibility indicators; Indicates the on-time delivery score, which is calculated based on the resource allocation response time compliance rate; Indicates the technical support capability score, which is calculated based on the troubleshooting efficiency and service level agreement (SLA) compliance rate; is the preset weight coefficient and satisfies Resource Supply Quality Rating Calculated based on hardware indicators such as CPU performance stability and memory compatibility; delivery on-time rate score Based on the resource allocation response time compliance rate (e.g. 95% of requests completed within 10 seconds will receive 100 points); technical support capability score Combined with fault resolution efficiency and SLA compliance, an alert is triggered when the score falls below 80. If the score falls below 70 for three consecutive times or a major incident occurs, a new supplier is automatically selected from the backup supplier database, switching resource sources through hot migration technology.

[0031] The module also maintains a resource mutual assistance alliance interface. When a large-scale network attack or computer room failure is detected, it queries the idle resources of each service provider through the alliance protocol and uses a greedy algorithm to prioritize core business needs.

[0032] The security protection module provides end-to-end security: AES-256 encryption is used at the data collection end, TLS 1.3 is used at the transport layer, and sensitive information is encrypted at the field level during storage (keys are rotated every 24 hours). The intrusion detection system combines anomaly detection and pattern matching: anomaly detection uses an autoencoder to learn normal traffic characteristics (reconstruction errors exceeding a threshold trigger an alarm), while pattern matching maintains an up-to-date attack signature library (such as DDoS and SQL injection patterns). The security audit module logs key operations (resource allocation, vendor switching) and supports traceability queries within six months.

[0033] Each module in the above system implements closed-loop management through the data middle platform. The data pre-processed by the data acquisition module is stored in the time series database. The prediction engine module reads the data regularly to generate prediction results, which are pushed to the resource scheduling module to perform three-level regulation; the scheduling results are fed back to the supplier management module to update the score, and the security protection module encrypts and verifies the security of all data interactions.

[0034] Example 2: A method for streaming video on a cloud phone based on the system described in Example 1. This method mainly solves the problem of delayed resource allocation on cloud phones through the technical route of "data collection-feature fusion-intelligent scheduling". Figure 5 As shown in the figure, 101-multi-source heterogeneous data collection and preprocessing step; 102-temporal and spatial feature fusion and resource demand prediction step based on preprocessed data; 103-dynamic resource pre-allocation and elastic regulation step based on prediction results. The software flow chart of this method is as follows Figure 2 As shown, it includes three core stages: multi-source data collection and preprocessing, spatiotemporal feature fusion prediction, and dynamic resource allocation and risk response. Each stage transmits information through standardized data structures to form an intelligent management closed loop.

[0035] The first stage is the multi-source heterogeneous data collection and preprocessing step.

[0036] The data collection process adopts a distributed probe architecture: an agent program is deployed on the cloud phone management platform to obtain the task creation timestamp, running time, CPU utilization curve (recorded every 100ms), memory allocation histogram and network bandwidth throughput log at a sampling frequency of 1 second; through the GPS positioning and network quality monitoring module of the CDN edge node, the regional latitude and longitude, operator network latency (average / peak), packet loss rate and regional user online density (statisticed once every minute) are obtained; a business tag collector is implanted in the video streaming application to capture semantic data such as live broadcast type (game / e-commerce / education) and cloud game category (3DMMO / casual mobile games) in real time.

[0037] Spatial data preprocessing specifically utilizes the Voronoi diagram method: First, all collected latitude and longitude coordinates are imported into a PostGIS spatial database. Delaunay triangulation is then used to generate Voronoi primitives, each representing a service area. Regional user online density is used as node weight, and a Gaussian kernel function is used to calculate the impact of regional distance on edge weights. Edge weights are dynamically adjusted based on the quality of the operator's network to reflect the actual correlation between resource demands between regions. For example, if the network latency from region A to region B increases by 10ms, the edge weight connecting regions A and B will decrease by 5%, reflecting the change in the correlation between resource demands between the two regions.

[0038] Data denoising and normalization: The denoising autoencoder is implemented using the TensorFlow framework. The number of neurons in the input layer is consistent with the feature dimension (e.g., 100 dimensions), and the bottleneck layer is compressed to 16 dimensions. The model is trained by minimizing the reconstruction error (MSE). The LSTM interpolation model uses an "input layer (10 dimensions) - LSTM layer (128 units) - fully connected layer (1-dimensional output)" structure based on the CPU resource consumption curve. The teacher-forcing strategy is used during training, and the learning rate is set to 0.001. Finally, the Min-Max normalization formula is used. Map all features to the interval [0,1]. X is the original data value, Xmin is the minimum value in the data set, Xmax is the maximum value in the data set, X’ is the normalized value.

[0039] In the second stage, the spatiotemporal feature fusion and resource demand prediction steps are as follows: Figure 3 shown.

[0040] Temporal feature extraction uses a combination of Bi-GRU and attention mechanism architecture: the Bi-GRU network processes time series with a length of 1440 (24 hours × 60 minutes), each layer of GRU units contains 128 neurons, and outputs a concatenated vector (256 dimensions) of the forward and reverse hidden states; the attention mechanism uses the weight matrix and Calculate the importance score of each time step (formula: ), the weight vector is obtained after Softmax normalization , the final time feature is , the dimension is 128. Among them, et Indicates that at time step t The importance score of Wh The weight matrix is ​​used to map the hidden state to the score space through nonlinear transformation, and the tanh hyperbolic tangent function is a commonly used activation function used to introduce nonlinearity. Wx The weight matrix is ​​used to transform the hidden state ht Mapped to the score space, ht At time step t The hidden state vector contains the information of this time step, b Bias vector, used to adjust the baseline value of the score, t The final time feature vector has a dimension of 128, ∑ t =1 T For all time steps t From 1 to T Perform the summation, at At time step t The weight vector is obtained by Softmax normalization, which indicates the importance of this time step. ht At time step t The hidden state vector of .

[0041] Spatial feature extraction based on GCN to build regional association graph: node feature matrix Contains indicators such as regional user density, network delay and packet loss rate (5-dimensional features per node), adjacency matrix It is composed of the Voronoi diagram edge weights and business relevance (trained by historical collaborative allocation data); the convolution operation of GCN is defined as ,in The first layer maps the features to 64 dimensions, and the second layer outputs a 32-dimensional spatial feature vector , H (l+1) :Indicates the l +1 layer node feature matrix, σ : represents the activation function, which is usually used to introduce nonlinearity so that the model can learn complex patterns. : represents the normalized degree matrix, which is the original degree matrix D After normalization, D Elements Di is a node i degree (i.e., the degree of the node i The number of connected edges), : represents the normalized adjacency matrix, which is the original adjacency matrix A Add the identity matrix I After that, we get ,This is done to take into account self-loops, i.e. connections from a node itself to itself, H (l) :Indicates the l The node feature matrix of the layer. W (l) :Indicates the l The weight matrix of the layer is a parameter matrix that needs to be trained. l The feature map of the layer is l +1 layer.

[0042] The feature fusion process is to combine the 128-dimensional time features and 32-dimensional spatial features The data is concatenated into a 160-dimensional vector, and the business type label is pre-trained with Word2Vec to generate a 32-dimensional encoding vector. ; The gated fusion unit generates a weight vector through the fully connected layer (192→160) and the Sigmoid activation function , the final fusion feature is ,in, f Represents the final fused feature vector, with a dimension of 160. g Represents the weight vector, which is generated by the gated fusion unit and the activation function, and the dimension is also 160. ⊙ represents the element-level multiplication. [ t ; s ] means that the time feature vector t and spatial eigenvectors s Splice to form a 160-dimensional vector, t Represents the time feature vector, with a dimension of 128. s Represents the spatial feature vector, the dimension is 32, b The encoding vector representing the business type label is generated through Word2Vec pre-training and has a dimension of 32, but may need to be expanded to 160 dimensions when splicing to match [ t ; s ] dimension, (1- g ) represents 1 minus the weight vector g Elements of , used to calculate and encode vectors bMultiply the weights, and finally, fuse the features f is a weighted combination of time-space features and business type label encoding vector, where g Decided t ; s ]and b The degree of contribution to the final fusion feature.

[0043] The prediction network is based on the Transformer architecture and contains a six-layer encoder. Each layer consists of an eight-head attention mechanism (each head has a dimension of 64) and a two-layer feedforward network (160→512→160). Position encoding uses sine-cosine functions. The loss function is a linear combination of Huber loss (to handle outliers) and weighted mean square error. Training uses an early stopping strategy (stopping if the validation set loss does not decrease after five rounds). The final output is a resource demand forecast vector for the next hour (predicted values ​​of CPU / memory / network bandwidth).

[0044] The third stage is the dynamic resource pre-allocation and elastic control steps, such as Figure 4 shown.

[0045] Short-term control: The HorizontalPodAutoscaler component of Kubernetes monitors the resource indicators of the cloud phone container. When the CPU utilization exceeds the 80% threshold, the VerticalPodAutoscaler (VPA) is triggered to dynamically adjust the requested resources, adding 0.5 vCPUs or 2GB of memory each time. The adjustment interval is set to 5 minutes to avoid frequent fluctuations.

[0046] Mid-term control: The task migration planner is optimized using a genetic algorithm. The chromosome encoding is a mapping of tasks to cloud phones (e.g., 100 tasks correspond to 100 genes). The fitness function is ( is the load balancing degree, is the migration bandwidth consumption, weight ), the algorithm generates the optimal migration plan after 100 generations of iteration. Pre-replication technology is used during the migration process to reduce service interruption time.

[0047] Long-term control: The resource application interface is connected to the API of service providers such as Alibaba Cloud and Tencent Cloud, and an incremental expansion strategy is adopted (formula: ,in is the smoothing coefficient, To forecast demand, is the current utilization rate), the initial application volume is 20% of the forecast value, and is adjusted every hour thereafter.

[0048] Supplier dynamic evaluation strictly follows the scoring model: regular (daily) calculation of supplier comprehensive scores ,when Send an early warning notice and ask the supplier to submit an improvement plan within 48 hours; if the score is If a major supply incident occurs (such as a resource interruption lasting more than 30 minutes), the supplier with the highest score will be selected from the alternative library, and resource supply will be switched through hot migration technology. During the switch, cache resources will be enabled to ensure business continuity.

[0049] The extreme event response mechanism builds a Bayesian network model based on historical failure data (cyber attacks, computer room power outages, natural disasters) to simulate changes in resource supply and demand under sudden situations; when a large-scale network attack (traffic exceeds 10 times the baseline) or a computer room failure (more than 30% of cloud phones lose connection) is detected, the resource mutual aid alliance interface is immediately activated, and idle resources within the alliance are allocated according to priority (live broadcast > cloud game > other), and non-core businesses are downgraded (such as reducing the streaming quality and delaying upload).

[0050] The above-mentioned method in this embodiment integrates a joint prediction model of spatiotemporal dimensions and business characteristics, overcoming the limitations of traditional methods that only consider time series. A three-level elastic scheduling mechanism based on prediction enables forward-looking resource allocation. Dynamic supplier management and resource mutual assistance alliances enhance system robustness. Actual application tests have shown that this method increases cloud phone CPU resource utilization from an average of 55% to over 75%, increases memory utilization by 28%, reduces the average video streaming latency from 450ms to 270ms, and stabilizes resource allocation response time within 500ms during promotional periods.

[0051] Example 3: The computer-readable storage medium uses an enterprise-class SSD array architecture with RAID 5 data redundancy to ensure program storage reliability and read / write performance. The video streaming method described for the program stored in the medium utilizes a microservices architecture design, encompassing six service modules: data acquisition, feature calculation, prediction engine, resource scheduling, supplier management, and security protection. Each module is encapsulated in a Docker container, supporting flexible deployment on the Kubernetes platform.

[0052] When the processor loads and executes the program from the storage medium, it first performs initial configuration: reading model parameters (such as the hidden layer dimensions of the Bi-GRU and the number of convolution kernels of the GCN), scheduling thresholds (such as triggering short-term control at 80% CPU utilization), and supplier scoring weights. It then enters a loop: the data collection service is launched every 10 seconds, acquiring raw data from various data sources. The preprocessing module is invoked to perform noise reduction, interpolation, and normalization, and the results are stored in the distributed file system (HDFS). The feature calculation service is triggered every 5 minutes, loading the latest 24 hours of preprocessed data. The Bi-GRU, GCN, and gated fusion units are invoked to generate fused features, which are then cached in Redis. Once the fused features are generated, the prediction service is immediately launched, using a Transformer network to generate a one-hour resource demand forecast and pushing the results to a message queue (RabbitMQ). The resource scheduling service listens to the message queue, executes a three-level control policy based on the forecast results, logs scheduling operations in a database, and feeds the results back to the supplier management service. The supplier scoring module is run at 2:00 AM daily to update supplier status. When a switch condition is triggered, the resource mutual aid alliance interface is invoked. The security protection module continuously monitors online public opinion and policy change interfaces, uses NLP technology to analyze risk events in real time, and immediately triggers the emergency dispatch process once an anomaly is detected.

[0053] In another aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0054] On the other hand, the present invention further discloses a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0055] It is understandable that the system, device and storage medium provided in the embodiments of the present invention correspond to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts of the above methods.

[0056] In the above embodiments, all or part of the embodiments can be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, hard disk, tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0057] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0058] Each embodiment in this specification is described in a related manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiment is generally similar to the method embodiment, so the description is relatively simple. For related parts, refer to the description of the method embodiment.

[0059] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A video streaming method based on cloud mobile phone, characterized in that: The following steps are involved: S1. Multi-source heterogeneous data collection and preprocessing: Real-time data collection from the cloud phone cluster management platform, CDN nodes, and user behavior analysis system includes task creation time, duration, and CPU / memory / network resource consumption curves; regional latitude and longitude, operator network quality, and regional user online density; Live broadcast type and cloud game category labels; Data denoising is performed using a denoising autoencoder, missing values ​​are interpolated using an LSTM model, and the data is normalized to the [0, 1] interval to generate preprocessed data; S2. Spatiotemporal feature fusion and resource demand prediction based on preprocessed data: Based on the preprocessed data generated in step S1, a bidirectional gated recurrent unit combined with an attention mechanism is used to extract temporal features to capture task periodicity, trend, and burstiness. A spatial relationship graph model is constructed, using regions as nodes. A graph convolutional network is used to explore the spatial propagation patterns of resource demand between regions. Temporal features, spatial features, and service type encoding vectors are input into a gated fusion unit to generate a fused feature vector. This fused feature vector is then input into a prediction network based on a Transformer architecture, which uses a multi-head attention mechanism and Huber loss function optimization to output resource demand prediction results. S3. Dynamic resource pre-allocation and elastic regulation based on prediction results: Based on the resource demand forecast results output in step S2, three-level resource allocation is performed: short-term regulation dynamically adjusts cloud phone resource quotas through container orchestration tools; medium-term regulation formulates cross-region task migration plans; Long-term regulation: Apply to cloud service providers to expand computing resources; Real-time access to online public opinion and policy change warning data, analyzing external risks through NLP technology; switching resource suppliers based on a dynamic supplier evaluation system, and establishing resource sharing agreements with high-quality suppliers; Implement transmission and storage encryption for the entire link data, and deploy an intrusion detection system to monitor traffic in real time.

2. The method according to claim 1, characterized in that The preprocessing of spatial data in step S1 also includes: constructing a Voronoi diagram based on regional longitude and latitude to divide the service area, using the regional user online density as the graph node weight, and dynamically adjusting the edge connection weight based on the operator network quality.

3. The method according to claim 1, characterized in that The spatiotemporal feature fusion in step S2 specifically includes: tensor splicing the temporal feature vector output by Bi-GRU and the spatial feature vector output by GCN; inputting the spliced ​​features and the business type coding vector into the gated fusion unit, generating feature selection weights through the Sigmoid function, and generating a fused feature vector by weighted summation.

4. The method according to claim 1, wherein The implementation of the supplier dynamic evaluation system in step S3 includes: defining a supplier comprehensive scoring formula: ； in: Indicates the supplier's comprehensive rating; Indicates the resource supply quality score, calculated based on CPU performance stability and memory compatibility indicators; Indicates the on-time delivery score, which is calculated based on the resource allocation response time compliance rate; Indicates the technical support capability score, which is calculated based on the troubleshooting efficiency and service level agreement (SLA) compliance rate; is the preset weight coefficient and satisfies ;when When the supplier is triggered, is the preset scoring threshold; if the supplier Second-rate Or if a major supply accident occurs, a new supplier will be called from the alternative supplier database as a replacement.

5. The method according to claim 1, wherein Step S3 also includes an extreme event response mechanism: building a resource supply risk simulation system to simulate supply and demand changes under sudden situations based on historical failure data; when a large-scale network attack or computer room failure is detected, idle resources of multiple cloud service providers are called upon through a resource mutual aid alliance.

6. A video streaming system based on cloud mobile phone, characterized in that: include: Data acquisition module: configured to perform step S1 of claim 1; Prediction engine module: integrating Bi-GRU, GCN and gated fusion unit, configured to perform step S2 of claim 1; Resource scheduling module: including an elastic scaling controller, a task migration planner and a resource application interface, configured to execute step S3 of claim 1; Supplier management module: built-in supplier rating model and resource mutual assistance alliance interface; Security protection module: Deploy AES-256 full-link encryption and real-time intrusion detection system.

7. The system according to claim 6, characterized in that The prediction engine module also includes: a time series cross-validation unit for dividing the training set and the test set, and optimizing the Transformer prediction network through a weighted mean square error function.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images