Live broadcast request scheduling method and system for active anomaly detection

Abstract

The invention relates to the technical field of live broadcast request scheduling, and provides a live broadcast request scheduling method and system for active anomaly detection, and the method comprises the steps: S1, designing a fault detection model employing a two-stage type of rule-based detection and learning-based detection according to the pre-collected historical load information of all servers under an edge cloud platform; s2, establishing a request distribution prediction model to predict the number and categories of requests in different time steps and different geographic areas in the future, and establishing a request service effect model to quantify the service effects of processing different categories of requests by different servers, including earning estimation and abnormal possibility evaluation; and S3, based on the fault detection model and the request service effect model, constructing a real-time request scheduling framework to schedule a user live broadcast service request arriving in real time. Through active anomaly detection and optimization of a real-time scheduling decision of a live broadcast request, the overall resource utilization rate of an edge cloud platform is maximized, and the influence of platform anomaly on a request service is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of live broadcast request scheduling, and in particular to a live broadcast request scheduling method and system with active anomaly detection. Background Art

[0002] With the widespread deployment of high-speed network technologies (such as LTE / 5G) and the popularity of advanced personal devices (such as the iPhone 14Pro), live streaming has become one of the most popular applications in recent years, with typical platforms including Twitch and Douyin. According to the Zippia report, by 2022, live streaming services accounted for 17% of global IP traffic and are expected to continue to grow rapidly. In addition, live streaming has gradually become part of the daily lives of Internet users, with users' total viewing time reaching 550 billion hours in 2021.

[0003] The rapid growth in demand for live streaming services has driven the rise of crowdsourced edge cloud platforms, which are used to cope with the increasing number of user requests. Crowdsourced edge cloud platforms serve a variety of live streaming applications by integrating widely distributed, heterogeneous computing and bandwidth resources. With their unique crowdsourcing nature, such platforms can expand server coverage at a lower cost and provide higher quality service experience through edge servers close to users, thus improving users' viewing experience.

[0004] In the edge cloud platform environment, making full use of computing resources and maximizing service capabilities under limited resource conditions is one of the core goals of platform operation. However, due to the heterogeneity and instability of the edge cloud platform, device and service anomalies often occur, affecting the scheduling effect. Therefore, an efficient request scheduling method should not only ensure the maximization of global resource utilization, but also combine active anomaly detection mechanisms to identify and avoid abnormal servers in advance. In addition, it is necessary to dynamically adjust the scheduling strategy according to the real-time situation to avoid extreme situations where resource utilization is too low or too high in a few servers, ensure the balanced allocation of system resources and the stability of scheduling, and thus optimize the benefits and performance of the overall platform.

[0005] At present, the live broadcast request scheduling methods commonly used in the industry are mainly based on the nearest scheduling strategy based on geographical location. Although such methods are simple to implement, they are limited in effectiveness when dealing with actual scenarios such as high device heterogeneity and changing business requirements in edge cloud platforms. In addition, these methods ignore the detection and processing of platform anomalies, making it difficult to achieve optimal global and local server resource utilization in complex and unstable environments. The specific problems are as follows:

[0006] (1) Single optimization goal: Existing methods usually only focus on the resource utilization of the global server, lack consideration of factors affecting server abnormalities and revenue, and cannot comprehensively measure the scheduling effect.

[0007] (2) Limited scenario applicability: Due to the highly heterogeneous resources of edge cloud devices, existing scheduling methods are difficult to cope with complex scenarios. The deployment effect of live broadcast requests (such as bandwidth consumption) is affected by multi-dimensional factors such as server hardware performance, geographical location, and business characteristics.

[0008] (3) Lack of proactive anomaly detection capabilities: Current scheduling methods are unable to predict or detect anomalies in the server or network before a request arrives, resulting in the inability to effectively adjust scheduling strategies, affecting the resource utilization and overall benefits of the platform. Summary of the invention

[0009] In view of the above problems, the purpose of the present invention is to provide a method and system for scheduling live broadcast requests with active anomaly detection, which can maximize the overall resource utilization of the edge cloud platform and reduce the impact of platform anomalies on request services by actively detecting anomalies and optimizing real-time scheduling decisions for live broadcast requests. In addition, this method is not only applicable to edge cloud scenarios, but can also be extended to multiple scenarios such as central cloud, IDC, CDN and terminal devices to optimize user request scheduling.

[0010] The above-mentioned object of the present invention is achieved through the following technical solutions:

[0011] A live broadcast request scheduling method for active anomaly detection includes the following steps:

[0012] S1: Based on the historical load information of all servers under the edge cloud platform collected in advance, a two-stage fault detection model is designed, which adopts rule-based detection and learning-based detection. First, the rule-based detection is used to quickly exclude servers with serious anomalies and screen out servers with potential anomalies. Then, the learning-based detection is used to re-evaluate the servers with potential anomalies.

[0013] S2: Establish a request distribution prediction model to predict the number and type of requests at different time steps and in different geographical areas in the future, and establish a request service effect model to quantify the service effect of different servers in processing different types of requests, including revenue estimation and abnormal possibility assessment;

[0014] S3: Based on the fault detection model and the request service effect model, a real-time request scheduling framework is constructed to schedule the real-time user live broadcast service requests.

[0015] Furthermore, in step S1, the rule-based detection is specifically:

[0016] The historical load information of each server is recorded using a sliding window S, where the monitoring time length is T, and the resource dimension N of the load sequence represents the number of variables, that is, the number of monitored resource types, s t,nis the load information at time point t and resource dimension n, and the sliding window S is defined as:

[0017] S={s t,n |t∈[1,T],n∈[1,N]}

[0018] In addition, select a healthy load sequence S h As a baseline, the rule-based detection follows the following two conditions:

[0019] A: Resource Fluctuation Rules

[0020] For each resource dimension n, calculate the variance fluctuation of the real-time resource usage rate and the healthy baseline. If the fluctuation exceeds the set threshold, it is determined that the current server has an abnormality in the current resource dimension n. The specific formula is:

[0021] Var(s t,n )-Var(s h,n )>θ v

[0022] Among them, Var(s t,n ) is the resource variance of the nth resource dimension in the current sliding window S, Var(s h,n ) is the resource variance of the healthy baseline, θ v is the set fluctuation threshold. If the condition is met, it means that the current resource dimension n has abnormal fluctuations;

[0023] B: Resource dependency rules

[0024] Calculate the correlation matrix between resource dimensions in the current sliding window S and compare it with the correlation of the healthy baseline. If the correlation difference exceeds the set threshold, it is determined that the current server has resource anomalies. The specific formula is:

[0025] ||R(S)-R(S h )||>θ r

[0026] Among them, R(S) is the correlation matrix of resource utilization in the current sliding window S, R(S h ) is the baseline correlation matrix, θ r is the threshold of correlation change. If the formula condition is met, it means that the resource usage pattern of the server is abnormal.

[0027] Based on resource fluctuation rules and resource correlation rules, quickly evaluate the status of the server:

[0028] If both conditions are met, the current server is judged to be seriously abnormal and unsuitable for scheduling, and the current server is excluded;

[0029] If both conditions are not met, the server is judged to be in a healthy state and the current server is allowed to participate in scheduling;

[0030] If only one of the conditions is met, there is a potential abnormality in the current server, and the learning-based detection step of the next stage is entered for re-evaluation.

[0031] Furthermore, in step S1, the learning-based detection is specifically:

[0032] A deep learning model based on the variational gated recurrent unit (VGRU) is used to model and analyze the multi-dimensional resource usage data of the server. The variational gated recurrent unit (VGRU) includes an inference phase and a generation phase:

[0033] The reasoning stage:

[0034] In the inference phase, the deep learning model uses information s according to the input resource usage information s at the current time step. t,n and the hidden state h at the previous time step t-1,n Infer the underlying random variable z t,n , the potential random variable is inferred by the following formula:

[0035] z t,n ∣s t,n ~N(μ z,t,n ,σ z,t,n )

[0036]

[0037] Among them, μ z,t,n and σ z,t,n are the mean and variance of the underlying random variable, respectively. and are the nonlinear transformation functions of the mean and variance, φ s is the feature vector extracted by the convolutional network, h t-1,n is the historical information of the previous time step;

[0038] At the same time, the reasoning stage integrates time series features through the mechanism of gated recurrent unit GRU to capture long-term dependencies in resource usage patterns;

[0039] The generation phase:

[0040] In the generation phase, the deep learning model uses the potential random variables z inferred by the inference phase. t,n , reconstruct the original input Used to calculate the reconstruction error to determine whether there is an anomaly. The formula for the generation process is:

[0041]

[0042] in, Represented by the potential random variable z t,n Generated reconstructed input, μ s,t,n and σ s,t,n are the mean and variance in the generation process, and is the nonlinear function of the decoder. The generated reconstructed data will be used to calculate the difference with the real input data to determine the abnormal state of the server. t-1,n is the historical information of the previous time step;

[0043] The deep learning model calculates the actual input data s t,n With the generated reconstructed data The reconstruction error between , the formula of the reconstruction error is:

[0044]

[0045] The reconstruction error reflects the degree of deviation between the current observation data and the normal mode learned by the deep learning model. If the reconstruction error exceeds the set threshold η, it is determined that the server is abnormal. The specific judgment rules are:

[0046]

[0047] Among them, η is the error threshold of anomaly detection, which is set based on the statistical distribution of historical data.

[0048] Furthermore, in step S2, the request distribution prediction model is established to predict the number and type of requests at different time steps and different geographical areas in the future, specifically:

[0049] The KMeans clustering method is used to classify historical live broadcast requests, and the heterogeneous requests are classified into multiple categories, and then the number of requests of each category is predicted respectively, wherein the KMeans clustering method is used to classify according to the request characteristics including content type, request platform and request time period;

[0050] Based on cluster analysis, a model combining spatiotemporal convolutional network (TCN) and graph convolutional network (GCN) is used to predict the number of requests at different times and regions for different categories of requests. The specific process is as follows:

[0051] Define a graph G=(V,E,A,R) containing multiple geographic regions, where V is a set of spatial nodes representing different regions, E is an edge set between regions, A is an adjacency matrix between nodes, and R is a historical request data matrix. The model predicts the number of requests for each region in the future time step through convolution operations. The formula is as follows:

[0052] H′=σ1(TCN1(R))⊙σ2(TCN2(R))

[0053]

[0054] Wherein, H′ is the feature matrix extracted by the spatiotemporal convolutional network TCN, TCN1(R) and TCN2(R) represent two different temporal convolution operations on the historical request matrix data matrix R, σ1 and σ2 are activation functions for performing nonlinear transformation on the convolution results, and ⊙ represents element-by-element multiplication;

[0055] D is a diagonal matrix, I is the identity matrix, A is the adjacency matrix for adaptive learning, is the normalization of the adjacency matrix A, Θ is the parameter matrix of the graph convolutional network, and H is the processed feature matrix;

[0056] W d is the weight matrix, b d is the bias vector, σ3 is the activation function, and finally the predicted number of requests at the future time t+1 is obtained.

[0057] Furthermore, in step S2, a request service effect model is established to quantify the service effects of different servers processing different types of requests, including revenue estimation and abnormal possibility assessment, specifically:

[0058] The request features and server attributes are matched through the cross-attention mechanism. The formula is as follows:

[0059] Q,V=φ enc1 (i,m,t),φ enc2 (e,B e ,L e )

[0060] α=softmax(Q)

[0061] Among them, Q represents the feature code of the request, V represents the feature code of the server, and φ enc1 and φ enc2 are two different encoding functions, i is the category of the request, m is the geographical location or area of ​​the request, t is the time of the request, e is the server identifier, and B e is the bandwidth capacity of the server, L e The historical status of the server, including historical load and abnormal result detection;

[0062] α is the attention weight, which is used to measure the compatibility between the server and the request. The server priority is dynamically adjusted through the attention mechanism to ensure that the most suitable server is selected for each request.

[0063] Based on the cross-attention mechanism, a request service effect matrix including a benefit matrix and an anomaly matrix is ​​generated. The benefit matrix is ​​used to quantify the potential benefits of different types of requests on different servers. The server bandwidth, location and load will affect the benefits. The anomaly matrix is ​​used to predict the anomalies that a request may encounter on a specific server, including device anomalies and service anomalies. The specific generation process is as follows:

[0064] A e,m,i =φ dec1 (Q+Dropout(α·V))

[0065] S e,n,i =φ dec2 (Q+Dropout(α·V))

[0066] Among them, A e,m,i represents the revenue generated by the i-th type of request on server e, S e,m,i Indicates the abnormal service status of the current request on server e, φ dec1 and φ dec2 is the decoding function, and Dropout is a technique to prevent overfitting.

[0067] Further, in step S3, based on the fault detection model and the request service effect model, a real-time request scheduling framework is constructed to schedule the real-time user live service requests, specifically:

[0068] First, for the live broadcast requests that come in real time, we use the clustering model to obtain the category of each request and record the geographic location of the request. represents the number of requests of the i-th category from region m;

[0069] Define the resource utilization maximization scheduling problem. Assume that there are E edge nodes and M locations. When facing real-time requests, the request scheduling strategy obtained in each round of scheduling is defined as

[0070] in represents the number of requests of type i served by server e at location m. Under this premise, the pre-scheduling problem of each scheduling cycle is:

[0071]

[0072] Formula (1) is the objective function, which aims to maximize the resource utilization of the entire edge cloud platform. Formula (2) ensures that the number of requests allocated is non-negative. Formula (3) ensures that all requests from any category i in any region m are All are served, and formula (4) ensures that the resource utilization of each server e does not exceed its optimal utilization threshold, avoiding server overload or low resource utilization;

[0073] E represents the number of edge nodes, M represents the number of locations, and N represents the number of requested categories. A represents the number of requests of type i served by server e at location m, e,m,i represents the revenue generated by the i-th type of request on server e, S e,m,i Indicates the abnormal service status of the current request on server e, B e represents a certain related resource amount of server e, D e is a weight coefficient associated with server e;

[0074] For all requests of any category i from any region m, U opt is the optimal utilization threshold.

[0075] Furthermore, step S3 also includes:

[0076] To simplify the problem, we first ignore the geographic location dimension of the request and Dimensionality reduction represents the number of requests of type i from all regions assigned to server e, while using represents the average revenue of server e when processing the i-th type of request. After the above simplification, we can simplify the above formula (1) as follows:

[0077]

[0078] Getting the intermediate results After that, we perform two processing steps to get PS:

[0079] A: Round to the nearest integer

[0080] B: Further refine the simplified two-dimensional strategy PS∈R according to the actual geographical distribution of the request. E×N Extension to three-dimensional strategy PS∈R E×M×N , in order to more accurately reflect the request distribution in different regions. The specific expansion formula is:

[0081]

[0082] After obtaining the PS, requests of corresponding categories and regions are assigned to corresponding servers, thereby maximizing the resource utilization of the edge cloud platform while avoiding scheduling to faulty servers.

[0083] A live broadcast request scheduling system for active anomaly detection for executing the live broadcast request scheduling method for active anomaly detection as described above, comprising:

[0084] The fault detection model building module is used to design a two-stage fault detection model using rule-based detection and learning-based detection based on the historical load information of all servers under the edge cloud platform collected in advance. First, the rule-based detection is used to quickly exclude servers with serious anomalies and screen out servers with potential anomalies, and then the learning-based detection is used to re-evaluate the servers with potential anomalies;

[0085] The service effect model building module is used to establish a request distribution prediction model to predict the number and type of requests at different time steps and different geographical areas in the future, and to establish a request service effect model to quantify the service effect of different servers processing different types of requests, including revenue estimation and abnormal possibility assessment;

[0086] The request invocation framework construction module is used to construct a real-time request scheduling framework based on the fault detection model and the request service effect model to schedule the real-time user live broadcast service requests.

[0087] A computer device comprises a memory and one or more processors, wherein the memory stores computer codes, and when the computer codes are executed by the one or more processors, the one or more processors execute the above method.

[0088] A computer-readable storage medium stores computer codes. When the computer codes are executed, the above method is executed.

[0089] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0090] (1) Resource utilization optimization model combined with anomaly detection: This method targets the resource utilization distribution of edge cloud servers, combines the active anomaly detection mechanism, and proposes a modeling method for the optimal utilization range. It can identify and exclude servers that may have anomalies before scheduling, ensuring resource utilization while optimizing benefits.

[0091] (2) Request prediction based on spatiotemporal and business characteristics: This method deeply explores the spatiotemporal and business characteristics of requests, combines the anomaly detection results, and proposes a method to fuse request characteristics through clustering methods and successfully predict the request service effect. This enables the system to estimate the scheduling effect in advance and optimize resource allocation.

[0092] (3) Double-constrained linear programming model combined with exception handling: This method models request scheduling as a linear programming problem with double constraints, and optimizes the global resource utilization of the platform and the utilization distribution of local servers at the same time. By combining anomaly detection data, abnormal servers can be effectively avoided, the stability and efficiency of scheduling can be guaranteed, and the optimization of real-time scheduling can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 The overall flow chart of the live broadcast request scheduling method for active anomaly detection of the present invention;

[0094] Figure 2 is a schematic diagram of a fault detection model of the present invention;

[0095] Figure 3 A schematic diagram of a request distribution prediction model of the present invention;

[0096] Figure 4 A schematic diagram of a service effect model for requesting the present invention;

[0097] Figure 5 This is an overall structural diagram of the live broadcast request scheduling system for active anomaly detection of the present invention. DETAILED DESCRIPTION

[0098] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0099] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0100] In the prior art, conventional request scheduling solutions have the following problems that need to be solved urgently:

[0101] (1) It only considers maximizing global resource utilization, without considering the constraints of local servers in real scenarios.

[0102] (2) The impact of unstable devices in the edge cloud on request scheduling is not considered.

[0103] (3) The impact of heterogeneous device attributes and complex request characteristics of edge clouds on request scheduling is not considered.

[0104] This method mainly has the following technical breakthroughs:

[0105] (1) A utilization optimal interval modeling method based on the resource utilization distribution of the edge cloud platform was designed, and the platform's actual profit factors were taken into consideration in the subsequent optimization problems.

[0106] (2) The key factors that affect the effect of request scheduling deployment are identified, potential abnormal scheduling is actively perceived, and an anomaly detection model based on request category and server attributes and a request scheduling effect prediction model are designed.

[0107] (3) Request scheduling is modeled as a nonlinear programming problem with multiple constraints, and a simplified strategy and branch-and-bound algorithm are adopted to solve request scheduling, thereby optimizing global resource utilization and reducing ineffective scheduling.

[0108] The method mainly includes the following invention points: (1) Identify potential faulty devices in the platform and build a detection model. (2) Build a real-time request scheduling framework and estimate all possible deployment performances through a prediction model to solve the problem of revenue optimization.

[0109] The professional terms involved in the present invention include:

[0110] K-means clustering, K-means clustering is a widely used unsupervised learning algorithm, which is used to divide the data set into several clusters (clusters), so that the data points in the same cluster have greater similarity, while the data points between different clusters are relatively different. The core idea of ​​K-means clustering is to find the K cluster centers (also called centroids) in the data set through an iterative method, and assign each data point to the cluster center closest to it, thereby achieving data clustering.

[0111] Integer Linear Programming,Integer Linear Programming is an optimization technique used to solve optimization problems involving integer variables under linear inequality or equality constraints. Mathematically, it can be expressed as:

[0112] minimize c T x

[0113] subject to Ax≤b

[0114]

[0115] Where c, A, b are known vectors or matrices, and x is the integer solution to be found. Integer linear programming is widely used in many fields such as supply chain optimization, network design and scheduling.

[0116] The branch and cut algorithm is an optimization algorithm for solving integer linear programming (ILP) problems that combines the branch and bound algorithm with the cutting plane method. This combination allows the algorithm to find integer optimal solutions more efficiently, especially for complex problems with a large number of integer and continuous variables. In the branch and cut algorithm, an initial solution is first obtained by solving a relaxed problem (i.e., a problem that ignores integer constraints). The algorithm then uses the cutting plane method to add new linear constraints (i.e., "cuts") to remove or reduce the possibility of non-integer solutions.

[0117] The following is described by specific examples:

[0118] First embodiment

[0119] like Figure 1 As shown, this embodiment provides a live broadcast request scheduling method for active anomaly detection, including the following steps:

[0120] S1: Based on the pre-collected historical load information of all servers under the edge cloud platform, a two-stage fault detection model is designed, which is rule-based detection and learning-based detection. First, the rule-based detection is used to quickly eliminate servers with serious anomalies and screen out servers with potential anomalies. Then, the learning-based detection is used to re-evaluate the servers with potential anomalies.

[0121] The main purpose of this step is to identify potential anomalies in the platform and design a detection model to identify servers that may fail in the platform. First, we collect historical load data of all servers under the edge cloud platform (including network bandwidth, CPU usage, and memory usage. Note that since this method is aimed at live broadcast services, the main load is network bandwidth consumption, so the resources in the following text refer to bandwidth) and design a fault detection model. The principle behind this is that when a device fails, it often cannot work properly, a certain load will mutate, and the correlation between various load indicators will also change.

[0122] like Figure 2 As shown in the figure, considering that the actual scheduling process is time-sensitive, the detection model should take up as little time as possible. For this purpose, we design a two-stage anomaly detector, including a rule-based detector (using the formulated rules to quickly screen out a part of obviously faulty servers) and a learning-based server (deep learning models are used to detect servers with potential faults that are difficult to detect).

[0123] S11: Rule-based detection, specifically:

[0124] The rule-based anomaly detection mechanism is used to quickly identify servers with possible hardware failures or resource anomalies in the edge cloud platform. As the first step of the scheduling strategy, it is used to exclude potential abnormal servers and ensure the stability and efficiency of scheduling. The detection mechanism evaluates the operating status of the server by monitoring the multi-dimensional use of server resources, combined with the set thresholds and fluctuation rules.

[0125] The historical load information of each server is recorded using a sliding window S, where the monitoring time length is T, and the resource dimension N of the load sequence represents the number of variables, that is, the number of monitored resource types, s t,n is the load information at time point t and resource dimension n, and the sliding window S is defined as:

[0126] S={s t,n |t∈[1,T],n∈[1,N]}

[0127] In addition, select a healthy load sequence S h As a baseline, the rule-based detection follows the following two conditions:

[0128] A: Resource Fluctuation Rules

[0129] For each resource dimension n, calculate the variance fluctuation of the real-time resource usage rate and the healthy baseline. If the fluctuation exceeds the set threshold, it is determined that the current server has an abnormality in the current resource dimension n. The specific formula is:

[0130] Var(s t,n )-Var(s h,n )>θ v

[0131] Among them, Var(s t,n ) is the resource variance of the nth resource dimension in the current sliding window S, Var(s h,n ) is the resource variance of the healthy baseline, θ v is the set fluctuation threshold. If the condition is met, it means that the current resource dimension n has abnormal fluctuations;

[0132] B: Resource dependency rules

[0133] The usage of server resources is often correlated. For example, under high load, CPU and bandwidth usage tend to increase simultaneously.

[0134] Therefore, the correlation matrix between the resource dimensions in the current sliding window S is calculated and compared with the correlation of the healthy baseline. If the correlation difference exceeds the set threshold, it is determined that the current server has resource anomalies. The specific formula is:

[0135] ||R(S)-R(Sh )||>θ r

[0136] Among them, R(S) is the correlation matrix of resource utilization in the current sliding window S, R(S h ) is the baseline correlation matrix, θ r is the threshold of correlation change. If the formula condition is met, it means that the resource usage pattern of the server is abnormal.

[0137] Furthermore, the meaning of the correlation matrix in this embodiment is: the correlation matrix is ​​a matrix used to describe the correlation between multiple variables. In the context of server resource usage, it describes the correlation between different resource dimensions (such as network bandwidth, CPU usage, memory usage, etc.). For example, it can represent the correlation between the change in network bandwidth usage and the change in CPU usage, and the correlation between the change in memory usage, etc. within the current monitoring window.

[0138] The calculation of the correlation matrix is ​​usually based on statistical methods. For server resource usage data, data of each resource dimension at each time point is collected within the current sliding window (i.e., within a specific time span and data point range).

[0139] Taking two resource dimensions A and B as an example, the commonly used method to calculate the correlation is to calculate their covariance and obtain the correlation coefficient through some standardization. For multiple resource dimensions, a matrix form will be formed. For example, if there are three resource dimensions A, B, and C, then the correlation matrix may be a 3×3 matrix, where the (i,j)th element represents the correlation coefficient between the i-th resource dimension and the j-th resource dimension (i,j=1,2,3). The calculation of this coefficient may be based on the sample covariance formula and combined with some standardization steps, such as dividing by the product of their respective standard deviations, to obtain a correlation coefficient between -1 and 1, thereby comprehensively describing the correlation relationship between resource dimensions. These calculations are based on real-time resource usage data collected within a sliding window.

[0140] After defining the resource fluctuation rule and resource correlation rule, quickly evaluate the server status according to the resource fluctuation rule and resource correlation rule:

[0141] If both conditions are met, the current server is judged to be seriously abnormal and unsuitable for scheduling, and the current server is excluded;

[0142] If both conditions are not met, the server is judged to be in a healthy state and the current server is allowed to participate in scheduling;

[0143] If only one of the conditions is met, there is a potential abnormality in the current server, and the learning-based detection step of the next stage is entered for re-evaluation.

[0144] S12: Learning-based detection, specifically:

[0145] Learning-based detection is used to handle servers that cannot be clearly identified through rule-based detection. This method uses deep learning models to identify complex abnormal patterns, especially suitable for capturing multi-dimensional resource usage behaviors with unstable and nonlinear characteristics, further improving detection accuracy and anomaly identification capabilities.

[0146] A deep learning model based on the Variational Gated Recurrent Unit (VGRU) is used to model and analyze the multi-dimensional resource usage data of the server. The model can effectively process time series data and identify potential anomalies by learning historical patterns. The Variational Gated Recurrent Unit VGRU includes an inference phase and a generation phase:

[0147] (1) The reasoning stage:

[0148] In the inference phase, the deep learning model uses information s according to the input resource usage information s at the current time step. t,n and the hidden state h at the previous time step t-1,n Infer the underlying random variable z t,n , the potential random variable is inferred by the following formula:

[0149] z t,n ∣s t,n ~N(μ z,t,n ,σ z,t,n )

[0150]

[0151] Among them, μ z,t,n and σ z,t,n are the mean and variance of the underlying random variable, respectively. and are the nonlinear transformation functions of the mean and variance, φ s is the feature vector extracted by the convolutional network, h t-1,n is the historical information of the previous time step;

[0152] At the same time, the reasoning stage integrates time series features through the mechanism of gated recurrent unit GRU to capture long-term dependencies in resource usage patterns;

[0153] (2) The generation stage:

[0154] In the generation phase, the deep learning model uses the potential random variables z inferred by the inference phase. t,n , reconstruct the original input Used to calculate the reconstruction error to determine whether there is an anomaly. The formula for the generation process is:

[0155]

[0156] in, Represented by the potential random variable z t,n Generated reconstructed input, μ s,t,n and σ s,t,n are the mean and variance in the generation process, and is the nonlinear function of the decoder. The generated reconstructed data will be used to calculate the difference with the real input data to determine the abnormal state of the server. t-1,n is the historical information of the previous time step;

[0157] In order to evaluate whether the server is abnormal, the deep learning model calculates the actual input data s t,n With the generated reconstructed data The reconstruction error between , the formula of the reconstruction error is:

[0158]

[0159] The reconstruction error reflects the degree of deviation between the current observation data and the normal mode learned by the deep learning model. If the reconstruction error exceeds the set threshold η, it is determined that the server is abnormal. The specific judgment rules are:

[0160]

[0161] Among them, η is the error threshold of anomaly detection, which is set based on the statistical distribution of historical data.

[0162] S2: Establish a request distribution prediction model to predict the number and category of requests at different time steps and in different geographical areas in the future, and establish a request service effect model to quantify the service effect of different servers in processing different categories of requests, including revenue estimation and abnormal possibility assessment.

[0163] Service effect prediction is a key step in the present invention, and its purpose is to estimate the service effect of the request before scheduling the request, including resource consumption, potential benefits and abnormal probability. By combining request characteristics and server attributes, the system can accurately judge the service effect of each request on different servers before scheduling decisions, thereby improving the accuracy of overall scheduling and system resource utilization.

[0164] S21: Establish a request distribution prediction model, specifically:

[0165] In order to improve the accuracy of request prediction, the present invention introduces a spatiotemporal feature prediction model, which combines historical request data and spatiotemporal distribution characteristics to predict future request distribution and scale.

[0166] In order to predict requests more efficiently, we first use the KMeans clustering method to classify historical live broadcast requests, classify heterogeneous requests into multiple categories, and then predict the number of requests of each category respectively. Among them, when using the KMeans clustering method for classification, classification is performed according to request features including content type, request platform and request time period; by clustering analysis of request features, heterogeneous requests are classified into multiple categories, thereby simplifying complex feature prediction tasks and improving prediction efficiency.

[0167] Based on cluster analysis, a model combining spatiotemporal convolutional network (TCN) and graph convolutional network (GCN) is used to predict the number of requests at different times and regions for different categories of requests. The TCN module is used to extract the time series features of requests and predict the number of requests at a certain moment in the future. The GCN module is used to capture the spatial correlation between different regions and predict the transfer and change of requests between regions through the graph convolutional network.

[0168] like Figure 3 As shown, the specific process is:

[0169] Define a graph G=(V,E,A,R) containing multiple geographic regions, where V is a set of spatial nodes representing different regions, E is an edge set between regions, A is an adjacency matrix between nodes, and R is a historical request data matrix. The model predicts the number of requests for each region in the future time step through convolution operations. The formula is as follows:

[0170] H′=σ1(TCN1(R))⊙σ2(TCN2(R)

[0171]

[0172] Wherein, H′ is the feature matrix extracted by the spatiotemporal convolutional network TCN, TCN1(R) and TCN2(R) represent two different temporal convolution operations on the historical request matrix data matrix R, σ1 and σ2 are activation functions for performing nonlinear transformation on the convolution results, and ⊙ represents element-by-element multiplication;

[0173] D is a diagonal matrix, I is the identity matrix, A is the adjacency matrix for adaptive learning, is the normalization of the adjacency matrix A, Θ is the parameter matrix of the graph convolutional network, and H is the processed feature matrix;

[0174] Wd is the weight matrix, b d is the bias vector, σ3 is the activation function, and finally the predicted number of requests at the future time t+1 is obtained.

[0175] Through the above model, the system can predict the number and category distribution of requests at different time steps and in different geographical areas in the future. These prediction results will serve as important inputs for the next step of service effect modeling, providing the estimated demand of different categories of requests in different regions, facilitating further resource scheduling decisions.

[0176] S22: Establish a request service effect model, specifically:

[0177] After completing the request prediction, the present invention further quantifies the service effect of different servers in processing a certain type of request, including revenue estimation and abnormal possibility assessment, through the service effect modeling step, combined with server characteristics and predicted request distribution.

[0178] The present invention uses the Cross-Attention Mechanism to dynamically match request features with server characteristics, thereby accurately predicting the service effect of the request. The key features of the server include:

[0179] Bandwidth capacity: The network transmission capability of the server.

[0180] Geographic location: The physical distance between the server and the user affects network latency.

[0181] Historical status: includes historical load, anomaly detection results, etc., which are used to evaluate the availability of the server.

[0182] The request features and server attributes are matched through the cross-attention mechanism. The formula is as follows:

[0183] Q,V=φ enc1 (i,m,t),φ enc2 (e,B e ,L e )

[0184] α=softmax(Q)

[0185] Among them, Q represents the feature code of the request, V represents the feature code of the server, and φ enc1 and φ enc2 are two different encoding functions, i is the category of the request, m is the geographical location or area of ​​the request, t is the time of the request, e is the server identifier, and B e is the bandwidth capacity of the server, L e The historical status of the server, including historical load and abnormal result detection;

[0186] α is the attention weight, which is used to measure the compatibility between the server and the request. The server priority is dynamically adjusted through the attention mechanism to ensure that the most suitable server is selected for each request.

[0187] like Figure 4 As shown, based on the cross-attention mechanism, a request service effect matrix including a benefit matrix (A) and an anomaly matrix (B) is generated, wherein the benefit matrix is ​​used to quantify the potential benefits of different types of requests on different servers. The server bandwidth, location and load will affect the benefits. The anomaly matrix is ​​used to predict the anomalies that a request may encounter on a specific server, including device anomalies and service anomalies. The specific generation process is as follows:

[0188] A e,m,i =φ dec1 (Q+Dropout(α·V))

[0189] S e,m,i =φ dec2 (Q+Dropout(α·V))

[0190] Among them, A e,m,i represents the revenue generated by the i-th type of request on server e, S e,m,i Indicates the abnormal service status of the current request on server e, φ dec1 and φ dec2 is the decoding function, and Dropout is a technique to prevent overfitting.

[0191] The generated profit matrix and anomaly matrix provide strong data support for scheduling decisions. During the scheduling process, the system will give priority to servers with the highest profit and lowest anomaly probability to provide services for requests, ensuring the efficiency and stability of the system.

[0192] Through service effect modeling, we can evaluate the service effect of each type of request in advance, optimize scheduling decisions, maximize the utilization of platform resources, and reduce the risk of service interruptions and anomalies.

[0193] S3: Based on the fault detection model and the request service effect model, a real-time request scheduling framework is constructed to schedule the real-time user live broadcast service requests.

[0194] After establishing the fault detection model and service effect prediction model, we use the following optimal scheduling algorithm to schedule the real-time user live broadcast service requests to maximize the resource utilization of the edge cloud platform and avoid invalid scheduling.

[0195] First, for the live broadcast requests that come in real time, we use the clustering model to obtain the category of each request and record the geographic location of the request. represents the number of requests of the i-th category from region m;

[0196] Then we define the resource utilization maximization scheduling problem. Assuming there are E edge nodes and M locations, when facing real-time requests, the request scheduling strategy finally obtained in each round of scheduling is defined as in represents the number of requests of type i served by server e at location m. Under this premise, the pre-scheduling problem of each scheduling cycle is:

[0197]

[0198] Formula (1) is the objective function, which aims to maximize the resource utilization of the entire edge cloud platform. Formula (2) ensures that the number of requests allocated is non-negative. Formula (3) ensures that all requests from any category i in any region m are All are served, and formula (4) ensures that the resource utilization of each server e does not exceed its optimal utilization threshold, avoiding server overload or low resource utilization.

[0199] E represents the number of edge nodes, M represents the number of locations, and N represents the number of requested categories. A represents the number of requests of type i served by server e at location m, e,m,i represents the revenue generated by the i-th type of request on server e, S e,m,i Indicates the abnormal service status of the current request on server e, B e represents a certain resource quantity of server e, D e is a weight coefficient associated with server e;

[0200] For all requests of any category i from any region m, U opt is the optimal utilization threshold.

[0201] Since x in formula (1) represents a non-negative integer, and the objective function and all constraints are linear, the problem constitutes an integer linear programming (ILP) problem. Although in theory there are analytical solutions to ILP problems, solving ILP is NP-Hard (i.e., the complexity is extremely high and there is no possibility of solving it in a finite time). Considering the huge scale of CCP's servers and requests, this leads to a huge solution space, and most ILP solutions cannot be solved in time.

[0202] To solve the above problems, we proposed a problem simplification strategy, ignoring the geographic location dimension of the request and Dimensionality reduction represents the number of requests of type i from all regions assigned to server e, while using represents the average revenue of server e when processing the i-th type of request. After the above simplification, we can simplify the above formula (1) as follows:

[0203]

[0204] The complexity of the new ILP is reduced by a factor of M = 59 (from 1 minute to 1 second or less). In addition, we relax the integer constraint on x and transform the problem into a linear programming problem. We then use the Branch and Bound (BB) algorithm to solve the problem. After that, we perform two processing steps to get PS:

[0205] A: Round to the nearest integer

[0206] B: Further refine the simplified two-dimensional strategy PS∈R according to the actual geographical distribution of the request. E×N Extension to three-dimensional strategy PS∈R E×M×N , in order to more accurately reflect the request distribution in different regions. The specific expansion formula is:

[0207]

[0208] This extension ensures that the pre-scheduling strategy can accurately reflect the distribution of requests in different regions, while further considering the possibility of service anomalies. After obtaining the PS, the requests of the corresponding categories and regions are assigned to the corresponding servers, thereby maximizing the resource utilization of the edge cloud platform while avoiding scheduling to faulty servers.

[0209] Second embodiment

[0210] like Figure 5 As shown, this embodiment provides a live broadcast request scheduling system for active anomaly detection for executing the live broadcast request scheduling method for active anomaly detection in the first embodiment, characterized by comprising:

[0211] The fault detection model building module is used to design a two-stage fault detection model using rule-based detection and learning-based detection based on the historical load information of all servers under the edge cloud platform collected in advance. First, the rule-based detection is used to quickly exclude servers with serious anomalies and screen out servers with potential anomalies, and then the learning-based detection is used to re-evaluate the servers with potential anomalies;

[0212] The service effect model building module is used to establish a request distribution prediction model to predict the number and type of requests at different time steps and different geographical areas in the future, and to establish a request service effect model to quantify the service effect of different servers processing different types of requests, including revenue estimation and abnormal possibility assessment;

[0213] The request invocation framework construction module is used to construct a real-time request scheduling framework based on the fault detection model and the request service effect model to schedule the real-time user live broadcast service requests.

[0214] A computer-readable storage medium stores a computer code. When the computer code is executed, the above method is executed. A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program. The program can be stored in a computer-readable storage medium. The storage medium can include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0215] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

[0216] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0217] It should be noted that the above embodiments can be freely combined as needed. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention.

Claims

1. A live broadcast request scheduling method with active anomaly detection, characterized in that: The following steps are involved: S1: Based on the historical load information of all servers under the edge cloud platform collected in advance, a two-stage fault detection model is designed, which adopts rule-based detection and learning-based detection. First, the rule-based detection is used to quickly exclude servers with serious anomalies and screen out servers with potential anomalies. Then, the learning-based detection is used to re-evaluate the servers with potential anomalies. S2: Establish a request distribution prediction model to predict the number and type of requests at different time steps and in different geographical areas in the future, and establish a request service effect model to quantify the service effect of different servers in processing different types of requests, including revenue estimation and abnormal possibility assessment; S3: Based on the fault detection model and the request service effect model, a real-time request scheduling framework is constructed to schedule the real-time user live broadcast service requests.

2. The method for scheduling live broadcast requests based on active anomaly detection according to claim 1, characterized in that: In step S1, the rule-based detection is specifically: The historical load information of each server is recorded using a sliding window S, where the monitoring time length is T, and the resource dimension N of the load sequence represents the number of variables, that is, the number of monitored resource types, s t,n is the load information at time point t and resource dimension n, and the sliding window S is defined as: S={s t,n ∣t∈[1,T],n∈[1,N]} In addition, select a healthy load sequence S h As a baseline, the rule-based detection follows the following two conditions: A: Resource Fluctuation Rules For each resource dimension n, calculate the variance fluctuation of the real-time resource usage rate and the healthy baseline. If the fluctuation exceeds the set threshold, it is determined that the current server has an abnormality in the current resource dimension n. The specific formula is: Was(s t,n )-Var(s h,n )>θ v Among them, Var(s t,n ) is the resource variance of the nth resource dimension in the current sliding window S, Var(s h,n ) is the resource variance of the healthy baseline, θ v is the set fluctuation threshold. If the condition is met, it means that the current resource dimension n has abnormal fluctuations; B: Resource dependency rules Calculate the correlation matrix between resource dimensions in the current sliding window S and compare it with the correlation of the healthy baseline. If the correlation difference exceeds the set threshold, it is determined that the current server has resource anomalies. The specific formula is: ||R(S)-R(S h )||>θ R Among them, R(S) is the correlation matrix of resource utilization in the current sliding window S, R(S h ) is the baseline correlation matrix, θ r is the threshold of correlation change. If the formula condition is met, it means that the resource usage pattern of the server is abnormal. Based on resource fluctuation rules and resource correlation rules, quickly evaluate the status of the server: If both conditions are met, the current server is judged to be seriously abnormal and unsuitable for scheduling, and the current server is excluded; If both conditions are not met, the server is judged to be in a healthy state and the current server is allowed to participate in scheduling; If only one of the conditions is met, there is a potential abnormality in the current server, and the learning-based detection step of the next stage is entered for re-evaluation.

3. The method for scheduling live broadcast requests based on active anomaly detection according to claim 2, characterized in that: In step S1, the learning-based detection is specifically: A deep learning model based on the variational gated recurrent unit (VGRU) is used to model and analyze the multi-dimensional resource usage data of the server. The variational gated recurrent unit (VGRU) includes an inference phase and a generation phase: The reasoning stage: In the inference phase, the deep learning model uses information s according to the input resource usage information s at the current time step. t,n and the hidden state h at the previous time step t-1,n Infer the underlying random variable z t,n , the potential random variable is inferred by the following formula: z t,n ∣s t,n ~N(μ z,t,n ,s z,t,n ) Among them, μ z,t,n and σ z,t,n are the mean and variance of the underlying random variable, respectively. and are the nonlinear transformation functions of the mean and variance, φ s is the feature vector extracted by the convolutional network, h t-1,n is the historical information of the previous time step; At the same time, the reasoning stage integrates time series features through the mechanism of gated recurrent unit GRU to capture long-term dependencies in resource usage patterns; The generation phase: In the generation phase, the deep learning model uses the potential random variables z inferred by the inference phase. t,n , reconstruct the original input Used to calculate the reconstruction error to determine whether there is an anomaly. The formula for the generation process is: in, Represented by the potential random variable z t,n Generated reconstructed input, μ s,t,n and σ s,t,n are the mean and variance in the generation process, and is the nonlinear function of the decoder. The generated reconstructed data will be used to calculate the difference with the real input data to determine the abnormal state of the server. t-1,n is the historical information of the previous time step; The deep learning model calculates the actual input data s t,n With the generated reconstructed data The reconstruction error between , the formula of the reconstruction error is: The reconstruction error reflects the degree of deviation between the current observation data and the normal mode learned by the deep learning model. If the reconstruction error exceeds the set threshold η, it is determined that the server is abnormal. The specific judgment rules are: Among them, η is the error threshold of anomaly detection, which is set based on the statistical distribution of historical data.

4. The method for scheduling live broadcast requests based on active anomaly detection according to claim 1, characterized in that: In step S2, the request distribution prediction model is established to predict the number and type of requests at different time steps and different geographical areas in the future, specifically: The KMeans clustering method is used to classify historical live broadcast requests, and the heterogeneous requests are classified into multiple categories, and then the number of requests of each category is predicted respectively, wherein the KMeans clustering method is used to classify according to the request characteristics including content type, request platform and request time period; Based on cluster analysis, a model combining spatiotemporal convolutional network (TCN) and graph convolutional network (GCN) is used to predict the number of requests at different times and regions for different categories of requests. The specific process is as follows: Define a graph G=(V,E,A,R) containing multiple geographic regions, where V is a set of spatial nodes representing different regions, E is an edge set between regions, A is an adjacency matrix between nodes, and R is a historical request data matrix. The model predicts the number of requests for each region in the future time step through convolution operations. The formula is as follows: H′=σ1(TCN1(R))⊙σ2(TCN2(R)) Wherein, H′ is the feature matrix extracted by the spatiotemporal convolutional network TCN, TCN1(R) and TCN2(R) represent two different temporal convolution operations on the historical request matrix data matrix R, σ1 and σ2 are activation functions for performing nonlinear transformation on the convolution results, and ⊙ represents element-by-element multiplication; D is a diagonal matrix, I is the identity matrix, A is the adjacency matrix for adaptive learning, is the normalization of the adjacency matrix A, Θ is the parameter matrix of the graph convolutional network, and H is the processed feature matrix; W d is the weight matrix, b d is the bias vector, σ3 is the activation function, and finally the predicted number of requests at the future time t+1 is obtained.

5. The method for scheduling live broadcast requests based on active anomaly detection according to claim 1, characterized in that: In step S2, a request service effect model is established to quantify the service effects of different servers processing different types of requests, including revenue estimation and abnormal possibility assessment, specifically: The request features and server attributes are matched through the cross-attention mechanism. The formula is as follows: Q,V=φ enc1 (i,m,t),φ enc2 (e,B e ,L e ) α=softmax(Q) Among them, Q represents the feature code of the request, V represents the feature code of the server, and φ enc1 and φ enc2 are two different encoding functions, i is the category of the request, m is the geographical location or area of ​​the request, t is the time of the request, e is the server identifier, and B e is the bandwidth capacity of the server, L e The historical status of the server, including historical load and abnormal result detection; α is the attention weight, which is used to measure the compatibility between the server and the request. The server priority is dynamically adjusted through the attention mechanism to ensure that the most suitable server is selected for each request. Based on the cross-attention mechanism, a request service effect matrix including a benefit matrix and an anomaly matrix is ​​generated. The benefit matrix is ​​used to quantify the potential benefits of different types of requests on different servers. The server bandwidth, location and load will affect the benefits. The anomaly matrix is ​​used to predict the anomalies that a request may encounter on a specific server, including device anomalies and service anomalies. The specific generation process is as follows: A e,m,i =φ dec1 (Q+Dropout(α·V)) S e,m,i =φ dec2 (Q+Dropout(α·V)) Among them, A e,m,i represents the revenue generated by the i-th type of request on server e, S e,m,i Indicates the abnormal service status of the current request on server e, φ dec1 and φ dec2 is the decoding function, and Dropout is a technique to prevent overfitting.

6. The method for scheduling live broadcast requests based on active anomaly detection according to claim 5, characterized in that: In step S3, based on the fault detection model and the request service effect model, a real-time request scheduling framework is constructed to schedule the real-time user live service requests, specifically: First, for the live broadcast requests that come in real time, we use the clustering model to obtain the category of each request and record the geographic location of the request. represents the number of requests of the i-th category from region m; Define the resource utilization maximization scheduling problem. Assume that there are E edge nodes and M locations. When facing real-time requests, the request scheduling strategy obtained in each round of scheduling is defined as in represents the number of requests of type i served by server e at location m. Under this premise, the pre-scheduling problem of each scheduling cycle is: Formula (1) is the objective function, which aims to maximize the resource utilization of the entire edge cloud platform. Formula (2) ensures that the number of requests allocated is non-negative. Formula (3) ensures that all requests from any category i in any region m are All are served, and formula (4) ensures that the resource utilization of each server e does not exceed its optimal utilization threshold, avoiding server overload or low resource utilization; E represents the number of edge nodes, M represents the number of locations, and N represents the number of requested categories. A represents the number of requests of type i served by server e at location m, e,m,i represents the revenue generated by the i-th type of request on server e, S e,m,i Indicates the abnormal service status of the current request on server e, B e represents a certain related resource amount of server e, D e is a weight coefficient associated with server e; For all requests of any category i from any region m, U opt is the optimal utilization threshold.

7. The method for scheduling live broadcast requests based on active anomaly detection according to claim 6, characterized in that: Step S3 also includes: To simplify the problem, we first ignore the geographic location dimension of the request and Dimensionality reduction represents the number of class i requests from all regions assigned to server e, while using represents the average revenue of server e when processing the i-th type of request. After the above simplification, we can simplify the above formula (1) as follows: Getting the intermediate results After that, we perform two processing steps to get PS: A: Round to the nearest integer B: Further refine the simplified two-dimensional strategy PS∈R according to the actual geographical distribution of the request. E×N Extension to three-dimensional strategy PS∈R E×M×N , in order to more accurately reflect the request distribution in different regions. The specific expansion formula is: After obtaining the PS, requests of corresponding categories and regions are assigned to corresponding servers, thereby maximizing the resource utilization of the edge cloud platform while avoiding scheduling to faulty servers.

8. A live broadcast request scheduling system for active anomaly detection for executing the live broadcast request scheduling method for active anomaly detection according to any one of claims 1 to 7, characterized in that: include: The fault detection model building module is used to design a two-stage fault detection model using rule-based detection and learning-based detection based on the historical load information of all servers under the edge cloud platform collected in advance. First, the rule-based detection is used to quickly exclude servers with serious anomalies and screen out servers with potential anomalies, and then the learning-based detection is used to re-evaluate the servers with potential anomalies; The service effect model building module is used to establish a request distribution prediction model to predict the number and type of requests at different time steps and different geographical areas in the future, and to establish a request service effect model to quantify the service effect of different servers processing different types of requests, including revenue estimation and abnormal possibility assessment; The request invocation framework construction module is used to construct a real-time request scheduling framework based on the fault detection model and the request service effect model to schedule the real-time user live broadcast service requests.

9. A computer device comprising a memory and one or more processors, wherein the memory stores computer codes, and when the computer codes are executed by the one or more processors, the one or more processors execute the method according to any one of claims 1 to 7. 10 . A computer-readable storage medium storing a computer code. When the computer code is executed, the method according to claim 1 is executed.

Images