CDN node service quality prediction method, service quality early warning method and management system

Abstract

The invention relates to the technical field of CDN, and provides a CDN node service quality prediction method, which comprises the steps of generating a static embedding vector based on static configuration information of a node; calculating a time sequence of a target performance index of the node in a historical time period based on a time sequence dynamic model to obtain a time sequence hidden state vector; and inputting the static embedded vector and the time sequence hidden state vector into a service quality prediction model, and calculating to obtain a predicted value of the target performance index in a future preset time period. Therefore, accurate prediction of the CDN node service quality can be realized.

Description

Technical Field

[0001] This application relates to the field of CDN technology, specifically to CDN node service quality prediction methods, service quality early warning methods, and management systems. Background Technology

[0002] With the rapid popularization of mobile internet and broadband networks, video-on-demand services have gradually become a major component of internet traffic. Industry statistics show that video services now account for over 80% of global internet traffic, and users are placing increasingly stringent demands on user experience metrics such as smooth video playback, first-frame loading latency, and buffering rates.

[0003] To support the video consumption needs of massive users and continuously improve the quality of experience (QoE), Content Delivery Networks (CDNs) are widely deployed at edge network nodes around the world, effectively reducing user access latency and improving content transmission efficiency by distributing content from the nearest location.

[0004] As CDN networks continue to expand, their operation and maintenance management becomes significantly more challenging. On one hand, CDN edge nodes typically number in the thousands or even tens of thousands, distributed across different geographical regions and connected to different carrier networks, with significant differences in hardware configuration, bandwidth capacity, and network environment among the nodes. On the other hand, video-on-demand service load exhibits distinct periodicity and burstiness, with significant differences in the operating status and performance of different nodes at different times.

[0005] Against this backdrop, how to achieve refined monitoring of CDN service quality and accurately predict the service quality of each node before anomalies occur, thereby identifying potential risks in advance, has become a key technical issue in ensuring the stable operation of businesses. Summary of the Invention

[0006] To achieve accurate prediction of CDN node service quality, this application provides a CDN node service quality prediction method. The method is applied to a CDN management system and includes: generating a static embedding vector based on the node's static configuration information, wherein the static configuration information includes physical attributes, including one or more of hardware configuration, network attributes, and geographical attributes; calculating the time series of the node's target performance indicators within a historical period based on a time-series dynamic model to obtain a time-series hidden state vector; inputting the static embedding vector and the time-series hidden state vector into a service quality prediction model to calculate the predicted value of the target performance indicator within a future preset period; wherein the calculation process of the predicted value includes constructing a Query using the static embedding vector, constructing a Key and a Value using the time-series hidden state vector, calculating an attention score through the dot product of the Query and Key, and normalizing it to obtain an attention weight, applying the attention weight to the Value to generate an attention output vector, and concatenating the static embedding vector, the time-series hidden state vector, and the attention output vector to construct a unified spatiotemporal fusion feature vector.

[0007] Based on the above method, the spatiotemporal fusion modeling mechanism implemented in this application can deeply fuse the static attribute features of CDN nodes with dynamic temporal features. For each CDN node, static information such as hardware configuration (e.g., CPU specifications, memory capacity, disk type), network attributes (e.g., access operator, bandwidth capacity, network layer), and geographical attributes (e.g., region, time zone) is first extracted to form a static embedded representation of the node. Simultaneously, the time series of node performance indicators is encoded to extract temporal dynamic features. The static embedding and temporal features are fused through an attention fusion layer, enabling the model to perceive the structural differences between different nodes. This allows the model to adaptively adjust the modeling strategy for different nodes during prediction, significantly improving the model's overall network generalization ability and thus enhancing the accuracy of prediction results.

[0008] In one implementation, the method for generating static embedding vectors based on node static configuration information includes: classifying the static configuration information of the nodes; for discrete attributes, determining attribute values ​​and querying an embedding table based on the attribute values ​​to obtain discrete attribute embedding vectors; for continuous attributes, standardizing the numerical values ​​to obtain continuous attribute vectors; concatenating the discrete attribute embedding vectors and the continuous attribute vectors, and performing regularization processing to obtain the static embedding vector of the node.

[0009] In one implementation, the method for calculating the time series based on the time-series dynamic model to obtain the time-series hidden state vector includes: performing sliding window sampling based on the time series to obtain multiple sub-series samples; initializing the hidden state; sequentially inputting the data of each time step of the sub-series samples into the GRU, so that the GRU updates the current hidden state according to the hidden state of the current input and the previous input; and when all time step data has been processed, extracting the hidden state of the last time step as the time-series hidden state vector.

[0010] In one implementation, the method for calculating the predicted value further includes: inputting the spatiotemporal fusion feature vector into a two-layer fully connected neural network for feature transformation and dimensionality reduction to obtain a high-level feature vector; mapping the high-level feature vector to the target prediction space through a linear output layer to generate a sequence of predicted service quality indicators for a future preset step size.

[0011] In one implementation, the static configuration information also includes domain name business attributes, which include one or more of the following: business type, traffic level, acceleration region, and customer level.

[0012] Based on this, quality prediction at the domain level can be achieved, thereby adapting to different monitoring needs.

[0013] Furthermore, this application embodiment also provides a CDN service quality early warning method, which is applied to a CDN management system and includes: collecting and preprocessing static configuration information of a target node and the time series of target performance indicators in a historical period; calculating the predicted value of the target performance indicator of the target node in a future preset period according to the above method; obtaining the dynamic alarm threshold corresponding to the target node; and comparing the predicted value with the dynamic alarm threshold to determine whether an early warning is triggered.

[0014] In one implementation, the method for calculating the dynamic alarm threshold includes: loading the historical operating baseline statistics of the target node; and dynamically calculating the dynamic alarm threshold of the target node based on the historical operating baseline statistics.

[0015] In one implementation, the historical operating baseline statistics are updated based on the EWMA algorithm.

[0016] Furthermore, this application embodiment also provides a CDN management system, wherein the above-mentioned prediction method predicts the service quality of nodes or the above-mentioned early warning method realizes CDN service quality early warning management. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the CDN node service quality prediction method provided in the embodiments of this application; Figure 2 A flowchart illustrating the method for generating static embedding vectors based on node static configuration information in this application embodiment; Figure 3 This is a flowchart of the method for calculating the temporal hidden state vector in the embodiments of this application; Figure 4 This is a flowchart illustrating the method for calculating the predicted value in an embodiment of this application; Figure 5 This is a flowchart of the CDN service quality early warning method provided in the embodiments of this application; Figure 6 This is a flowchart illustrating the method for calculating the dynamic alarm threshold in this application embodiment. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by this application in order to achieve the intended purpose of the invention, the following detailed description of the specific implementation methods, structures, features and effects of this application is provided in conjunction with the accompanying drawings and preferred embodiments.

[0020] Please refer to Figure 1 The CDN node service quality prediction method provided in this application embodiment is used to accurately predict the service quality of each node server in the CDN network. It can be applied to the CDN management system and specifically includes the following steps.

[0021] S101, Generate static embedding vectors based on the static configuration information of nodes.

[0022] In one implementation, the static configuration information includes physical attributes, which include one or more of the following: hardware configuration (such as CPU specifications, memory capacity, disk type), network attributes (such as access operator, bandwidth capacity, network level), and geographical attributes (such as location region, time zone).

[0023] In another implementation, the static configuration information may further include domain name business attributes, which may include one or more of the following: business type, traffic level, acceleration region, and customer level. By incorporating domain name business attributes into the static configuration information, node-level and domain-level service quality predictions can be achieved, thereby facilitating focused monitoring of specific accelerated domain names.

[0024] Please refer to Figure 2 The method for generating static embedding vectors based on the static configuration information of nodes includes the following steps.

[0025] S201 categorizes the static configuration information of nodes.

[0026] Based on the nature of each attribute in the static configuration information, they can be divided into two categories: discrete attributes and continuous attributes. In one example, categorical attributes can be identified as discrete attributes, and numerical attributes as continuous attributes.

[0027] For discrete attributes, proceed to step S202; for continuous attributes, proceed to step S203.

[0028] S202, determine the attribute values, and query the embedding table based on the attribute values ​​to obtain discrete attribute embedding vectors.

[0029] In implementation, corresponding attribute values ​​can be pre-set for each discrete attribute, and an independent embedding table can be constructed.

[0030] In one example, the size of the value space of a discrete attribute is set to |V|, and the embedding dimension is d. e Then the embedding table of this attribute is a learnable parameter matrix E∈R |V|×de .

[0031] For a discrete attribute value v∈V, its embedding vector is e. v =E v That is, the v-th row of the embedding table. Embedding vectors of multiple discrete attributes can be concatenated as: e discrete =[e v1 ;e v2 ;⋯;e vk ] Here, the semicolon indicates a vector concatenation operation, and k is the number of discrete attributes.

[0032] For example, the operator attribute of node "node003" is "China Telecom", the data center type is "edge data center", and the embedded tables maintained by the system can be found in Tables 1 and 2 for examples: Table 1

[0033] Table 2

[0034] Therefore, the discrete attribute embedding vector of this node is the concatenation of two embedding vectors: e discrete = [0.23,-0.15,0.42,0.08,-0.34,0.61].

[0035] S203, standardize the numerical values ​​to obtain a continuous attribute vector.

[0036] In implementation, Z-Score standardization can be used for continuous attributes: x std,i =(x i -μ i ) / σ i Where, μ i and σ i Let be the mean and standard deviation of the i-th attribute in the training set. Here, the training set is the dataset corresponding to that attribute.

[0037] Multiple continuous attributes form a continuous attribute vector: c=[x std,1 ,x std,2 ,⋯,x std,m ] Where m is the number of continuous attributes.

[0038] For example, the statistical parameters of the number of CPU cores in the training set are: mean μ = 32 cores, standard deviation σ = 16 cores, and the number of CPU cores in node "node003" is 64 cores, after standardization: x std =(64-32) / 16=2.0 This value indicates that the number of CPU cores on this node is 2 standard deviations higher than the average, classifying it as a high-configuration node.

[0039] S204 concatenates the discrete attribute embedding vector and the continuous attribute vector, and after regularization, obtains the static embedding vector of the node.

[0040] In implementation, discrete attribute embedding vectors can be concatenated with continuous attribute vectors to form the static embedding vector of a node: s=[e discrete ;c], where the semicolon is for concatenation.

[0041] To prevent overfitting, Dropout regularization is applied after fusion. Dropout regularization is a commonly used technique in neural network training to prevent model overfitting and thus improve its generalization ability.

[0042] S102, based on the time series dynamic model, calculates the time series of the target performance indicators of the node in the historical period to obtain the time series hidden state vector.

[0043] The historical time period can be the time period corresponding to a preset duration from the current moment, such as the previous 10 minutes. The target performance metric refers to the performance metrics that the system needs to monitor, including but not limited to response time, first frame latency, error rate, bandwidth utilization, stuttering rate, number of connections, CPU utilization, and memory utilization. The time series of the target performance metrics within the historical time period consists of the data corresponding to the target performance metrics collected within that historical time period.

[0044] For example, the stuttering rate (in %) of the domain "vod.example.com" over the first 10 minutes was collected on node "node001", as shown in Table 3. Table 3

[0045] Please refer to Figure 3 In this embodiment, the method for calculating the time series based on the time series dynamic model to obtain the time series hidden state vector includes the following steps.

[0046] S301, based on time series, performs sliding window sampling to obtain multiple subsequence samples.

[0047] In implementation, the time series data can be normalized first, and then sampled according to a preset sliding window to obtain multiple subsequence samples. Normalization ensures the consistency and comparability of the input data, improving the model's convergence speed and performance.

[0048] By using the sliding window sampling technique, long time series can be divided into multiple shorter subsequence samples. Each subsequence sample contains data at a certain time step. In this way, continuous time series data can be transformed into a sample form suitable for model processing, while preserving the local features and temporal relationships of the time series.

[0049] For example, by setting the window length w=5 and the prediction steps k=2, the stuttering rate time series in the previous example can be constructed into the subsequence samples in Table 4: Table 4

[0050] S302, Initialize hidden state.

[0051] In implementation, hidden states are used to store and pass on historical information from the sequence. An initial value is initialized for the hidden states to ensure the model can process time series data progressively from the initial state.

[0052] S303, input the data of each time step of the subsequence sample into the GRU in sequence, so that the GRU updates the current hidden state according to the hidden state of the current input and the previous input.

[0053] In practice, the time series dynamic model can be implemented based on the GRU (Gated Recurrent Unit) neural network, which can effectively capture long-term dependencies in time series through gating mechanisms.

[0054] GRU controls the flow of information through reset and update gates. The reset gate determines how much information from the previous time step's hidden state needs to be forgotten, while the update gate determines how much new information from the current input needs to be added to the hidden state. During the processing of data at each time step, GRU calculates and updates the current hidden state based on the current input and the hidden state of the previous input using a gating mechanism.

[0055] Taking sample 1 above as an example, the calculation process of GRU is shown in Table 5: Table 5

[0056] It is worth noting that when there are multiple target performance metrics, the input value of the subsequence samples can be set as a multi-dimensional vector containing each target performance metric. For example, when monitoring both response time and stuttering rate simultaneously, the input for each time step is a two-dimensional vector containing the corresponding metric value. In this way, GRU can simultaneously capture the changing trends of the two metrics and their synergistic relationship; for example, a longer response time is often accompanied by an increase in stuttering rate.

[0057] S304. After all time step data has been processed, extract the hidden state of the last time step as the time-series hidden state vector.

[0058] The final hidden state contains the time-series information of all subsequence samples and can be used as a feature representation of that sample.

[0059] In one example, the feature vectors corresponding to each subsequence sample can be calculated based on the above method, and then aggregated, such as by averaging or summing, and the aggregated result can be used as the time series hidden state vector.

[0060] In another example, GRU computation can be performed on each subsequence sample in sequence, and the computation result of the last time step can be used as the time-series hidden state vector.

[0061] In a practical application example, the GRU gated loop unit controls the flow of information by updating and resetting gates, for the input x at time step t. t and the hidden state h at the previous momentt-1 The calculation process of GRU is as follows: Update Gate: z t =σ(W z x t +U z h t-1 +b z ) Reset Gate: r t =σ(W r x t +U r h t-1 +b r ) Candidate hidden states: Current hidden state: Where σ is the Sigmoid activation function, ⊙ is element-wise multiplication, and W, U, and b are learnable parameters.

[0062] To improve the model's expressive power, multiple layers of GRU can be stacked. Let the first... l The input of the layer is h t (l -1) The output is h t (l ) Then we have:

[0063] The final time step hidden state h of the last layer T (L) This is the output time-series feature vector.

[0064] For multi-indicator joint modeling, the input at each time step is a vector composed of multiple indicator values.

[0065] Where n is the number of indicators.

[0066] Model training can be performed using supervised learning. For an original time series of length T, a sliding window method is used to construct input-target sample pairs (X, Y). The input window length is set to w, and the prediction step size is k. For time t, the input sequence is: X t =[x t-w+1 ,x t-w+2 ,…,x t ] The corresponding target label sequence Y t =[y t+1 ,y t+2 ,…,y t+k], where y is the key performance indicator to be predicted (such as response time). By sliding over the original sequence with a preset step size, a large number of sample pairs can be generated for model training.

[0067] Based on this, this embodiment adopts a multi-layer time series modeling structure based on GRU to achieve joint modeling of multiple indicator time series. Through update gate and reset gate mechanisms, the model can adaptively balance historical information and current input in the time dimension, effectively depicting the dynamic changes of indicators. For multi-indicator input scenarios, each indicator is combined into a vector at the same time step as the model input, and a sliding window method is used to construct supervised learning sample pairs. This allows the model to cyclically encode the input sequence step by step, and the hidden state of the last layer at the final time step is used as the time series hidden state vector, providing high-quality dynamic features for subsequent multi-step prediction and service quality early warning.

[0068] It is understood that the static configuration information in this embodiment is all collected based on the corresponding nodes and preprocessed. The data collection and preprocessing methods will be described in detail below and will not be repeated here.

[0069] S103, input the static embedding vector and the time-series hidden state vector into the service quality prediction model to calculate the predicted value of the target performance index in the future preset time period.

[0070] Among them, the service quality prediction model is pre-trained and used to predict the performance of the target performance indicators of nodes in the future within a preset period of time, thereby providing a decision-making basis for node monitoring and operation and maintenance.

[0071] Please refer to Figure 4 The method for calculating the predicted value specifically includes the following steps.

[0072] S401 receives the input of the static embedding vector and the time-series hidden state vector.

[0073] The input layer of the service quality prediction model consists of a static embedding vector and a time-series hidden state vector.

[0074] S402, calculates the attention weights of each time step data in the temporal hidden state vector based on the attention mechanism.

[0075] Static attributes and temporal dynamic features of nodes belong to different feature spaces, and their influence on prediction results differs significantly. Static attributes characterize the structural features and configuration conditions of nodes and do not change over time, while temporal features reflect the evolution of indicators over time, exhibiting significant dynamism and uncertainty. Due to differences in resource allocation and operating environments among different nodes, their focus on historical temporal information during prediction should also differ. Simply concatenating static and temporal features is insufficient to characterize the dependencies and modulating effects between them. Therefore, introducing a fusion mechanism that adaptively adjusts the degree of attention to temporal features based on node attributes can further achieve effective modeling and joint representation of heterogeneous features.

[0076] In implementation, to characterize the influence of the static structural features of different nodes on time series prediction, a scaled dot product attention mechanism can be used. A query is constructed using a static embedding vector, and key and value are constructed using temporal hidden state vectors. Then, the attention score is calculated by the dot product of the query and key, and the attention weight is obtained after Softmax normalization.

[0077] Specifically, firstly, by applying linear transformations to the static embedding vector and the temporal dynamic latent state vector respectively, the original features are mapped to a unified attention representation space, generating corresponding Query, Key, and Value vectors. The Query characterizes the static attributes' need for temporal information, the Key represents the matchability of temporal features, and the Value carries the temporal feature content that actually participates in the fusion.

[0078] Then, by calculating the dot product between Query and Key, the correlation between static attributes and temporal features is measured, and the dot product result is scaled to alleviate the training instability caused by excessively large values ​​in high-dimensional space, thereby obtaining the attention score matrix.

[0079] Finally, the attention scores are normalized using the Softmax function and converted into attention weights that satisfy the probability distribution characteristics. These weights represent the importance of each temporal feature component to the prediction task under the current node attribute conditions.

[0080] Based on this, the adaptive modulation of static structural features on temporal dynamic features is realized, enabling the model to dynamically adjust its focus on historical temporal information according to the differences in the static attributes of nodes, thereby improving the accuracy of prediction results.

[0081] S403: Based on the attention weights, the temporal hidden state vectors are weighted and summed to generate the attention output vector.

[0082] In practice, the attention weights calculated in the above steps can be applied to Value to generate an attention output vector.

[0083] Specifically, the normalized attention weights can be used to perform a weighted summation of the Value vector to generate an attention output vector. This vector comprehensively reflects the selective attention of static attributes to temporal dynamic features and serves as an important input for subsequent spatiotemporal feature fusion and predictive modeling.

[0084] S404 concatenates the static embedding vector, the temporal hidden state vector, and the attention output vector to construct a spatiotemporal fusion feature vector.

[0085] In practice, the static embedding vector, the temporal hidden state vector, and the attention output vector can be concatenated in a predetermined order to construct a spatiotemporal fusion feature vector. This feature vector contains the static structural information of the node, the historical temporal dynamic information, and the interaction between the two, which can provide a comprehensive feature representation for subsequent prediction.

[0086] S405 inputs the spatiotemporal fusion feature vectors sequentially into two fully connected neural networks for feature transformation and dimensionality reduction to obtain high-level feature vectors.

[0087] In implementation, the first fully connected layer can use the ReLU activation function to introduce nonlinear expressive power and suppress the risk of model overfitting through the Dropout regularization mechanism; the second fully connected layer further extracts higher-order feature representations to enhance the model's ability to model complex nonlinear relationships. The dimensions of the two hidden layers can be set to 256 and 128 respectively to achieve a balance between expressive power and computational efficiency.

[0088] S406 maps the high-level feature vectors to the target prediction space through a linear output layer, generating a sequence of predicted service quality index values ​​for a preset step size in the future.

[0089] In implementation, the high-level feature vectors after multi-layer feature transformation are mapped to the target prediction space through a linear output layer to generate a sequence of predicted service quality indicators for a preset number of time steps in the future. In one example, the preset step size is 5 minutes.

[0090] The multi-step prediction mechanism enables the model to characterize the changing trends of service quality indicators in advance, providing the system with sufficient time lead time.

[0091] The above method steps will be explained in detail below with specific application examples.

[0092] Suppose we need to predict the quality of service (QoS) of the domain name live.example.com on node003. The static attributes of this node are encoded to obtain a static embedding vector s∈R.6 .

[0093] Simultaneously, based on the most recent time window, the temporal hidden state vector h of three consecutive time steps is output. t-2 , h t-1 , h t ∈R 4 .

[0094] Specific examples are as follows: s= [0.23, -0.15, 0.42, 2.00, 0.45, 1.20] h t-2 = [0.45, 0.28, -0.12, 0.33] h t-1 = [0.39, 0.31, -0.08, 0.29] h t = [0.41, 0.26, -0.10, 0.30] Step 1: Construct the Query, Key, and Value vectors.

[0095] By projecting the static embedding vector and the temporal hidden state vector onto the attention space using a linear mapping, we obtain: Q=W q s K=W k [h t-2 , h t-1 , h t ] V=W v [h t-2 , h t-1 , h t ] Among them, W q W k W v This is a learnable parameter matrix, where Q, K, and V correspond to Query, Key, and Value, respectively. For ease of illustration, specific projection values ​​are omitted here.

[0096] Step 2: Calculate the scaling dot product attention weights.

[0097] The degree of matching between static features and time-series features at each time step is calculated using the dot product:

[0098] Assuming the calculated attention weights are α=[0.72, 0.18, 0.10], the results show that, under the current node attribute conditions, the model pays more attention to the temporal change patterns of earlier time steps.

[0099] Step 3: Calculate the attention output vector.

[0100] We obtain the attention output by weighting the Value vector using attention weights: a=0.72v t-2 +0.18v t-1 +0.10v t Suppose the attention output vector is a=[0.32, 0.20, -0.09, 0.24], which represents the key temporal dynamic features obtained by aggregation under static attribute modulation.

[0101] Step 4: Spatiotemporal feature fusion and prediction input construction.

[0102] The static embedding vector, the temporal hidden state vector at the current time step, and the attention output vector are concatenated to form the final spatiotemporal fusion feature vector: f=[s; h t ; a]∈R 14 This 14-dimensional feature vector simultaneously contains node static attribute information, recent temporal dynamic information, and the interaction features between the two.

[0103] The spatiotemporal fusion feature vector f is input into the subsequent multilayer perceptron prediction network. After nonlinear mapping, the predicted value sequence of service quality indicators for the next 5 minutes is output. The prediction results can be used to identify potential service quality degradation trends in advance.

[0104] Therefore, the CDN node service quality prediction method provided in this application, by combining the static characteristics and dynamic temporal characteristics of nodes to predict the quality of target performance indicators, can improve the accuracy of prediction. Furthermore, based on the spatiotemporal fusion mechanism, by deeply fusing the static attributes of nodes with the temporal dynamic characteristics, it effectively solves the problem of insufficient generalization ability of traditional simple temporal models.

[0105] Furthermore, static attribute encoding enables the model to perceive structural differences between different nodes, and the attention fusion mechanism enables the interpretation of temporal features to be adapted to the attribute characteristics of nodes. This design allows a single model to be uniformly deployed across the entire network without the need to train the model separately for each node or each type of node, which greatly reduces the training and maintenance costs of the model.

[0106] Meanwhile, the fusion model can share the learned general knowledge among different nodes, and has good generalization ability for newly added nodes or nodes with configuration changes, and can be put into use without retraining.

[0107] This application also provides a CDN service quality early warning method, applied to a CDN management system. Please refer to [link / reference]. Figure 5 The method specifically includes the following steps.

[0108] S501 collects and preprocesses the static configuration information of the target node and the time series of the target performance indicators over historical periods.

[0109] In implementation, the original information of the target node can be obtained through methods such as API retrieval, message queue subscription, and log file collection. For example, the Host field in the request can be extracted as the accelerated domain name identifier by parsing the access log of the CDN node, thereby realizing data collection at the domain level.

[0110] It is understandable that the raw information collected may contain anomalies or be of poor quality, thus requiring preprocessing. Preprocessing methods include data cleaning, missing value imputation, time alignment, and normalization. The specific preprocessing method should be selected according to the different data types, and this application is not limited to this.

[0111] After preprocessing, the raw information yields static configuration information and time series.

[0112] S502, based on static configuration attributes and time series calculations, obtains the predicted values ​​of the target performance indicators of the target node within a preset future time period.

[0113] The predicted values ​​can be calculated using the node service quality prediction method provided in the above embodiments, and will not be repeated here.

[0114] S503, obtain the dynamic alarm threshold corresponding to the target performance index of the target node.

[0115] In one implementation, the method for calculating the dynamic alarm threshold specifically includes the following steps.

[0116] S601, load the historical operating baseline statistics of the target node.

[0117] Among them, the historical operating baseline statistics include the mean and variance of the exponentially weighted moving average of each target performance indicator.

[0118] Since different CDN nodes have different hardware configurations and service loads, their corresponding performance baselines are also different. In implementation, the historical operating baseline statistics of each node can be adaptively calculated and statistically analyzed based on the EWMA algorithm, so as to realize the continuous tracking of the dynamic performance baseline of each node, and then realize the automatic adjustment of alarm thresholds based on historical data.

[0119] Specifically, the EWMA algorithm is used to continuously update the historical baseline statistics of each node's target performance indicators, including mean updates and variance updates based on the following formulas: μ t =αx t +(1-α)⋅μ t-1 σ t 2 =α(x t -μ t ) 2 +(1-α)⋅σ t-1 2 Where α is the smoothing coefficient, used to control the decay rate of historical information; x t μ is the observed value at time t, i.e., the currently collected index value; t Let be the baseline mean at time t, and be the mean updated based on the EWMA algorithm; σ t is the baseline standard deviation at time t, and is the standard deviation updated based on the EWMA algorithm.

[0120] S602 dynamically calculates the dynamic alarm threshold of the target performance index of the target node based on historical operating baseline statistical information.

[0121] In practice, dynamic alarm thresholds can be calculated based on confidence levels.

[0122] Specifically, the dynamic alarm threshold T can be calculated based on the following formula. upper and lower limit T lower : T upper =μ+z γ σ T lower =μ-z γ σ Among them, z γ This is the z-score corresponding to the confidence level, such as z=2.576 for 99% confidence.

[0123] S504 compares the predicted value with the dynamic alarm threshold to determine whether an alert is triggered.

[0124] During implementation, threshold comparisons can be performed on each time step in the prediction sequence. Based on this, multiple alarm strategies can be supported: single-point triggering (triggered when any prediction point exceeds the threshold), continuous triggering (triggered only when n consecutive points exceed the threshold), and trend triggering (triggered when the overall prediction sequence shows a deteriorating trend). Time points that meet the alarm conditions are marked as anomalies, and information such as prediction values, thresholds, and deviations are recorded to help technicians quickly grasp the abnormal situations at nodes and thus handle them in a timely and accurate manner.

[0125] Understandably, based on the above methods, it is also possible to monitor anomalies in target domains. In one example, the generated alert information can be routed to the corresponding operations and maintenance recipients according to the alert granularity. Node-level alerts are pushed to the infrastructure operations and maintenance team, and domain-level alerts are pushed to the business operations and maintenance team. This allows the corresponding professional teams to quickly verify and respond to relevant anomalies in advance, thereby preventing the actual occurrence of anomalies.

[0126] The above method will be further explained below with specific application examples.

[0127] Assuming the system uses the historical average stutter rate of 1.5% and its fluctuation range as the operating baseline, when a new observation value of 2.0% is collected, the EWMA algorithm is used to smooth the baseline, so that the new data has a limited impact on the baseline, resulting in a new mean of 1.6% and a standard deviation of approximately 0.32%.

[0128] Based on the updated baseline statistics, dynamic alarm thresholds were calculated, with the Warning threshold at approximately 2.43% and the Critical threshold at approximately 2.66%.

[0129] If the prediction lag rate reaches 2.8%, exceeding the Critical threshold, a serious performance risk is identified, triggering a Critical-level warning. This mechanism can adapt to differences in node performance while enabling early identification of abnormal risks.

[0130] Furthermore, if the current prediction does not trigger an alarm, the node baseline statistics can be updated using the actual observations of the current period. The EWMA algorithm is used to incrementally update the mean and variance, enabling the baseline to adapt to the slow drift of the business load and thus adapt to changes in the node's state.

[0131] Based on the above technical solution, a deep spatiotemporal network model is used to achieve multi-step prediction of service quality indicators, enabling the system to have proactive early warning capabilities. Unlike traditional methods that can only passively alert after an anomaly has occurred, this application embodiment can identify potential degradation trends while service quality is still at a normal level, issuing early warning signals several minutes in advance. This "early warning" capability provides the operation and maintenance team with a valuable time window, enabling them to take preventative measures, such as traffic scheduling, resource expansion, and failover, before the problem actually affects the user experience. This elevates the early warning timeliness from "post-event response" to "pre-event intervention," significantly enhancing the initiative and controllability of CDN operation and maintenance.

[0132] Furthermore, by introducing an adaptive threshold warning mechanism, personalized alarm thresholds can be established for each CDN node, effectively solving the adaptation problem of traditional uniform threshold methods in heterogeneous network environments. The adaptive threshold is dynamically calculated based on the node's historical operational data, accurately reflecting the performance baseline differences between different nodes caused by factors such as hardware configuration, network environment, and service characteristics, thereby improving the accuracy of warning judgments.

[0133] The above description is merely a preferred embodiment of this application and is not intended to limit this application in any way. Although this application has been disclosed above with reference to preferred embodiments, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the scope of the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. A CDN node service quality prediction method, characterized in that, The method is applied to a CDN management system and includes: Static embedding vectors are generated based on the static configuration information of nodes, wherein the static configuration information includes physical attributes, and the physical attributes include one or more of hardware configuration, network attributes, and geographic attributes; The time series of target performance indicators of the node in the historical period is calculated based on the time series dynamic model to obtain the time series hidden state vector. The static embedding vector and the temporal hidden state vector are input into the service quality prediction model to calculate the predicted value of the target performance index within a preset future time period; wherein... The process of calculating the predicted value includes constructing a Query with the static embedding vector, constructing a Key and a Value with the temporal hidden state vector, calculating an attention score by the dot product of the Query and the Key, and normalizing it to obtain an attention weight. The attention weight is applied to the Value to generate an attention output vector. The static embedding vector, the temporal hidden state vector, and the attention output vector are concatenated to construct a unified spatiotemporal fusion feature vector.

2. The method according to claim 1, characterized in that, The method for generating static embedding vectors based on node-based static configuration information includes: The static configuration information of the nodes is categorized; For discrete attributes, attribute values ​​are determined, and the embedding table is queried based on the attribute values ​​to obtain discrete attribute embedding vectors; For continuous attributes, the numerical values ​​are standardized to obtain a continuous attribute vector; The discrete attribute embedding vector and the continuous attribute vector are concatenated and then regularized to obtain the static embedding vector of the node.

3. The method according to claim 1, characterized in that, The method for calculating the time series hidden state vector based on the time series dynamic model includes: Sliding window sampling is performed based on the time series to obtain multiple sub-sequence samples; Initialize the hidden state; The data of each time step of the subsequence sample is sequentially input into the GRU, so that the GRU updates the current hidden state according to the current input and the hidden state of the previous input. Once all time-step data has been processed, the hidden state of the last time-step is extracted as the temporal hidden state vector.

4. The method according to claim 1, characterized in that, The method for calculating the predicted value also includes: The spatiotemporal fusion feature vectors are sequentially input into two fully connected neural networks for feature transformation and dimensionality reduction to obtain high-level feature vectors. The high-level feature vectors are mapped to the target prediction space through a linear output layer to generate a sequence of predicted service quality index values ​​for a future preset step size.

5. The method according to claim 1, characterized in that, The static configuration information also includes domain name business attributes, which include one or more of the following: business type, traffic level, acceleration region, and customer level.

6. A CND service quality early warning method, characterized in that, The method is applied to a CDN management system and includes: Collect and preprocess the static configuration information of the target nodes and the time series of the target performance indicators over historical periods; The predicted value of the target performance index within a preset future time period is calculated by the method according to any one of claims 1 to 5; Obtain the dynamic alarm threshold corresponding to the target performance index of the target node; The predicted value is compared with the dynamic alarm threshold to determine whether an early warning is triggered.

7. The method according to claim 6, characterized in that, The calculation method for the dynamic alarm threshold includes: Load the historical operational baseline statistics of the target node; The dynamic alarm threshold is dynamically calculated based on the historical operating baseline statistics.

8. The method according to claim 7, characterized in that, The historical operating baseline statistics are calculated and updated based on the EWMA algorithm.

9. A CDN management system, characterized in that, The system predicts the quality of service of nodes based on the method of any one of claims 1 to 5.

10. A CDN management system, characterized in that, The system implements CND service quality early warning management based on the method described in any one of claims 6 to 8.

Images