An overseas fusion CDN fault prediction method based on deep learning

Abstract

The application discloses an overseas fusion CDN fault prediction method based on deep learning, outputs a sliding window time sequence tensor; a bidirectional gate recurrent unit with different expansion coefficients is used to traverse the sliding window time sequence tensor in the forward and backward directions and output a global context embedding sequence; an aligned context embedding representation is generated and normalized, and a normalized aligned context embedding representation is output; an optimized multi-task transfer prediction model is obtained; a fault probability prediction result is compared with a dynamic fault threshold to generate a fault early warning signal, and an overseas fusion CDN intelligent operation and maintenance action is automatically triggered according to the fault early warning signal. The application improves operation and maintenance response efficiency and network service stability, and meets the high availability and high quality access requirements of global users.

Description

TECHNICAL FIELD

[0001] The application relates to the field of overseas fusion CDN technology, in particular to an overseas fusion CDN fault prediction method based on deep learning. BACKGROUND

[0002] With the continuous expansion of global Internet service demand, a content distribution network becomes infrastructure for guaranteeing efficient access of cross-regional and cross-operator users. In the overseas fusion content distribution network scenario, edge nodes and links in different regions are highly fragmented, network index data is scattered and the sampling environment is heterogeneous, and traditional fault prediction methods face many technical challenges.

[0003] Most of the existing content distribution network fault prediction adopts a single-domain modeling method, mainly relies on local historical data for model training, and it is difficult to realize effective migration and generalization of the model in a multi-regional or heterogeneous operator environment. The sample size of the target domain is often extremely limited at the initial stage of overseas node deployment, resulting in a significant decrease in prediction performance in the cold start stage of the model. The existing sequence prediction method generally uses a unidirectional recurrent neural network or a statistical rule based on a fixed threshold, and the capture ability is insufficient for frequent short-term jitter, long-term trend changes and complex time sequence dependence in the transoceanic link and multi-region environment.

[0004] The existing technology usually ignores the differences in network characteristics, sampling distribution and index mode between the source domain and the target domain, lacks effective domain adaptation and parameter sharing mechanism, and causes the prediction accuracy of the model to be greatly reduced after migration to a new overseas node. For the identification of fault precursors in a complex link environment, traditional technology is difficult to extract the precursor features of minor abnormalities from time series data, and it is easy to miss the best fault warning window. SUMMARY

[0005] One object of the application is to provide an overseas fusion CDN fault prediction method based on deep learning. The application improves the operation and maintenance response efficiency and network service stability, and meets the high availability and high quality access requirements of global users.

[0006] According to an overseas fusion CDN fault prediction method based on deep learning, the method comprises the following steps:

[0007] Collecting overseas fusion CDN network index data and preprocessing, outputting a sliding window time series tensor;

[0008] Inputting the sliding window time series tensor into a multi-resolution dilated bidirectional gated recurrent unit shared feature encoder, traversing the sliding window time series tensor in the forward and backward directions through the bidirectional gated recurrent unit with different dilation coefficients and outputting a global context embedding sequence;

[0009] The cross-domain attention is calculated based on the global context embedding sequence to fuse the source domain global context embedding sequence and the target domain global context embedding sequence, to generate an aligned context embedding representation, and the aligned context embedding representation is normalized to output a normalized aligned context embedding representation.

[0010] The normalized aligned context embedding representation is input into a multi-task transfer prediction model, and the multi-task transfer prediction model includes a fault probability classification task branch, a maximum load window regression task branch, and a network delay distribution segmentation regression task branch.

[0011] The multi-task transfer prediction model is trained to obtain an optimized multi-task transfer prediction model.

[0012] The standardized overseas converged CDN network index data collected in real time in the target domain is input into the optimized multi-task transfer prediction model to obtain a fault probability prediction result, a dynamic fault threshold is calculated based on the fault probability prediction result combined with historical threshold statistics and alarm sensitivity strategies, and the fault probability prediction result is compared with the dynamic fault threshold to generate a fault warning signal, and an overseas converged CDN intelligent operation action is automatically triggered according to the fault warning signal.

[0013] Optionally, the construction of the sliding window time series tensor comprises:

[0014] The overseas converged CDN network index data is collected in real time at a fixed sampling frequency, and the overseas converged CDN network index data includes a packet loss rate sequence, a round-trip delay sequence, a link jitter sequence, and a connection establishment time sequence.

[0015] The collected overseas converged CDN network index data is mapped to multiple time zones according to the collection time stamp of each time slice, the time stamp in each region is converted into a unified coordinated universal time time stamp, and a linear interpolation method is used to synchronize the time of data from different sources, to obtain time-aligned overseas converged CDN network index data.

[0016] The time-aligned overseas converged CDN network index data is normalized for each type of index to obtain normalized overseas converged CDN network index data.

[0017] The normalized overseas converged CDN network index data is divided by a sliding window, the normalized data is continuously sliced according to a fixed length time window, a Kalman smoothing algorithm is used to estimate and correct the abnormal fluctuations in the time series, to obtain a sliding window time series tensor.

[0018] Optionally, the multi-resolution dilated bidirectional gated recurrent unit shared feature encoder comprises:

[0019] The sliding window time series tensor is sequentially input into the multi-resolution dilated bidirectional gated recurrent unit shared feature encoder.

[0020] The multi-resolution dilatation bidirectional gated recurrent unit shared feature encoder constructs dilatation bidirectional gated recurrent units under multiple dilatation coefficient settings, performs feature encoding on each sliding window time series tensor, and for each dilatation coefficient, uses the corresponding dilatation bidirectional gated recurrent unit to traverse the sliding window time series tensor in both forward and backward directions with the dilatation interval as the step size, starting from the first time slice, to obtain the dilatation context representation.

[0021] By concatenating the dilated context representations obtained under all different dilation coefficients along the feature dimension, a multi-resolution context fusion representation is obtained.

[0022] A linear transformation is used to map the multi-resolution context fusion representation to the shared context embedding space, generating a shared context embedding representation. The shared context embedding representations of all sliding windows are arranged in order of sliding window number to construct a global context embedding sequence at the sliding window level.

[0023] Optionally, the output normalized aligned context embedding representation includes:

[0024] Obtain the global context embedding sequence at the sliding window level of the source domain and the global context embedding sequence at the sliding window level of the target domain, respectively;

[0025] A cross-domain attention mechanism is adopted to associate each target domain sliding window level global context embedding sequence with the reference window context embedding representation in the source domain sliding window level global context embedding sequence. Weights are assigned according to the attention mechanism, and the feature information of the reference window context embedding representation in the source domain sliding window level global context embedding sequence is used to perform weighted aggregation on the target domain sliding window level global context embedding sequence to obtain the cross-domain attention output of each target domain sliding window.

[0026] The global context embedding sequence at the sliding window level of each target domain is concatenated with the corresponding cross-domain attention output according to the feature dimension to obtain the aligned context embedding representation;

[0027] Domain-specific batch normalization is adopted. Independent normalization parameters are set for the overseas converged CDN network index data of different regions. The normalized aligned context embedding representation is obtained by subtracting the mean of the current region, dividing by the standard deviation of the current region, multiplying by the scaling factor and adding the offset factor.

[0028] ;

[0029] in, This represents the alignment context embedding vector for the t-th time slice. This represents the normalized output vector. , Let represent the mean and variance of the current batch in the target domain, respectively. For numerical stability, T is the total time slice. Indicates element-wise multiplication. This represents the scaling factor, which is calculated based on the standard deviation of each feature dimension within the same batch for all aligned context embeddings of the current target domain. The offset factor is calculated based on the mean of each feature dimension of all aligned context embeddings in the current target domain within the same batch.

[0030] Optionally, the multi-task migration prediction model includes:

[0031] The normalized aligned context embedding representation is used as input and fed into a unified multi-task transfer prediction model. The normalized aligned context embedding representation is then subjected to task-shared encoding to obtain a task-shared feature sequence.

[0032] The shared feature sequence of the task is input into the fault probability classification task branch. Attention-weighted pooling is applied in the time dimension to obtain the sliding window feature representation. The sliding window feature representation is input into the fully connected layer and then into the softmax function to output the fault occurrence probability.

[0033] Input the task-shared feature sequence into the maximum load window regression task branch, and output the maximum load index value within the sliding window;

[0034] Input the task-shared feature sequence into the delay distribution and segment the regression task branch to output the predicted probability of multiple delay intervals within the sliding window;

[0035] The delayed distribution piecewise regression task branch extracts delayed context features, which are the input to the delayed distribution regression task. The delayed distribution regression task multiplies these features with the delayed distribution regression weight matrix and adds the delayed distribution regression bias vector to output the predicted probabilities of multiple delayed intervals within the sliding window. The predicted probabilities of the delayed intervals are multiple real numbers, each representing an estimated value for the corresponding delayed interval.

[0036] Optionally, training the multi-task transfer prediction model includes:

[0037] During the training phase, a gradient reversal layer is inserted before the multi-task transfer prediction model and connected to the domain discriminator. By minimizing the joint objective of the domain discriminant loss and the multi-task loss through adversarial training, the normalized aligned context embedding representation learns domain-independent features, resulting in optimized multi-task transfer prediction model parameters. During the training phase, window-level contrastive learning between the source domain and the target domain is performed on the normalized aligned context embedding representation in parallel. Positive and negative sample pairs are used to construct a contrastive learning loss, which is jointly optimized with the domain discriminant loss and the multi-task loss to enhance cross-domain discrimination capability, further optimizing the multi-task transfer prediction model parameters.

[0038] Optionally, the automatic triggering of overseas converged CDN intelligent operation and maintenance actions based on fault warning signals includes:

[0039] The standardized overseas converged CDN network indicator data collected in real time during the inference phase of the target domain are segmented according to the sliding window strategy consistent with the training phase and converted into a sliding window time series tensor of the target domain.

[0040] Input the target domain sliding window time series tensor into the optimized multi-task migration prediction model to obtain the fault prediction probability at the sliding window level.

[0041] Based on the fault prediction probability, combined with the historical statistical distribution of fault occurrence samples and the alarm sensitivity parameters set for the target domain, a dynamic fault threshold is calculated.

[0042] The fault prediction probability under the current sliding window is compared with the dynamic fault threshold. When the fault prediction probability is greater than the dynamic fault threshold, the current sliding window is determined to be in a high-risk state, and a fault warning signal is generated.

[0043] Based on the judgment results of the fault warning signal, the overseas converged CDN intelligent operation and maintenance action is automatically triggered.

[0044] Optionally, the intelligent operation and maintenance actions include the following three categories:

[0045] DNS resolution switching: When the target domain edge node is judged to be in a high-risk state for three consecutive sliding windows, and the probability of the most recent failure prediction is greater than 0.9, DNS cache refresh and primary / backup address weight switching are automatically initiated, and access traffic is preferentially guided to the backup edge node with low predicted risk.

[0046] Link traffic migration: When the alarm level included in the fault warning signal is severe, and the current predicted load value of the target domain edge node exceeds 80% of the historical average, some traffic will be migrated to the link covered by the predicted low-risk node in the same region, so that the link load of the target node is reduced to a safe range.

[0047] Cache Reconstruction: When the predicted latency distribution of the target edge node in the fault warning signal fluctuates abnormally in multiple intervals, and the predicted latency exceeds the historical peak standard deviation three times in the highest interval, the edge cache cleanup and core resource hot reconstruction operation are automatically triggered, prioritizing the retention of high-frequency access content and rebuilding the cache structure of low-hit content.

[0048] The beneficial effects of this invention are:

[0049] This invention employs a method combining bidirectional gated recurrent units (BRUs) and multi-task transfer learning. By designing a multi-resolution dilated bidirectional gated recurrent unit encoder and a cross-domain attention fusion mechanism, it achieves efficient generalization in multi-regional, multi-operator, and small-sample environments. Relying on domain-adaptive batch normalization and adversarial training strategies, it can maintain a target node fault prediction recall rate of over 90% even in cold-start scenarios where the target domain accounts for only 10% of the source domain's sample size. This effectively improves the model's transferability and practicality in overseas fragmented CDN environments, significantly reduces the manual cost of deploying and adjusting new node models, and enhances the agility of cross-domain intelligent CDN operation and maintenance.

[0050] The multi-resolution dilatational bidirectional gated cyclic encoder proposed in this invention can simultaneously capture short-term high-frequency fluctuations and long-term trend information of network indicator data. It effectively models the temporal dependency relationship before and after minor anomalies and sudden failures through bidirectional structure. Combined with multi-task transfer learning, it incorporates three major network state indicators, namely failure probability, load peak, and delay distribution, into a unified shared representation space for collaborative modeling. This improves the model's sensitivity to the leading features of minor anomalies. According to actual engineering evaluation, it can issue high-confidence warnings on average 15 minutes in advance in overseas link high-fluctuation scenarios, significantly extending the window for automatic traffic switching and protection during operation and maintenance.

[0051] This invention enables real-time status monitoring, risk assessment, and dynamic response of CDN nodes and links. The system can automatically switch DNS resolution weights, migrate link traffic, and rebuild edge caches according to different alarm levels, effectively reducing the risk of service interruption caused by node failures. In large-scale overseas CDN intelligent operation and maintenance scenarios, it significantly improves operation and maintenance response efficiency and network service stability, meeting the high availability and high quality access needs of global users. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart of a deep learning-based overseas fusion CDN fault prediction method proposed in this invention. Detailed Implementation

[0054] Example 1:

[0055] refer to Figure 1 A deep learning-based method for predicting faults in overseas integrated CDNs, comprising:

[0056] Collect overseas converged CDN network indicator data, preprocess it, and output a sliding window time series tensor;

[0057] In this embodiment, the construction of the sliding window time series tensor includes:

[0058] Overseas converged CDN network indicator data are collected in real time at a fixed sampling frequency. The overseas converged CDN network indicator data includes packet loss rate sequence, round-trip delay sequence, link jitter sequence, and connection establishment time sequence.

[0059] The packet loss rate sequence represents the average packet loss rate within each time slice, in percentage; the round-trip delay sequence represents the average round-trip delay within each time slice, in milliseconds; the link jitter sequence represents the standard deviation of link delay within each time slice, in milliseconds; the connection establishment time sequence represents the average connection establishment time within each time slice, in milliseconds; each time slice is numbered in ascending order of its sequence number.

[0060] The collected overseas converged CDN network indicator data is time-mapped to multiple time zones according to the collection timestamp of each time slice. The timestamps under each region are converted into a unified Coordinated Universal Time timestamp. Then, the linear interpolation method is used to perform time synchronization processing on the data from different sources to obtain time-aligned overseas converged CDN network indicator data.

[0061] For the time-aligned overseas converged CDN network indicator data, normalization is performed on each type of indicator to obtain normalized overseas converged CDN network indicator data.

[0062] The normalization process is as follows: the mean of each type of overseas converged CDN network indicator over all time slices is used to subtract the indicator value for each time slice, and the standard deviation of each type of overseas converged CDN network indicator over all time slices is used to normalize the difference, so as to obtain the normalized packet loss rate, normalized round-trip latency, normalized link jitter, and normalized connection establishment time.

[0063] The normalized overseas converged CDN network index data is divided into sliding window segments. The normalized data is continuously sliced ​​according to a fixed-length time window. The Kalman smoothing algorithm is used to estimate and correct abnormal fluctuations in the time series, and the sliding window time series tensor is obtained.

[0064] Each sliding window time series tensor consists of normalized packet loss rate, normalized round-trip time, normalized link jitter, and normalized connection establishment time for several consecutive time slices within the window. Each row corresponds to the four types of normalized index data for one time slice. The total number of sliding window time series tensors is determined by the total length of normalized data, the window length, and the sliding step size.

[0065] The Kalman smoothing algorithm models the state transition relationship between the current state estimate and the previous state estimate of each normalized metric in each time slice, and updates the state estimate of each metric by combining the current observation and observation noise, thereby obtaining the packet loss rate state estimate, round-trip delay state estimate, link jitter state estimate, and connection establishment time state estimate for each time slice.

[0066] The sliding window time series tensor is input into the multi-resolution dilated bidirectional gated recurrent unit shared feature encoder. The sliding window time series tensor is traversed in both forward and backward directions by the bidirectional gated recurrent unit with different dilation coefficients, and the global context embedding sequence is output.

[0067] In this embodiment, the multi-resolution dilatational bidirectional gated cyclic unit shares a feature encoder, including:

[0068] The sliding window time series tensor is sequentially input into the multi-resolution dilated bidirectional gated recurrent unit shared feature encoder.

[0069] Each sliding window time series tensor consists of normalized packet loss rate state estimates, normalized round-trip time state estimates, normalized link jitter state estimates, and normalized connection establishment time state estimates for consecutive time slices. The k-th sliding window time series tensor contains four types of normalized index data for each time slice. Each row represents an index for a time slice, and the order of all rows is consistent with the order of the time slices.

[0070] The multi-resolution dilatation bidirectional gated recurrent unit shared feature encoder constructs dilatation bidirectional gated recurrent units under multiple dilatation coefficient settings, performs feature encoding on each sliding window time series tensor, and for each dilatation coefficient, uses the corresponding dilatation bidirectional gated recurrent unit to traverse the sliding window time series tensor in both forward and backward directions with the dilatation interval as the step size, starting from the first time slice, to obtain the dilatation context representation.

[0071] In Example 1, dilated bidirectional gated recurrent units are constructed under multiple dilation coefficient settings. Feature encoding is performed on each sliding window time series tensor. For each dilation coefficient, the corresponding dilated bidirectional gated recurrent unit starts from the first time slice and samples the input sequence data in the time slice sequence in a skip manner with the dilation interval as the step size. During the forward traversal, the dilated bidirectional gated recurrent unit collects the input features of one time slice in chronological order, starting from the first time slice, and sends the sequentially obtained input features to the gated unit to update the forward hidden state sequence until the end of the sequence. During the backward traversal, the dilated bidirectional gated recurrent unit collects the input features of one time slice in reverse chronological order, starting from the last time slice, and sends the sequentially obtained input features to the gated unit to update the backward hidden state sequence until the beginning of the sequence. For each time slice, the forward hidden state in the forward hidden state sequence and the backward hidden state in the backward hidden state sequence are concatenated to form the dilated context representation of the time slice. The set of dilated context representations of all time slices constitutes the dilated context representation under the dilation coefficient.

[0072] By concatenating the dilated context representations obtained under all different dilation coefficients along the feature dimension, a multi-resolution context fusion representation is obtained.

[0073] Multi-resolution context fusion representation merges the dilation context representations under each dilation coefficient in time slice order, so that the representation of each time slice contains parallel feature information under multiple dilation-aware scales.

[0074] A linear transformation is used to map the multi-resolution context fusion representation to the shared context embedding space, generating a shared context embedding representation. The shared context embedding representations of all sliding windows are arranged in order of sliding window number to construct a global context embedding sequence at the sliding window level.

[0075] The linear transformation is determined by both weight and bias parameters. The output shared context embedding representation corresponds one-to-one with the input time slice order. Each row represents the context embedding feature of the time slice, and the embedding features of all time slices have the same dimension.

[0076] Cross-domain attention is computed based on the global context embedding sequence to fuse the source domain global context embedding sequence and the target domain global context embedding sequence, generating an aligned context embedding representation, which is then normalized and output as a normalized aligned context embedding representation.

[0077] In this embodiment, the output normalized aligned context embedding representation includes:

[0078] Obtain the global context embedding sequence at the sliding window level of the source domain and the global context embedding sequence at the sliding window level of the target domain, respectively;

[0079] The source domain sliding window level global context embedding sequence is composed of the shared context embedding representations of all source domain sliding windows arranged in sequence. The target domain sliding window level global context embedding sequence is composed of the shared context embedding representations of all target domain sliding windows arranged in sequence. The shared context embedding representation of each sliding window is obtained by arranging the sliding window context features obtained in the previous step in chronological order.

[0080] A cross-domain attention mechanism is adopted to associate each target domain sliding window level global context embedding sequence with the reference window context embedding representation in the source domain sliding window level global context embedding sequence. Weights are assigned according to the attention mechanism, and the feature information of the reference window context embedding representation in the source domain sliding window level global context embedding sequence is used to perform weighted aggregation on the target domain sliding window level global context embedding sequence to obtain the cross-domain attention output of each target domain sliding window.

[0081] The global context embedding sequence at the sliding window level of each target domain is concatenated with the corresponding cross-domain attention output according to the feature dimension to obtain the aligned context embedding representation;

[0082] Aligned context embedding means that each target domain sliding window level global context embedding sequence contains both the target domain's own feature information and the source domain's supplementary feature information.

[0083] Domain-specific batch normalization is adopted. Independent normalization parameters are set for the overseas converged CDN network index data of different regions. The normalized aligned context embedding representation is obtained by subtracting the mean of the current region, dividing by the standard deviation of the current region, multiplying by the scaling factor and adding the offset factor.

[0084] ;

[0085] in, This represents the alignment context embedding vector for the t-th time slice. This represents the normalized output vector. , Let represent the mean and variance of the current batch in the target domain, respectively. For numerical stability, T is the total time slice. Indicates element-wise multiplication. This represents the scaling factor, which is calculated based on the standard deviation of each feature dimension within the same batch for all aligned context embeddings of the current target domain. The offset factor is calculated based on the mean of each feature dimension of all aligned context embeddings in the current target domain within the same batch.

[0086] The normalized aligned context embedding is used to represent the input to the multi-task migration prediction model, which includes a fault probability classification task branch, a maximum load window regression task branch, and a network latency distribution piecewise regression task branch.

[0087] In this embodiment, the multi-task migration prediction model includes:

[0088] The normalized aligned context embedding representation is used as input and fed into a unified multi-task transfer prediction model. The normalized aligned context embedding representation is then subjected to task-shared encoding to obtain a task-shared feature sequence.

[0089] The task-shared encoding process performs a unified feature transformation on the normalized context embedding vector of each time slice. The unified feature transformation process is as follows: multiply each normalized context embedding vector with the shared encoding weight matrix and then add the shared encoding bias vector to obtain the task-shared feature sequence. The feature dimension of each time slice in the task-shared feature sequence is the shared feature dimension.

[0090] The shared feature sequence of the task is input into the fault probability classification task branch. Attention-weighted pooling is applied in the time dimension to obtain the sliding window feature representation. The sliding window feature representation is input into the fully connected layer and then into the softmax function to output the fault occurrence probability.

[0091] The attention-weighted pooling operation is as follows: For the task-shared feature vector of each time slice, normalize all time slices by taking the dot product with the attention weight vector to obtain the attention score of each time slice. Then, multiply the task-shared feature vector of each time slice with its corresponding attention score and sum them in the time slice dimension to obtain the sliding window feature representation.

[0092] The probability of failure occurrence is a binary probability vector, where each term represents the predicted probability that the failure has not occurred and the predicted probability that the failure has occurred under the current sliding window.

[0093] Input the task-shared feature sequence into the maximum load window regression task branch, and output the maximum load index value within the sliding window;

[0094] The maximum load window regression task branch uses convolutional pooling to extract load context features. The load context features extracted by convolutional pooling are the input of the regression task. The maximum load window regression task outputs the maximum load index value within the sliding window by multiplying it with the regression weight vector and adding the regression bias vector. The maximum load index value is a real number that reflects the peak level of the load within the current sliding window.

[0095] Input the task-shared feature sequence into the delay distribution and segment the regression task branch to output the predicted probability of multiple delay intervals within the sliding window;

[0096] The delayed distribution piecewise regression task branch extracts delayed context features, which are the input to the delayed distribution regression task. The delayed distribution regression task multiplies these features with the delayed distribution regression weight matrix and adds the delayed distribution regression bias vector to output the predicted probabilities of multiple delayed intervals within the sliding window. The predicted probabilities of the delayed intervals are multiple real numbers, each representing an estimated value for the corresponding delayed interval.

[0097] The multi-task migration prediction model is trained to obtain the optimized multi-task migration prediction model;

[0098] In this embodiment, training the multi-task migration prediction model includes:

[0099] During the training phase, a gradient reversal layer is inserted before the multi-task transfer prediction model and connected to the domain discriminator. By minimizing the joint objective of the domain discriminant loss and the multi-task loss through adversarial training, the normalized aligned context embedding representation learns domain-independent features, resulting in optimized multi-task transfer prediction model parameters. During the training phase, window-level contrastive learning between the source domain and the target domain is performed on the normalized aligned context embedding representation in parallel. Positive and negative sample pairs are used to construct a contrastive learning loss, which is jointly optimized with the domain discriminant loss and the multi-task loss to enhance cross-domain discrimination capability, further optimizing the multi-task transfer prediction model parameters.

[0100] The standardized overseas converged CDN network indicator data collected in real time from the target domain is input into the optimized multi-task migration prediction model to obtain the failure probability prediction result. Based on the failure probability prediction result, combined with historical threshold statistics and alarm sensitivity strategy, the dynamic failure threshold is calculated. The failure probability prediction result is compared with the dynamic failure threshold to generate a failure warning signal. The overseas converged CDN intelligent operation and maintenance action is automatically triggered according to the failure warning signal.

[0101] In this embodiment, the overseas converged CDN intelligent operation and maintenance actions are automatically triggered based on the fault warning signal, including:

[0102] The standardized overseas converged CDN network indicator data collected in real time during the inference phase of the target domain are segmented according to the sliding window strategy consistent with the training phase and converted into a sliding window time series tensor of the target domain.

[0103] Input the target domain sliding window time series tensor into the optimized multi-task migration prediction model to obtain the fault prediction probability at the sliding window level.

[0104] Based on the fault prediction probability, combined with the historical statistical distribution of fault occurrence samples and the alarm sensitivity parameters set for the target domain, a dynamic fault threshold is calculated.

[0105] The dynamic fault threshold is updated using a sliding weighting method, with higher weight for recent windows and gradually decreasing weight for longer windows, ensuring that the dynamic fault threshold can reflect the fluctuation of the fault probability baseline in the current stage.

[0106] The fault prediction probability under the current sliding window is compared with the dynamic fault threshold. When the fault prediction probability is greater than the dynamic fault threshold, the current sliding window is determined to be in a high-risk state, and a fault warning signal is generated.

[0107] The fault warning signal carries a sliding window timestamp, fault probability value, CDN edge node identifier, and current alarm level.

[0108] Based on the judgment results of the fault warning signal, the overseas converged CDN intelligent operation and maintenance action is automatically triggered.

[0109] In this embodiment, intelligent operation and maintenance actions include the following three categories:

[0110] DNS resolution switching: When the target domain edge node is judged to be in a high-risk state for three consecutive sliding windows, and the probability of the most recent failure prediction is greater than 0.9, DNS cache refresh and primary / backup address weight switching are automatically initiated, and access traffic is preferentially guided to the backup edge node with low predicted risk.

[0111] Link traffic migration: When the alarm level included in the fault warning signal is severe, and the current predicted load value of the target domain edge node exceeds 80% of the historical average, some traffic will be migrated to the link covered by the predicted low-risk node in the same region, so that the link load of the target node is reduced to a safe range.

[0112] Cache Reconstruction: When the predicted latency distribution of the target edge node in the fault warning signal fluctuates abnormally in multiple intervals, and the predicted latency exceeds the historical peak standard deviation three times in the highest interval, the edge cache cleanup and core resource hot reconstruction operation are automatically triggered, prioritizing the retention of high-frequency access content and rebuilding the cache structure of low-hit content.

[0113] The intelligent operation and maintenance actions can be combined or executed individually as needed. The execution method can be automatic or semi-automatic after manual review. The purpose is to reduce the risk of service interruption caused by potential node failures and improve the stability and access availability of overseas integrated CDN in cross-border and multi-carrier scenarios.

[0114] Example 2:

[0115] This embodiment takes the deployment of a new overseas integrated CDN intelligent operation and maintenance platform by a large international Internet content service provider in Southeast Asia and North America as an example. In the actual operation and maintenance scenario, the service provider plans to deploy the existing CDN core scheduling system to the newly expanded Southeast Asian region (countries A, B, and C) and North American region (countries D, E, and F). The goal is to provide smooth online video, live streaming, game acceleration, and large file distribution services for local and international users. Due to the great differences in network operators, laws and policies, link topology, and user access patterns in these two regions, the node distribution is fragmented and the data samples are highly uneven. Traditional methods have obvious shortcomings in cold start of new nodes and early warning of complex links.

[0116] The team deployed standardized network telemetry probes for each overseas CDN edge node (taking SG-Edge-03 in country A and US-Edge-17 in country D as examples). Every 5 seconds, the team collected the following multi-dimensional network metrics in real time: packet loss rate (%), round-trip time (ms), link jitter (ms), connection establishment time (ms), TCP retransmission rate (%), and cache hit rate (%). Taking SG-Edge-03 in country A as an example, 17,280 valid data points were collected in the past 24 hours. Statistics showed that the node's average packet loss rate was 0.51%, with a peak of 6.72%; the average RTT was 26.7ms, with a fluctuation range of 16.4~47.9ms; the average TCP retransmission rate was 2.3%, with occasional spikes reaching 9.8%. Compared to North American nodes, the network quality of country A nodes is more sensitive to the impact of transoceanic cable traffic, and the data distribution exhibits stronger heteroscedasticity.

[0117] For the indicator data, all original time-series data were aligned to UTC based on timestamps. Taking 02:00 AM on April 8, 2024 (SG time) as an example, the local time of the node in country A was accurately mapped to the globally unified UTC timeline through multi-time zone synchronization. To reduce the randomness and noise of the indicator data, the Kalman filtering algorithm was used to smooth all time-series data. The actual results show that after filtering, the original abnormal fluctuations (such as a sudden increase in the loss rate of individual packets) were effectively corrected, and missing data points were automatically filled by interpolation, which improved the continuity and robustness of the sequence.

[0118] All aligned and smoothed data are divided into multiple time series tensors using a sliding window of length 120 (10 minutes) and a step size of 12 (1 minute). Compared with the traditional single-point sampling method (e.g., once every 30 minutes), the data sample size in this embodiment is increased by more than 10 times, providing a more sufficient feature base for the model.

[0119] The sliding window time series is input into a multi-resolution dilated bidirectional gated recurrent unit (RNN) shared feature encoder. Taking window number #158 (April 8, 2024, 15:20-15:30) as an example, after modeling with three sets of dilation coefficients (d=1, d=2, d=4), the model can simultaneously extract the pattern of small fluctuations in packet loss rate at the beginning of the window (first 4 minutes) and significant increases in RTT and jitter at the end (last 2 minutes). The bidirectional structure ensures that the model can capture both short-term high-frequency anomalies and track long-term trend shifts, which is far superior to the structure of traditional unidirectional RNNs that can only trace the past and cannot perceive future dependencies. The context embedding sequence output by the feature encoding is linearly transformed and batch normalized before being fed into the downstream multi-task prediction branch.

[0120] In the multi-task transfer learning stage, this embodiment uses historical samples from two major regions, North America and Southeast Asia, as the source domain and the new SG-Edge-03 node as the target domain. The comparison data is as follows: the North American source domain has 40,000 complete training samples, while the SG-Edge-03 target domain initially only has 3,500 sliding window samples. Through the multi-task transfer learning mechanism of this invention, the contextual embedding features of the source domain and the target domain participate in gradient update after alignment. Less than 10% of the samples in the target domain can support the cold start of the model and rapidly improve the prediction accuracy.

[0121] Specifically, when there were only 3,500 samples in the target domain initially, the fault prediction recall rate of the model of this invention reached 91.6%, which is significantly higher than the 59.4% of the traditional pre-transfer model.

[0122] In the actual testing phase, taking the network operation and maintenance log from 16:00 to 19:00 on April 8, 2024 as an example, the SG-Edge-03 node experienced two real failure events. The first was from 16:44 to 16:55, with the link packet loss consistently exceeding 4.8% and the cache hit rate plummeting, causing a brief interruption in the live stream. The second was from 18:10 to 18:16, with a sudden increase in RTT and link jitter, and the TCP retransmission rate continuously exceeding 8%. The method of this invention issued high-confidence alarms 19 minutes (16:25) and 17 minutes (17:53) before the failure occurred, respectively. DNS resolution and link traffic scheduling were completed automatically, and the user experience did not experience large-scale interruptions. Under the same conditions, the traditional method based on unidirectional GRU + fixed threshold only issued alarms 4 minutes in advance for the first time and only 2 minutes in advance for the second time, and frequently missed alarms in scenarios with multiple abnormal indicators, leading to passive manual intervention on-site.

[0123] Further comparison from a multi-task collaboration perspective reveals that the method of this invention can not only simultaneously output the fault probability (P_fault=0.93 / 0.87), maximum load (peak 876Mbps / 943Mbps), and delay distribution prediction (P90=54ms / 62ms), but also automatically and dynamically adjust the loss weights of each branch, achieving rapid adaptation under data distribution drift. Traditional methods have a mean squared error of 17.3Mbps for maximum load prediction and an average error of 13.6ms for delay P90 prediction; while the multi-task model in this embodiment has a maximum load prediction error of only 7.5Mbps and a delay P90 error reduced to 4.2ms, resulting in more accurate predictions and stronger generalization.

[0124] In terms of cross-domain migration and cold start, the method of this invention also has the following data advantages: when the number of samples collected in the SG-Edge-03 target domain increases from 1,000 to 3,500, the model recall rate rapidly increases from 85.2% to 91.6%. Compared with the traditional migration model which only improves by less than 6%, the migration agility of this method is significantly higher.

[0125] The method of this invention remains stable under large-scale nodes. Taking the North American US-Edge-17 node on April 9, 2024 as an example, the average lead time for fault prediction over 48 consecutive hours was 15.7 minutes and the F1 score was 0.912, both of which are higher than the lead time (5.6 minutes) and F1 score (0.764) of the traditional method. Test data of other new nodes deployed in the same batch (MY-Edge-09 in country B and US-Edge-21 in country E) also showed that the false alarm rate decreased by 40%, the false alarm rate decreased by 37%, and the average response time for automatically triggering DNS resolution, link switching and cache reconstruction was shortened to 4.2 seconds.

[0126] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for predicting faults in overseas integrated CDNs, characterized in that, include: Collect overseas converged CDN network indicator data, preprocess it, and output a sliding window time series tensor; The sliding window time series tensor is input into the multi-resolution dilated bidirectional gated recurrent unit shared feature encoder. The sliding window time series tensor is traversed in both forward and backward directions by the bidirectional gated recurrent unit with different dilation coefficients, and the global context embedding sequence is output. Cross-domain attention is computed based on the global context embedding sequence to fuse the source domain global context embedding sequence and the target domain global context embedding sequence, generating an aligned context embedding representation, which is then normalized and output as a normalized aligned context embedding representation. The normalized aligned context embedding is used to represent the input to the multi-task migration prediction model, which includes a fault probability classification task branch, a maximum load window regression task branch, and a network latency distribution piecewise regression task branch. The multi-task transfer prediction model is trained to obtain the optimized multi-task transfer prediction model; Standardized overseas converged CDN network indicator data collected in real time from the target domain is input into the optimized multi-task migration prediction model to obtain the failure probability prediction result. Based on the failure probability prediction result, combined with historical threshold statistics and alarm sensitivity strategy, a dynamic failure threshold is calculated. The failure probability prediction result is compared with the dynamic failure threshold to generate a failure warning signal. Based on the failure warning signal, the overseas converged CDN intelligent operation and maintenance actions are automatically triggered. The step of inputting the sliding window time series tensor into a multi-resolution dilated bidirectional gated recurrent unit shared feature encoder, and traversing the sliding window time series tensor in both forward and backward directions through a bidirectional gated recurrent unit with different dilation coefficients to output a global context embedding sequence includes: The sliding window time series tensor is sequentially input into the multi-resolution dilated bidirectional gated recurrent unit shared feature encoder. The multi-resolution dilatation bidirectional gated recurrent unit shared feature encoder constructs dilatation bidirectional gated recurrent units under multiple dilatation coefficient settings, performs feature encoding on each sliding window time series tensor, and for each dilatation coefficient, uses the corresponding dilatation bidirectional gated recurrent unit to traverse the sliding window time series tensor in both forward and backward directions with the dilatation interval as the step size, starting from the first time slice, to obtain the dilatation context representation. By concatenating the dilated context representations obtained under all different dilation coefficients along the feature dimension, a multi-resolution context fusion representation is obtained. A linear transformation is used to map the multi-resolution context fusion representation to the shared context embedding space, generating a shared context embedding representation. The shared context embedding representations of all sliding windows are arranged in order of sliding window number to construct a global context embedding sequence at the sliding window level.

2. The method for predicting faults in an overseas fusion CDN based on deep learning according to claim 1, characterized in that, The construction of the sliding window time series tensor includes: Overseas converged CDN network indicator data are collected in real time at a fixed sampling frequency. The overseas converged CDN network indicator data includes packet loss rate sequence, round-trip delay sequence, link jitter sequence, and connection establishment time sequence. The collected overseas converged CDN network indicator data is time-mapped to multiple time zones according to the collection timestamp of each time slice. The timestamps under each region are converted into a unified Coordinated Universal Time stamp. Then, the linear interpolation method is used to perform time synchronization processing on the data from different sources to obtain time-aligned overseas converged CDN network indicator data. For the time-aligned overseas converged CDN network indicator data, normalization is performed on each type of indicator to obtain normalized overseas converged CDN network indicator data. The normalized overseas converged CDN network index data is divided into sliding window segments. The normalized data is continuously sliced ​​according to a fixed-length time window. The Kalman smoothing algorithm is used to estimate and correct abnormal fluctuations in the time series, and the sliding window time series tensor is obtained.

3. The method for predicting faults in an overseas fusion CDN based on deep learning according to claim 1, characterized in that, The output normalized aligned context embedding representation includes: Obtain the global context embedding sequence at the sliding window level of the source domain and the global context embedding sequence at the sliding window level of the target domain, respectively; A cross-domain attention mechanism is adopted to associate each target domain sliding window level global context embedding sequence with the reference window context embedding representation in the source domain sliding window level global context embedding sequence. Weights are assigned according to the attention mechanism, and the feature information of the reference window context embedding representation in the source domain sliding window level global context embedding sequence is used to perform weighted aggregation on the target domain sliding window level global context embedding sequence to obtain the cross-domain attention output of each target domain sliding window. The global context embedding sequence at the sliding window level of each target domain is concatenated with the corresponding cross-domain attention output according to the feature dimension to obtain the aligned context embedding representation; Domain-specific batch normalization is adopted. Independent normalization parameters are set for the overseas converged CDN network index data of different regions. The normalized aligned context embedding representation is obtained by subtracting the mean of the current region, dividing by the standard deviation of the current region, multiplying by the scaling factor and adding the offset factor. ； in, This represents the alignment context embedding vector for the t-th time slice. This represents the normalized output vector. , Let represent the mean and variance of the current batch in the target domain, respectively. For numerical stability, T is the total time slice. Indicates element-wise multiplication. This represents the scaling factor, which is calculated based on the standard deviation of each feature dimension within the same batch for all aligned context embeddings of the current target domain. The offset factor is calculated based on the mean of each feature dimension of all aligned context embeddings in the current target domain within the same batch.

4. The method for predicting faults in an overseas fusion CDN based on deep learning according to claim 1, characterized in that, The multi-task migration prediction model includes: The normalized aligned context embedding representation is used as input and fed into a unified multi-task transfer prediction model. The normalized aligned context embedding representation is then subjected to task-shared encoding to obtain a task-shared feature sequence. The shared feature sequence of the task is input into the fault probability classification task branch. Attention-weighted pooling is applied in the time dimension to obtain the sliding window feature representation. The sliding window feature representation is input into the fully connected layer and then into the softmax function to output the fault occurrence probability. Input the task-shared feature sequence into the maximum load window regression task branch, and output the maximum load index value within the sliding window; Input the task-shared feature sequence into the delay distribution and segment the regression task branch to output the predicted probability of multiple delay intervals within the sliding window; The delayed distribution piecewise regression task branch extracts delayed context features, which serve as the input to the delayed distribution regression task. The delayed distribution regression task multiplies these features with the delayed distribution regression weight matrix and adds the delayed distribution regression bias vector to output the predicted probabilities of multiple delayed intervals within the sliding window. The predicted probabilities of each delayed interval are multiple real numbers, each representing an estimated value for the corresponding delayed interval.

5. The method for predicting faults in an overseas fusion CDN based on deep learning according to claim 1, characterized in that, The training of the multi-task transfer prediction model includes: During the training phase, a gradient reversal layer is inserted before the multi-task transfer prediction model and connected to the domain discriminator. By minimizing the joint objective of the domain discriminant loss and the multi-task loss through adversarial training, the normalized aligned context embedding representation learns domain-independent features, resulting in optimized multi-task transfer prediction model parameters. During the training phase, window-level contrastive learning between the source domain and the target domain is performed on the normalized aligned context embedding representation in parallel. Positive and negative sample pairs are used to construct a contrastive learning loss, which is jointly optimized with the domain discriminant loss and the multi-task loss to enhance cross-domain discrimination capability, further optimizing the multi-task transfer prediction model parameters.

6. The method for predicting faults in an overseas fusion CDN based on deep learning according to claim 1, characterized in that, The automatic triggering of overseas converged CDN intelligent operation and maintenance actions based on fault warning signals includes: The standardized overseas converged CDN network indicator data collected in real time during the inference phase of the target domain are segmented according to the sliding window strategy consistent with the training phase and converted into a sliding window time series tensor of the target domain. Input the target domain sliding window time series tensor into the optimized multi-task migration prediction model to obtain the fault prediction probability at the sliding window level. Based on the fault prediction probability, combined with the historical statistical distribution of fault occurrence samples and the alarm sensitivity parameters set for the target domain, a dynamic fault threshold is calculated. The fault prediction probability under the current sliding window is compared with the dynamic fault threshold. When the fault prediction probability is greater than the dynamic fault threshold, the current sliding window is determined to be in a high-risk state, and a fault warning signal is generated. Based on the judgment results of the fault warning signal, the overseas converged CDN intelligent operation and maintenance action is automatically triggered.

7. The method for predicting faults in an overseas fusion CDN based on deep learning according to claim 1, characterized in that, The intelligent operation and maintenance actions include the following three categories: DNS resolution switching: When the target domain edge node is judged to be in a high-risk state for three consecutive sliding windows, and the probability of the most recent failure prediction is greater than 0.9, DNS cache refresh and primary / backup address weight switching are automatically initiated, and access traffic is preferentially guided to the backup edge node with low predicted risk. Link traffic migration: When the alarm level included in the fault warning signal is severe, and the current predicted load value of the target domain edge node exceeds 80% of the historical average, some traffic will be migrated to the link covered by the predicted low-risk node in the same region, so that the link load of the target node is reduced to a safe range. Cache Reconstruction: When the predicted latency distribution of the target edge node in the fault warning signal fluctuates abnormally in multiple intervals, and the predicted latency exceeds the historical peak standard deviation three times in the highest interval, the edge cache cleanup and core resource hot reconstruction operation are automatically triggered, prioritizing the retention of high-frequency access content and rebuilding the cache structure of low-hit content.

Images