Big data server early warning system based on artificial intelligence

Abstract

The application relates to the field of artificial intelligence and big data processing technology, and discloses a big data server early warning system based on artificial intelligence, which comprises a data acquisition module, a embedding modeling module and an abnormal confidence calculation module.The data acquisition module is used for collecting running indexes of a server cluster at multiple time points and forming a time sequence index vector sequence.The embedding modeling module is used for receiving the time sequence index vector sequence and inputting the same into a trained neural network model for low-dimensional potential space embedding.The neural network model comprises an encoder and a decoder based on a variational self-encoding structure.The abnormal confidence calculation module is used for receiving a potential space representation output by the embedding modeling module and calculating a corresponding abnormal confidence score based on an encoding offset between multiple perturbation samples and original input in the potential space.The application realizes the stability improvement of server state recognition and the high-precision response of abnormal early warning by constructing a dynamic scoring and judging mechanism based on potential space perturbation offset.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and big data processing technology, specifically to an artificial intelligence-based big data server early warning system. Background Technology

[0002] In modern data centers and large-scale server clusters, real-time monitoring and early warning of server operational status have become crucial for ensuring business continuity and service availability. If server anomalies are not detected in time, they can lead to performance bottlenecks, service interruptions, data loss, or even system-wide crashes. Therefore, building a timely, accurate, and adaptive server early warning system has become a real need in the operation and maintenance of various internet platforms, cloud computing service providers, and high-availability financial systems.

[0003] Existing server alerting methods are mostly based on static rule systems and threshold settings. The main advantages of these approaches are their simplicity, low deployment cost, and effectiveness in some application scenarios with stable structures and minimal load changes. For example, common detection methods based on sliding window statistics or moving averages can respond quickly to sudden changes in system load, and some machine learning algorithms, such as decision trees, also demonstrate high accuracy when handling low-dimensional features. These methods achieve good recognition results in scenarios with well-defined detection rules or significant differences in anomalous samples.

[0004] However, these traditional technical approaches still have many problems. On the one hand, static threshold mechanisms lack adaptability and are prone to misjudgment when faced with data drift or periodic load fluctuations, making them inflexible. On the other hand, direct modeling of the original indicator space is very sensitive to input perturbations, resulting in poor model output stability and unreliable identification results. Furthermore, relying on single-point scoring methods cannot fully reflect the model's response characteristics to anomalies and has weak anti-interference capabilities. More importantly, shallow feature extraction methods ignore the deep structural relationships between indicators, making it difficult to capture potential chains of abnormal behavior. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an AI-based big data server early warning system, which solves the problems of anomaly detection relying on static thresholds, models being sensitive to input perturbations, scoring mechanisms lacking robustness, and the inability to accurately model the potential structural relationships between high-dimensional indicators.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an artificial intelligence-based big data server early warning system, comprising: The data acquisition module is used to collect the operating metrics of the server cluster at multiple points in time and construct them into a time-series metric vector sequence. An embedding modeling module is used to receive the time-series index vector sequence and input it into a trained neural network model for low-dimensional latent space embedding. The neural network model includes an encoder and a decoder based on a variational autoencoder structure. An anomaly confidence calculation module is used to receive the latent space representation output by the embedded modeling module, and calculate the corresponding anomaly confidence score based on the encoding offset between multiple perturbation samples and the original input in the latent space. The early warning judgment module is used to receive the abnormal confidence score, construct a dynamic threshold based on the statistical information of historical confidence scores, and compare the current confidence score with the dynamic threshold to determine whether to generate a server abnormality early warning signal.

[0007] Preferably, the data acquisition module includes: The performance metrics collection unit is used to collect multi-dimensional performance metrics from the server cluster, including CPU utilization, memory usage, disk speed, and network traffic. The time-series construction unit is used to organize the running indicators into indicator vectors in chronological order, forming a continuous time-series indicator vector sequence. The indicator graph construction unit is used to construct an undirected weighted graph based on the correlation calculation results between the indicators in the indicator vector. The correlation calculation uses the Pearson correlation coefficient, and the edge weights are the absolute values ​​of the correlation coefficients.

[0008] Preferably, the embedded modeling module includes: The encoder unit is used to input and encode a time-series index vector sequence into a low-dimensional latent space representation. The encoder is based on a variational autoencoder structure and outputs latent space distribution parameters. Decoder unit, used to receive the latent space representation and reconstruct the index vector; The graph regularization unit is used to apply Laplacian regularization to the latent space representation based on the index correlation graph during the embedding learning process. The optimized loss function includes reconstruction error, KL divergence term, and graph regularization term, where the graph regularization term is: ; in, For Laplace matrix, The trace of a matrix is ​​the sum of the elements on its diagonal. Representation matrix transpose, This represents the latent representation matrix.

[0009] Preferably, the anomaly confidence calculation module includes: The perturbation generation unit is used to generate multiple perturbation samples based on the original time series index vector. The perturbation samples are obtained by adding Gaussian noise with zero mean and a set variance. The encoding offset calculation unit is used to input the perturbation sample and the original input to the encoder unit respectively, obtain the corresponding latent space representation, and calculate the offset between the perturbation sample and the latent space representation of the original sample; The confidence score unit is used to calculate the average of the encoding offset results of all perturbation samples and generate an anomaly confidence score, which is used to reflect the sensitivity of the sample to the perturbation.

[0010] Preferably, the early warning judgment module includes: The dynamic threshold calculation unit is used to calculate the mean and standard deviation of historical anomaly confidence scores within a preset window, and then calculate the dynamic threshold. The dynamic threshold calculation formula is as follows: ; in, The standard deviation vector of the latent variables, also output by the encoder network, controls the scale of the latent space distribution. This represents the mean vector of the latent variables. This is the sensitivity coefficient. This indicates a dynamic threshold.

[0011] The anomaly detection unit is used to compare the current anomaly confidence score with a dynamic threshold. When the anomaly confidence score is greater than the dynamic threshold, a server anomaly warning signal is output.

[0012] Preferably, the indicator graph construction unit includes: The correlation calculation unit is used to perform Pearson correlation calculation between any two indicators in the time series indicator vector sequence to obtain the correlation coefficient matrix. The weighted graph generation unit is used to construct an undirected weighted graph by using the absolute value of the correlation coefficient matrix as the edge weight. Each node in the graph corresponds to an indicator dimension, and each edge represents the correlation strength between indicators.

[0013] Preferably, the encoder unit includes: The parameter extraction unit is used to extract the mean vector and variance vector of the input vector, which are used to construct the distribution of the latent variables. The reparameter sampling unit is used to sample the latent variables based on the mean and variance using the reparameter technique to achieve a differentiable training process. The embedded output unit is used to output the sampled latent variables as a latent space representation to the decoder unit and subsequent modules.

[0014] Preferably, the encoding offset calculation unit includes: The original encoding generation unit is used to input the original input index vector into the encoder unit to obtain its corresponding latent space representation; The perturbation coding generation unit is used to input multiple perturbation samples into the encoder to generate corresponding perturbation latent representations. The offset calculation unit is used to calculate the Euclidean distance between each perturbation potential representation and the original potential representation, and stores all distance values ​​for subsequent scoring.

[0015] Preferably, the confidence scoring unit includes: The offset aggregation unit is used to receive the potential representation offset of each perturbation sample and calculate its average value. The confidence generation unit is used to output the average offset value as the abnormal confidence score of the input vector, which reflects the sensitivity of the input to disturbances.

[0016] Preferably, the anomaly detection unit includes: The threshold comparison unit is used to compare the current anomaly confidence score with the dynamic threshold. The early warning signal output unit is used to generate an abnormal alarm signal when the current confidence score is greater than the dynamic threshold, and send the signal to the system operation and maintenance platform for users to receive and handle.

[0017] This invention provides an artificial intelligence-based early warning system for big data servers. It has the following beneficial effects: 1. This invention employs a technical solution that constructs anomaly confidence scores based on the response offset of perturbed samples in the latent space, achieving a dynamic reflection of the sensitivity of server operating status. Compared to existing technologies that conventionally rely on reconstruction errors or static thresholds for judgment, this solves the problems of unstable anomaly pattern recognition and difficulty in adapting to changing data characteristics.

[0018] 2. By introducing a variational autoencoder structure to model the latent spatial embedding of time-series index sequences, this invention achieves effective compressed representation and structure preservation of high-dimensional complex index data. Compared with traditional methods based on manual feature engineering, it significantly reduces the difficulty of feature modeling and avoids representation bias caused by subjective selection.

[0019] 3. This invention proposes a dynamic threshold construction mechanism based on historical scoring distribution, achieving the effect of enabling the early warning judgment process to adapt to data drift. Traditional methods mostly set static threshold values, which are prone to false alarms or missed alarms. However, this technical solution can automatically adjust the judgment boundary according to changes in the operating environment, thereby solving this deficiency.

[0020] 4. By introducing a response averaging strategy for multiple perturbation samples into the confidence scoring mechanism, this invention enhances the robustness of anomaly detection from the perspective of model stability. Unlike existing methods where single-point scoring is easily affected by random noise, this scheme improves the consistency and repeatability of scoring judgments overall by aggregating multiple offset measures, resulting in more stable performance. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a framework diagram of the data acquisition module of the present invention; Figure 3 This is a framework diagram of the embedded modeling module of the present invention; Figure 4 This is a framework diagram of the anomaly confidence calculation module of the present invention; Figure 5 This is a framework diagram of the early warning judgment module of the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see the appendix Figure 1 -Appendix Figure 5 This invention provides an artificial intelligence-based big data server early warning system, comprising: The data acquisition module is used to collect the operating metrics of the server cluster at multiple points in time and construct them into a time-series metric vector sequence. In the proposed AI-based big data server early warning system, the data acquisition module, as the data input front-end of the entire system, bears the crucial responsibility of acquiring server operating status and constructing a temporal structure. This module directly interacts with the subsequent embedded modeling module, and its output data structure must meet the input requirements of neural network models for temporal continuity and dimensional consistency. Generally, the collected operating indicators should be representative of the server's health status and capable of supporting feature correlation analysis in downstream potential space modeling and anomaly confidence assessment processes.

[0024] In some embodiments, the data acquisition module is deployed in a local monitoring agent of the server cluster or on a centralized edge gateway node, and periodically collects the operating status data of the target server by calling a standardized monitoring interface.

[0025] Specifically, in this embodiment, the performance indicator acquisition unit is used to acquire multi-dimensional performance indicators of each target server in real-time or near real-time. These indicators may include, but are not limited to, CPU utilization, memory usage, disk read / write speed, network inflow / outflow rate, number of active threads, and system load. Each indicator value can be acquired at a uniform sampling period, such as once every 30 seconds or once per minute, to ensure consistency of the time series. Alternatively, a data caching and batch upload mechanism can be introduced to alleviate the upload bandwidth pressure on edge devices.

[0026] The time-series construction unit then aggregates the collected operational metrics in chronological order. Typically, this process uses a sliding window mechanism to prune the metric sequence. Metrics within each time window constitute a vector sample, and multiple consecutive samples form the time series of the metrics. For example, in a typical setup, the system can set the window length to 10, resulting in each sample sequence containing a one-dimensional vector sequence of operational metrics from the past 10 time steps.

[0027] In one possible implementation, the system organizes these indicator data using a vectorized data structure. The time-series indicators for the entire time window form a matrix: ; in, This represents the sample matrix of indicator sequences within the current time window. Indicates the window length. Indicates the current time step. Indicates the first A time-step index vector containing all monitoring index values ​​collected at that time. This matrix will serve as input to the subsequent neural network embedding modeling module.

[0028] Furthermore, in some embodiments, to improve the interpretability of the inter-indicator structure during the modeling phase, the system can construct an undirected weighted graph structure based on the correlation between indicators. Specifically, the indicator graph construction unit calculates the correlation between each pair of indicators and uses this as the edge weight to construct the graph structure. In a specific implementation, the Pearson correlation coefficient is used. Calculate the first The first indicator and the first The correlation between the indicators is defined as follows: ; in, Indicates the first The data collected at the [number]th time point Individual indicator values, This represents the mean of the indicator within the sampling window. This represents the total number of sampling time steps. Indicates the first The data collected at the [number]th time point The values ​​of each indicator, Indicates the first The average value of each indicator within the current sample time window. This represents the total number of time steps within the sample time window, i.e., the number of time points contained in the sample.

[0029] In addition, to improve the stability and robustness of the data, some embodiments will perform normalization or standardization on the original index sequence. Generally, z-score standardization can be used to ensure that each index sequence has a mean of 0 and a standard deviation of 1, which is beneficial for the stable propagation of gradients during the subsequent modeling process of the neural network.

[0030] In some highly scalable application scenarios, the data acquisition module can also support a synchronization and alignment mechanism between multiple nodes to enable collaborative analysis between servers. This mechanism achieves consistency in indicator collection time through unified timestamps and global time synchronization (such as using NTP), ensuring that modeling errors are not caused by time misalignment when constructing time series and correlation graph structures.

[0031] An embedding modeling module is used to receive the time-series index vector sequence and input it into a trained neural network model for low-dimensional latent space embedding. The neural network model includes an encoder and a decoder based on a variational autoencoder structure. In the AI-based big data server early warning system of this invention, the embedded modeling module is a crucial intermediate layer connecting data acquisition and anomaly identification. It is responsible for transforming the acquired high-dimensional time-series indicator vector sequence into a low-dimensional latent space representation to support subsequent quantitative modeling of anomalous behavior. This module directly receives the time-series indicator sequence constructed by the data acquisition module and performs representation learning through a trained neural network model. Generally, to ensure the interpretability and continuity of the latent space's representation distribution, the system employs a neural network model based on a Variational Autoencoder (VAE) structure for embedded modeling.

[0032] In this embodiment, the embedded modeling module includes an encoder unit, a decoder unit, and a graph regularization unit. The input to this module is a preprocessed index time series matrix, and the output is a fixed-dimensional latent space vector, whose dimension is much smaller than the original input space.

[0033] Specifically, the encoder unit first receives a sequence of time-series index vectors as input. The encoder extracts features from the input through a multi-layer neural network and outputs the distribution parameters of the latent variables. To make the model training differentiable, the system introduces a reparameterization trick during modeling, defining the sampling process of the latent variables as: ; in, For the latent representation vector, It is a random vector that follows a standard normal distribution. This represents element-wise multiplication. The standard deviation vector of the latent variables, also output by the encoder network, controls the scale of the latent space distribution. The mean vector of the latent variables is calculated by the encoder network based on the input samples and controls the location center of the latent spatial distribution.

[0034] In one possible implementation, both the encoder and decoder employ convolutional neural networks (CNNs) or gated recurrent units (GRUs) as their basic building blocks to enhance their ability to model temporal context. For example, when the input is a time-series index matrix, key change features within a local time window can be extracted through one-dimensional convolution, and then mapped to the distributed parameter space through a fully connected layer.

[0035] The decoder unit is used to process the latent variables obtained from the above sampling. Reconstruct it into a sequence of index vectors with the same dimensions as the input, and then combine it with the original input. They jointly participate in the calculation of reconstruction error. Reconstruction error is often measured using mean squared error (MSE), i.e.: ; in, This represents the sample matrix of indicator sequences within the current time window. Represents the reconstruction loss, used to measure the degree of difference between the original input and the model output. The representation is composed of latent representations The reconstructed sample matrix generated by the decoder network has dimensions of... same.

[0036] In this embodiment, to enhance the model's ability to model the index structure at the latent space level, a graph regularization unit is introduced. This unit applies graph structure constraints to the latent representation based on the undirected weighted graph of indices constructed in the data acquisition module. Specifically, it is used to apply Laplacian regularization to the latent space representation based on the index correlation graph during the embedding learning process, optimizing the loss function, which includes reconstruction error, KL divergence term, and graph regularization term, wherein the graph regularization term is: ; in, For Laplace matrix, The trace of a matrix is ​​the sum of the elements on its diagonal. Representation matrix transpose, This represents the latent representation matrix.

[0037] The overall loss function consists of three parts: ; in, Represents the reconstruction loss, used to measure the degree of difference between the original input and the model output. This is the KL divergence term for the latent variables, measuring the degree of deviation between the latent distribution and the standard normal distribution. and These are the weight coefficients for the KL term and the graph regularization term, respectively. This is the overall loss function.

[0038] In some embodiments, to improve system training efficiency, the latent space dimensions can be debugged layer by layer, and structural optimization techniques such as BatchNormalization and Dropout can be introduced to alleviate overfitting. In actual deployment, this module can complete parameter learning during the training phase, and only the forward inference path needs to be retained during the deployment phase to efficiently embed newly collected data.

[0039] As an alternative, latent space representation can also be used to construct dynamic trajectory analysis of indicator change trends, thereby assisting in the sensitivity assessment and perturbation robustness calculation of downstream modules.

[0040] The embedded modeling module not only realizes the mapping from the original indicator space to the low-dimensional abstract space, but also provides the system with a foundation of potential variables to characterize the server's operating status, ensuring that the entire early warning system has the ability to characterize, reconstruct, and reason.

[0041] An anomaly confidence calculation module is used to receive the latent space representation output by the embedded modeling module, and calculate the corresponding anomaly confidence score based on the encoding offset between multiple perturbation samples and the original input in the latent space. In the AI-based big data server early warning system proposed in this invention, the anomaly confidence calculation module, as a direct follow-up to the embedded modeling module, plays a crucial role in constructing an anomaly confidence scoring mechanism for server states based on perturbation responses in the latent space. This module assesses the sensitivity of input samples to minor perturbations, quantifies the stability of their latent space representation, and thus provides robustness analysis metrics for anomaly detection. Generally, this module does not directly model the original time-series data but instead fully utilizes the low-dimensional latent representation output from the embedded modeling stage for analysis, thereby achieving efficient and structured anomaly assessment.

[0042] In this embodiment, the anomaly confidence calculation module includes a perturbation generation unit, an encoding offset calculation unit, and a confidence scoring unit, corresponding to the construction of perturbation samples, the extraction of latent space offsets, and the generation of confidence scores, respectively. The input of this module is the latent space representation vector of the original sample output by the embedding modeling module, and it also receives multiple derived samples generated by adding small perturbations to the original input.

[0043] Specifically, in the perturbation generation unit, the system constructs a perturbation sample set based on the original time-series index sequence. Let the original input index matrix be... The perturbation samples are generated by adding random noise element by element to this matrix, i.e.: ; in, This represents the sample matrix of indicator sequences within the current time window. Indicates the first One perturbation sample, Indicates the first A perturbation noise matrix.

[0044] In one possible implementation, the number of perturbation samples can be dynamically adjusted based on computational resources. Typically, the value ranges from 10 to 100 to balance estimation accuracy and computational efficiency. The perturbation only affects the index matrix and does not affect the original model parameters.

[0045] Subsequently, in the encoding offset calculation unit, each perturbation sample and the original input are input into the encoder respectively to extract the corresponding latent spatial representation. Let the original latent representation be denoted as... , No. The potential representation of a perturbation sample is as follows: The encoding offset is calculated. and The Euclidean distance is obtained as follows: ; in, Indicates the first The latent space representation vector of each perturbed sample after being encoded. To record the origin, Indicates the first The encoding offset of a perturbed sample in the latent space, i.e., the Euclidean distance between the perturbed sample and the latent representation of the original sample.

[0046] Alternatively, other distance metrics, such as cosine distance, can be introduced, but in this invention, Euclidean distance is mainly used as the basis for offset measurement to simply and stably reflect the response changes after disturbance.

[0047] In the confidence scoring unit, the system aggregates the offset values ​​corresponding to all perturbed samples. Typically, the aggregation method involves calculating the average of all offset values ​​to obtain the anomaly confidence score for that sample. ; in, The anomaly confidence score is assigned to the current input index sequence, reflecting its sensitivity to small perturbations in the latent space. Indicates the first The encoding offset of a perturbed sample in the latent space, i.e., the Euclidean distance between the perturbed sample and the latent representation of the original sample. This represents the total number of perturbation samples generated, that is, the number of perturbation samples constructed based on the current input samples.

[0048] In some embodiments, the system may normalize the score for use by the subsequent threshold comparison module. Meanwhile, to suppress the influence of extreme perturbation samples on the score, robust statistical strategies such as median or truncated average may be employed during the scoring process.

[0049] As a technical extension, the system can further record the distribution information of disturbance offsets to construct a sample uncertainty map, thereby improving the transparency and interpretability of the early warning model.

[0050] The early warning judgment module is used to receive the abnormal confidence score, construct a dynamic threshold based on the statistical information of historical confidence scores, and compare the current confidence score with the dynamic threshold to determine whether to generate a server abnormality early warning signal. In the proposed server early warning system based on latent spatial embedding and perturbation response analysis, the early warning judgment module, as a downstream functional unit of the anomaly confidence scoring module, is responsible for processing the confidence value calculated by the encoding offset and constructing a dynamic threshold mechanism based on historical scoring data to achieve adaptive anomaly identification of the server status. This module not only completes the final classification judgment of the current input sample status but also provides decision support for the generation of early warning signals, making it a key link in realizing the online early warning function of the entire system.

[0051] In this embodiment, the early warning judgment module mainly includes a scoring cache unit, a dynamic threshold construction unit, and an early warning judgment unit, which are used to maintain the historical confidence score sequence, calculate the statistical discrimination boundary in real time, and perform state classification judgment, respectively.

[0052] Specifically, the scoring cache unit is used to store outlier confidence scores within a historical time window. Let the length of this time window be... Then at any time This module maintains a scoring sequence: ; in, For the rating sequence, Indicates the first Anomaly confidence scores at each time point This is the length of the time window. Generally, The settings can be configured according to the server's operating cycle, for example, selecting scoring data within the last 60 minutes to ensure the representativeness and timeliness of the statistical sample.

[0053] The dynamic threshold construction unit applies the above scoring sequence Perform statistical modeling. A common method is to calculate the mean and standard deviation, and then construct the confidence interval boundaries based on these. Definition: ; in, The time window length, Indicates the current time point The moving average of the rating sequence is the average rating within the time window. Indicates the current time point The moving standard deviation of the rating series is used to measure the degree of fluctuation in ratings within a time window. Indicates a point in time The abnormal confidence score.

[0054] In one possible implementation, the dynamic threshold calculation formula is: ; in, The standard deviation vector of the latent variables, also output by the encoder network, controls the scale of the latent space distribution. This represents the mean vector of the latent variables. This is the sensitivity coefficient. This indicates a dynamic threshold.

[0055] In some embodiments, to improve the stability of the early warning mechanism, the system may further introduce mechanisms such as moving average smoothing and distribution drift adaptation. For example, an exponentially weighted average can be introduced into the scoring sequence: ; in, This represents the smoothing coefficient, used to enhance the responsiveness to recent fluctuations. Indicates the current time point based on The calculated smoothed mean, Indicates the previous moment of Smoothed mean.

[0056] Alternatively, the system can also set multi-level judgment thresholds for servers with different behavior types. For example, for online service nodes with high real-time requirements.

[0057] Furthermore, after an early warning signal is generated, the system can also link with historical scoring patterns to determine whether there is a persistent anomaly. For example, if multiple early warnings are triggered consecutively within a certain time window, it can be inferred that the current server is in a state of continuous degradation, automatically adjusting the response level or triggering manual intervention.

[0058] The dynamic threshold mechanism in the early warning judgment module constructs a discrimination method that adapts to the server operating environment. Compared with a fixed threshold, this mechanism can effectively cope with fluctuations in the scoring baseline over different time periods, thus providing a more discriminative basis for anomaly detection.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A big data server early warning system based on artificial intelligence, characterized in that, include: The data acquisition module is used to collect the operating metrics of the server cluster at multiple points in time and construct them into a time-series metric vector sequence. An embedding modeling module is used to receive the time-series index vector sequence and input it into a trained neural network model for low-dimensional latent space embedding. The neural network model includes an encoder and a decoder based on a variational autoencoder structure. An anomaly confidence calculation module is used to receive the latent space representation output by the embedded modeling module, and calculate the corresponding anomaly confidence score based on the encoding offset between multiple perturbation samples and the original input in the latent space. The early warning judgment module is used to receive the abnormal confidence score, construct a dynamic threshold based on the statistical information of historical confidence scores, and compare the current confidence score with the dynamic threshold to determine whether to generate a server abnormality early warning signal.

2. The big data server early warning system based on artificial intelligence according to claim 1, characterized in that, The data acquisition module includes: The performance metrics collection unit is used to collect multi-dimensional performance metrics from the server cluster, including CPU utilization, memory usage, disk speed, and network traffic. The time-series construction unit is used to organize the running indicators into indicator vectors in chronological order, forming a continuous time-series indicator vector sequence. The indicator graph construction unit is used to construct an undirected weighted graph based on the correlation calculation results between the indicators in the indicator vector, wherein the correlation calculation uses the Pearson correlation coefficient and the edge weight is the absolute value of the correlation coefficient.

3. The artificial intelligence-based big data server early warning system according to claim 1, characterized in that, The embedded modeling module includes: The encoder unit is used to input and encode a time-series index vector sequence into a low-dimensional latent space representation. The encoder is based on a variational autoencoder structure and outputs latent space distribution parameters. Decoder unit, used to receive the latent space representation and reconstruct the index vector; The graph regularization unit is used to apply Laplacian regularization to the latent space representation based on the index correlation graph during the embedding learning process. The optimized loss function includes reconstruction error, KL divergence term, and graph regularization term, where the graph regularization term is: ； in, For Laplace matrix, The trace of a matrix is ​​the sum of the elements on its diagonal. Representation matrix transpose, This represents the latent representation matrix.

4. The artificial intelligence-based big data server early warning system according to claim 1, characterized in that, The anomaly confidence calculation module includes: The perturbation generation unit is used to generate multiple perturbation samples based on the original time series index vector. The perturbation samples are obtained by adding Gaussian noise with zero mean and a set variance. The encoding offset calculation unit is used to input the perturbation sample and the original input to the encoder unit respectively, obtain the corresponding latent space representation, and calculate the offset between the perturbation sample and the latent space representation of the original sample; The confidence score unit is used to calculate the average of the encoding offset results of all perturbation samples and generate an anomaly confidence score, which is used to reflect the sensitivity of the sample to the perturbation.

5. The big data server early warning system based on artificial intelligence according to claim 1, characterized in that, The early warning judgment module includes: The dynamic threshold calculation unit is used to calculate the mean and standard deviation of historical anomaly confidence scores within a preset window, and then calculate the dynamic threshold. The dynamic threshold calculation formula is as follows: ； in, The standard deviation vector of the latent variables, also output by the encoder network, controls the scale of the latent space distribution. This represents the mean vector of the latent variables. This is the sensitivity coefficient. Indicates a dynamic threshold; The anomaly detection unit is used to compare the current anomaly confidence score with a dynamic threshold. When the anomaly confidence score is greater than the dynamic threshold, a server anomaly warning signal is output.

6. The artificial intelligence-based big data server early warning system according to claim 2, characterized in that, The indicator chart construction unit includes: The correlation calculation unit is used to perform Pearson correlation calculation between any two indicators in the time series indicator vector sequence to obtain the correlation coefficient matrix. The weighted graph generation unit is used to construct an undirected weighted graph by using the absolute value of the correlation coefficient matrix as the edge weight. Each node in the graph corresponds to an indicator dimension, and each edge represents the correlation strength between indicators.

7. The artificial intelligence-based big data server early warning system according to claim 3, characterized in that, The encoder unit includes: The parameter extraction unit is used to extract the mean vector and variance vector of the input vector, which are used to construct the distribution of the latent variables. The reparameter sampling unit is used to sample the latent variables based on the mean and variance using the reparameter technique to achieve a differentiable training process. The embedded output unit is used to output the sampled latent variables as a latent space representation to the decoder unit and subsequent modules.

8. The artificial intelligence-based big data server early warning system according to claim 4, characterized in that, The encoding offset calculation unit includes: The original encoding generation unit is used to input the original input index vector into the encoder unit to obtain its corresponding latent space representation; The perturbation coding generation unit is used to input multiple perturbation samples into the encoder to generate corresponding perturbation latent representations. The offset calculation unit is used to calculate the Euclidean distance between each perturbation potential representation and the original potential representation, and stores all distance values ​​for subsequent scoring.

9. The artificial intelligence-based big data server early warning system according to claim 4, characterized in that, The confidence score unit includes: The offset aggregation unit is used to receive the potential representation offset of each perturbation sample and calculate its average value. The confidence generation unit is used to output the average offset value as the abnormal confidence score of the input vector, which reflects the sensitivity of the input to disturbances.

10. The artificial intelligence-based big data server early warning system according to claim 5, characterized in that, The anomaly detection unit includes: The threshold comparison unit is used to compare the current anomaly confidence score with the dynamic threshold. The early warning signal output unit is used to generate an abnormal alarm signal when the current confidence score is greater than the dynamic threshold, and send the signal to the system operation and maintenance platform for users to receive and handle.

Images