A machine learning based cloud node performance prediction method and system

Abstract

The application relates to the technical field of cloud computing and machine learning, and particularly discloses a cloud node performance prediction method and system based on machine learning, which comprises the following steps: collecting multi-source performance data of a cloud node, including system layer indexes, application layer indexes and infrastructure layer indexes, performing timestamp alignment, missing value interpolation and abnormal value marking, and generating multi-dimensional performance time sequence data; adopting a causal discovery method combining Granger causality test and a PCMCI algorithm; constructing a causal relationship graph by combining the Granger causality test and the PCMCI algorithm, performing feature selection based on the causal relationship, eliminating false correlation features from the root, proposing a multi-dimensional composite performance scoring mechanism, dynamically distributing index weights based on causal centrality, and introducing a causal interaction effect term to capture the synergistic and antagonistic effects between indexes, so that the generalization ability of the model in a distribution drift scenario is significantly improved.

Description

Technical Field

[0001] This invention relates to the fields of cloud computing and machine learning technology, and in particular to a method and system for predicting cloud node performance based on machine learning. Background Technology

[0002] With the rapid development of the cloud computing industry and the continuous expansion of cloud data center scale, cloud nodes, as the basic computing units of cloud computing infrastructure, directly affect the quality of upper-layer business services and user experience. Accurately predicting the performance change trend of cloud nodes has important engineering value for resource elastic scheduling, fault prevention and early warning, capacity planning and cost optimization.

[0003] Existing machine learning-based cloud node performance prediction solutions have the following technical shortcomings: First, single-dimensional prediction has limitations. It only models a single performance indicator independently, which cannot capture the complex correlation and coupling effects between indicators and makes it difficult to reflect the true overall performance of the node.

[0004] Secondly, the model has insufficient adaptability. It uses a unified model to predict all heterogeneous nodes, ignoring the differences in hardware, load, etc. between nodes, which leads to large fluctuations in prediction accuracy.

[0005] Then, there is a cold start problem. Newly deployed nodes lack historical data and need to accumulate data over a long period of time to achieve usable prediction accuracy, thus creating a prediction blind spot.

[0006] In addition, the lack of interpretability means that most deep learning prediction models are black boxes, making it difficult for operations and maintenance personnel to understand the basis for the prediction results and thus difficult to build trust.

[0007] Finally, ignoring causal relationships and modeling based on statistical correlation can easily introduce spurious correlation features, resulting in insufficient generalization ability of the model when faced with data distribution drift.

[0008] Therefore, a cloud node performance prediction method and system based on machine learning is proposed. Summary of the Invention

[0009] The purpose of this invention is to provide a cloud node performance prediction method and system based on machine learning to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a cloud node performance prediction method and system based on machine learning, wherein the cloud node performance prediction method based on machine learning includes the following steps: S1: Collect multi-source performance data of cloud nodes, including system layer indicators, application layer indicators and infrastructure layer indicators, perform timestamp alignment, missing value imputation and outlier marking, and generate multi-dimensional performance time series data; S2: A causal discovery method combining Granger causality test and PCMCI algorithm is adopted to explore the causal relationships between various performance indicators and construct a directed acyclic causal relationship graph; S3: Construct a heterogeneous feature profile for each cloud node, including static and dynamic features, and generate a node feature vector through encoding and fusion; S4: Construct a model pool containing multiple heterogeneous prediction models, build a feature input set for causal perception based on the causal relationship graph, and train the meta-learner to dynamically select the optimal combination of prediction models and hyperparameter configuration according to the node profile; S5: For a new node, calculate its feature similarity with existing nodes, select the source node with the highest similarity for cross-node transfer learning, and use a progressive transfer weight mechanism to achieve a smooth transition. S6: Calculate the causal centrality weights of the indicators based on the causal relationship graph, define a multidimensional composite performance score, integrate the predicted values ​​of each indicator and the interaction effects between indicators, and output the prediction results and an interpretability report based on the causal path. S7: Continuously monitors prediction errors and triggers cascaded adaptive updates of the causal relationship graph, prediction model, and node profile when data distribution drift is detected.

[0011] Preferably, the system-level metrics in S1 include at least one of CPU utilization, memory utilization, disk I / O throughput, network bandwidth utilization, and average system load. The application layer metrics include at least one of request response time, transaction throughput, error rate, and queue depth; The infrastructure layer metrics include at least one of node temperature, power consumption, fan speed, and storage health.

[0012] Preferably, the causal discovery method described in S2 includes: ADF stationarity test was performed on the time series data of each indicator, and differencing was performed on non-stationary series. Granger causality test was used to screen candidate causal pairs; The PCMCI algorithm is used to remove spurious causal relationships within the conditional independence framework. Based on conditional mutual information, causal strength weights are assigned to directed edges in the causal graph.

[0013] Preferably, the static features described in S3 include at least one of hardware configuration, operating system, and deployment location; The dynamic characteristics include at least one of load type distribution, performance statistics, and fluctuation cycle characteristics; Static and dynamic features are encoded and fused using a multilayer perceptron.

[0014] Preferably, the model pool in S4 includes at least two of LSTM, Attention-LSTM, Informer, TCN, ARIMA, Prophet, and LightGBM; The feature input set for causal perception is constructed based on the causal parent nodes and ancestor nodes of the target index in the causal relationship graph.

[0015] Preferably, the meta-learner in S5 is an XGBoost-based classifier, which is trained with node feature vectors as input and historical prediction accuracy of each model as supervision signal; the prediction results of the recommendation model are weighted and integrated according to the adaptation score to obtain the final prediction value.

[0016] Preferably, the cross-node transfer learning described in S5 includes: Calculate the cosine similarity between the feature vectors of the new node and the existing nodes, and select the K nodes with the highest similarity as source nodes, where K is a preset positive integer; use the similarity-weighted average of the source node model parameters as initialization parameters and fine-tune them. The progressive migration weights are: ; in, , , To accumulate days for data, The attenuation coefficient is the final predicted value. ; in, This represents the final fused prediction value of cloud node performance metrics. These are migration prediction values ​​obtained based on migrations from similar source nodes. This refers to locally fine-tuned predictions obtained by training with a small amount of local data from the new node.

[0017] Preferably, the calculation formula for the multidimensional composite performance score in S6 is as follows: ; in, For multidimensional composite performance scoring, The total number of performance indicators participating in the overall score. As an indicator Causal centrality weights For the first One original performance indicator, This is the normalized value of the indicator's predicted value. For the first The predicted values ​​of each performance indicator after normalization. This refers to the causal interaction effect term between indicators; ; in, To represent the set of directed edges in a causal graph Each causal edge in the middle, Let be the set of all directed causal edges in the causal graph. To the indicators Indicator Causal strength weights and Indicators and Normalized predicted values.

[0018] Preferably, the interpretability described in S6 includes: The Top-K causal paths that contribute most to changes in CPS and their contribution rates, current and predicted values ​​of each path link, and root cause analysis recommendations based on causal paths; S7 uses the ADWIN adaptive window algorithm to detect data distribution drift. When drift is detected, the following steps are executed sequentially: First, update the causal relationship graph, then update the prediction model, and finally update the node profile. Record the update time, the reason for the update, and the prediction accuracy before and after each update.

[0019] A machine learning-based cloud node performance prediction system includes a data acquisition module, a causal discovery module, a node profiling module, an adaptive prediction module, a transfer learning module, a scoring and output module, and an adaptive update module. The data acquisition module is used to collect multi-source performance data of cloud nodes and perform alignment, interpolation, and outlier marking. The causality discovery module is used to construct a directed acyclic causal graph using Granger causality tests and the PCMCI algorithm. The node profiling module is used to construct heterogeneous feature profiles of cloud nodes and generate node feature vectors. The adaptive prediction module includes a model pool and a meta-learner. It dynamically selects the optimal prediction model based on the node profile and constructs a feature input set for causal perception based on the causal relationship graph. The transfer learning module is used to build predictive capabilities for new nodes through cross-node transfer learning, and adopts a progressive transfer weight mechanism to achieve a smooth transition. The scoring and output module is used to calculate multidimensional composite performance scores, output prediction results and interpretability reports based on causal paths; The adaptive update module is used to monitor data distribution drift through the ADWIN algorithm and trigger cascaded adaptive updates of the causal relationship graph, prediction model, and node profile.

[0020] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs a causal relationship graph by combining Granger causality test and PCMCI algorithm, selects features based on causal relationships to eliminate spurious correlation features from the root, proposes a multi-dimensional composite performance scoring mechanism, dynamically allocates index weights based on causal centrality, and introduces causal interaction effect terms to capture the synergistic and antagonistic effects between indicators, significantly improving the model's generalization ability in distribution drift scenarios.

[0021] 2. This invention constructs heterogeneous feature profiles of nodes, trains a meta-learner to dynamically select the optimal combination of prediction models based on the profiles, and achieves accurate predictions that vary from node to node. It designs a cross-node transfer learning strategy based on profile similarity and a progressive transfer weight mechanism, so that new nodes can achieve usable prediction accuracy with only a very small amount of data.

[0022] 3. This invention utilizes causal relationship graphs to automatically generate causal path-level interpretability reports, and uses the ADWIN algorithm to detect data distribution drift in real time, automatically triggering full-link cascade updates to ensure that the prediction system always adapts to the latest data distribution. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Fig. 1 This is a flowchart of the method of the present invention; Fig. 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0025] The technical solutions of 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.

[0026] Please see Figs. 1-2 The present invention provides a technical solution: Example 1: A cloud node performance prediction method based on machine learning. The overall process of the prediction method is as follows: collect multi-source performance data and perform alignment processing; after completing the data processing, construct a causal relationship graph and node profile to drive the selection and training of an adaptive prediction model; then handle the cold start through cross-node transfer learning; subsequently calculate a multi-dimensional composite performance score and output an interpretability report; finally, continuously monitor data distribution drift and trigger adaptive model updates.

[0027] Step 1: Multi-source heterogeneous performance data acquisition and game simulation By deploying a lightweight data collection agent on each cloud node, performance metrics are collected from three levels at a set frequency of once every ten seconds.

[0028] System-level metrics are obtained by reading the proc file system, and specifically include CPU utilization, memory utilization, disk I / O throughput, network bandwidth utilization, and average system load.

[0029] CPU utilization is categorized into user mode, system mode, and I / O wait mode. Memory utilization includes detailed data on cache and buffer. Disk I / O throughput includes read / write speed, IOPS, and queue depth.

[0030] Network bandwidth utilization is differentiated between inbound and outbound data, and the average system load includes three dimensions: one minute, five minutes, and fifteen minutes.

[0031] Application layer metrics are obtained through application probes, including request response time, transaction throughput, error rate, and message queue depth. Request response time includes P50, P90, and P99 percentile values, and transaction throughput includes TPS and QPS values.

[0032] Infrastructure layer metrics are obtained through the IPMI and BMC interfaces, specifically including node CPU temperature, overall power consumption, fan speed, and storage device health. Storage device health is characterized using the SMART metric.

[0033] Data preprocessing operations include timestamp alignment, missing value imputation, and outlier labeling. Timestamps are uniformly aligned to millisecond-level UTC timestamps. Missing values ​​are handled using linear imputation. Segments with more than five consecutive missing times are filled with nearest-neighbor historical values. Outliers are labeled using the three-sigma criterion. After preprocessing, a multidimensional performance time-series data matrix D is output.

[0034] Step 2: Automatic Construction of Causal Relationship Graph The construction of a causal relationship graph consists of four sub-steps.

[0035] The first step is stationarity processing. The ADF test is used to perform a stationarity test on the time series data. For non-stationary series, first-order differencing is performed until the series passes the stationarity test. The maximum order of differencing is set to second.

[0036] The second step is Granger causality screening. All indicator combinations are iterated through, and restricted models using only historical values ​​of a single indicator and unrestricted models using historical values ​​of two indicators are constructed. The goodness of fit of the two types of models is compared using the F-test, and candidate causal relationship pairs with a significance level less than 0.05 are retained.

[0037] The third step is PCMCI refinement. Under the conditional independence framework, it examines whether candidate causal pairs remain conditionally dependent given all other indicators, thus eliminating spurious causal associations.

[0038] The fourth step is to construct the causal graph. Using performance metrics as nodes and causal relationships as directed edges, a directed acyclic causal graph is constructed. Causal strength weights are assigned to the directed edges based on conditional mutual information. A valid update window is set for the causal graph, with a window duration of seven days, and updates are performed at fixed intervals.

[0039] Step 3: Node Heterogeneous Feature Profiling Components The construction of heterogeneous feature profiles for nodes involves three steps.

[0040] The first step is static feature extraction. Static features include hardware configuration, operating system, and deployment location. Hardware configuration includes CPU model benchmark score, number of cores, memory capacity, and disk type. Operating system includes system type and version number. Deployment location includes availability zone number and rack location number. Disk type is represented using one-hot encoding.

[0041] The second step is dynamic feature extraction, which uses performance data from the most recent 24 hours as input to calculate load type distribution, performance statistics, and fluctuation period characteristics. The load type distribution includes the percentage of compute-intensive, I / O-intensive, and network-intensive loads. The performance statistics include mean, variance, kurtosis, and skewness. The fluctuation period characteristics are extracted using the FFT method to obtain the main period and corresponding amplitude.

[0042] The third step is image fusion, which concatenates static and dynamic feature vectors and performs encoding fusion through a two-layer multilayer perceptron. The first layer has a dimension of 256, and the second layer has a dimension of 128. The activation function is the ReLU function, and the output after encoding fusion is a 128-dimensional node feature vector.

[0043] Step 4: Construction and Dynamic Selection of Adaptive Predictive Model Pool The construction and dynamic selection of the adaptive prediction model pool involves four steps.

[0044] The first step is to build the model pool, which contains seven pre-set prediction models: LSTM with 128 hidden units, Attention-LSTM, Informer, TCN with eight layers, ARIMA with automatic order determination, Prophet, and LightGBM.

[0045] The second step is causal perception feature selection. Based on the causal relationship graph, the time series data corresponding to the causal parent node and ancestor node of the target indicator are selected and combined with the historical data of the target indicator itself to form the model input set.

[0046] The third step is the construction and training of the meta-learner. The meta-learner uses the XGBoost model, taking node feature vectors as input and historical MAPE values ​​of each model as supervision signals for training. During the inference phase, the adaptation scores of each model are output, and the top three models are selected for hyperparameter configuration through Bayesian optimization.

[0047] The fourth step is ensemble prediction, which involves performing weighted ensemble calculations on the selected models according to their fit scores to obtain the final prediction results.

[0048] The formula is: Step 5: Cross-node transfer learning and cold start handling Collect initial performance data of the new node for no less than 24 hours, construct an initial feature vector based on the data, calculate the cosine similarity between the feature vector of the new node and the feature vectors of all existing nodes, and select the three nodes with the highest similarity as source nodes.

[0049] The source node model parameters are calculated by weighted average based on similarity. The result is used as the initialization parameters for the new node model. Fine-tuning is performed using a small amount of data from the new node, with the fine-tuning learning rate set to one-tenth of the original model's learning rate.

[0050] A gradual migration weighting method is used to achieve a smooth transition in the prediction results. As the data accumulation time of the new nodes increases, the migration weights gradually decay, and the final prediction result is obtained by weighted fusion of the migration prediction result and the local fine-tuning prediction result. The time for the weights to decay to near zero is approximately ninety-two days.

[0051] Step Six: Multidimensional Compliance Performance Scoring and Predictive Output First, the causal centrality weight of the indicator is calculated. The causal centrality weight is determined by the ratio of the sum of the causal strength weights of the corresponding child nodes of the indicator to the sum of the causal strength weights of all directed edges. Normalization is performed on each indicator, with positive indicators being normalized positively and negative indicators being normalized negatively.

[0052] The multidimensional composite performance score is obtained by weighted summation of the independent contribution terms and the interaction effect terms of the indicators. The interaction effect term is obtained by summing the product of the causal strength weights corresponding to the directed edges in the causal relationship graph and the normalized predicted values ​​of the two indicators.

[0053] The interaction effect coefficient was determined within a preset range through cross-validation and was used to balance the proportion of independent contributions and interaction contributions in the total score.

[0054] An interpretability report is generated based on the prediction results. The report selects the multiple causal paths that contribute the most to the changes in the multidimensional composite performance score, marks the current and predicted values ​​of each link in the path, and provides root cause analysis suggestions.

[0055] Step 7: Model Adaptive Update and Drift Detection The ADWIN adaptive window algorithm is used to continuously monitor data distribution drift. When distribution drift is detected, cascading update operations are performed in a fixed order: first, the causal relationship graph is updated; second, the prediction model is updated; and finally, the node profile is updated. The system records the time, triggering reason, and change in prediction accuracy before and after each update, and supports version tracing and querying.

[0056] Implementation 2: A cloud node performance prediction system based on machine learning, including a data acquisition module, a causal discovery module, a node profiling module, an adaptive prediction module, a transfer learning module, a scoring and output module, and an adaptive update module; The data acquisition module is used to collect performance data from the system layer, application layer, and infrastructure layer of cloud nodes. It performs timestamp alignment, missing value imputation, and outlier marking on the collected data, and outputs standardized multidimensional performance time-series data.

[0057] The causal discovery module is used to perform stationarity processing on performance time series data. It mines causal relationships between indicators through Granger causality test and PCMCI algorithm, and constructs and maintains a weighted directed acyclic causal relationship graph.

[0058] The node profiling module is used to extract the static and dynamic features of cloud nodes, encode and fuse the two types of features, and generate node feature vectors to represent the heterogeneous attributes of nodes.

[0059] The adaptive prediction module is used to maintain a model pool consisting of multiple prediction models. It selects input features based on the causal relationship graph, matches the optimal model through a meta-learner, and performs weighted integration of the selected models to obtain the performance prediction results.

[0060] The transfer learning module is used to calculate the feature similarity between new nodes and existing nodes, select similar nodes to complete the model parameter transfer, and combine the transfer results with the local fine-tuning results by progressive weight fusion to solve the cold start problem of new node prediction.

[0061] The scoring and output module is used to calculate index weights based on the causal relationship graph, normalize the predicted indexes, calculate multi-dimensional composite performance scores, and generate a prediction interpretability report based on causal paths.

[0062] The adaptive update module is used to monitor data distribution drift in real time. When drift is detected, the causal relationship graph, prediction model and node profile are updated in sequence, the update information is recorded and version traceability is supported.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cloud node performance prediction method based on machine learning, characterized in that: The machine learning-based cloud node performance prediction method includes the following steps: S1: Collect multi-source performance data of cloud nodes, including system layer indicators, application layer indicators and infrastructure layer indicators, perform timestamp alignment, missing value imputation and outlier marking, and generate multi-dimensional performance time series data; S2: A causal discovery method combining Granger causality test and PCMCI algorithm is adopted to explore the causal relationships between various performance indicators and construct a directed acyclic causal relationship graph; S3: Construct a heterogeneous feature profile for each cloud node, including static and dynamic features, and generate a node feature vector through encoding and fusion; S4: Construct a model pool containing multiple heterogeneous prediction models, build a feature input set for causal perception based on the causal relationship graph, and train the meta-learner to dynamically select the optimal combination of prediction models and hyperparameter configuration according to the node profile; S5: For a new node, calculate its feature similarity with existing nodes, select the source node with the highest similarity for cross-node transfer learning, and use a progressive transfer weight mechanism to achieve a smooth transition. S6: Calculate the causal centrality weights of the indicators based on the causal relationship graph, define a multidimensional composite performance score, integrate the predicted values ​​of each indicator and the interaction effects between indicators, and output the prediction results and an interpretability report based on the causal path. S7: Continuously monitors prediction errors and triggers cascaded adaptive updates of the causal relationship graph, prediction model, and node profile when data distribution drift is detected.

2. The cloud node performance prediction method based on machine learning according to claim 1, characterized in that: The system-level metrics mentioned in S1 include at least one of CPU utilization, memory utilization, disk I / O throughput, network bandwidth utilization, and average system load. The application layer metrics include at least one of request response time, transaction throughput, error rate, and queue depth; The infrastructure layer metrics include at least one of node temperature, power consumption, fan speed, and storage health.

3. The cloud node performance prediction method based on machine learning according to claim 1, characterized in that: The causal discovery methods described in S2 include: ADF stationarity test was performed on the time series data of each indicator, and differencing was performed on non-stationary series. Granger causality test was used to screen candidate causal pairs; The PCMCI algorithm is used to remove spurious causal relationships within the conditional independence framework. Based on conditional mutual information, causal strength weights are assigned to directed edges in the causal graph.

4. The cloud node performance prediction method based on machine learning according to claim 1, characterized in that: The static features described in S3 include at least one of hardware configuration, operating system, and deployment location; The dynamic characteristics include at least one of load type distribution, performance statistics, and fluctuation cycle characteristics; Static and dynamic features are encoded and fused using a multilayer perceptron.

5. The cloud node performance prediction method based on machine learning according to claim 1, characterized in that: The model pool described in S4 includes at least two of LSTM, Attention-LSTM, Informer, TCN, ARIMA, Prophet, and LightGBM; The feature input set for causal perception is constructed based on the causal parent nodes and ancestor nodes of the target index in the causal relationship graph.

6. The cloud node performance prediction method based on machine learning according to claim 1, characterized in that: The meta-learner described in S5 is an XGBoost-based classifier that is trained using node feature vectors as input and historical prediction accuracy of each model as supervision signals. The prediction results of the recommendation model are weighted and integrated according to the adaptation score to obtain the final prediction value.

7. The cloud node performance prediction method based on machine learning according to claim 1, characterized in that: Cross-node transfer learning as described in S5 includes: Calculate the cosine similarity between the feature vectors of the new node and the existing nodes, and select the K nodes with the highest similarity as source nodes, where K is a preset positive integer; use the similarity-weighted average of the source node model parameters as initialization parameters and fine-tune them. The progressive migration weights are: ; in, , , To accumulate days for data, The attenuation coefficient is the final predicted value. ; in, This represents the final fused prediction value of cloud node performance metrics. These are migration prediction values ​​obtained based on migrations from similar source nodes. This refers to locally fine-tuned predictions obtained by training with a small amount of local data from the new node.

8. The cloud node performance prediction method based on machine learning according to claim 1, characterized in that: The calculation formula for the multidimensional composite performance score mentioned in S6 is as follows: ； in, For multidimensional composite performance scoring, The total number of performance indicators participating in the overall score. As an indicator Causal centrality weights For the first One original performance indicator, This is the normalized value of the indicator's predicted value. For the first The predicted values ​​of each performance indicator after normalization. This refers to the causal interaction effect term between indicators; ； in, To represent the set of directed edges in a causal graph Each causal edge in the middle, Let be the set of all directed causal edges in the causal graph. To the indicators Indicator Causal strength weights and Indicators and Normalized predicted values.

9. The cloud node performance prediction method based on machine learning according to claim 1, characterized in that: The interpretability described in S6 includes: The Top-K causal paths that contribute most to changes in CPS and their contribution rates, current and predicted values ​​of each path link, and root cause analysis recommendations based on causal paths; S7 uses the ADWIN adaptive window algorithm to detect data distribution drift. When drift is detected, the following steps are executed sequentially: First, update the causal relationship graph, then update the prediction model, and finally update the node profile. Record the update time, the reason for the update, and the prediction accuracy before and after each update.

10. A cloud node performance prediction system based on machine learning, comprising a cloud node performance prediction method based on machine learning according to any one of claims 1-9, characterized in that: The machine learning-based cloud node performance prediction system includes a data acquisition module, a causal discovery module, a node profiling module, an adaptive prediction module, a transfer learning module, a scoring and output module, and an adaptive update module. The data acquisition module is used to collect multi-source performance data of cloud nodes and perform alignment, interpolation, and outlier marking. The causality discovery module is used to construct a directed acyclic causal graph using Granger causality tests and the PCMCI algorithm. The node profiling module is used to construct heterogeneous feature profiles of cloud nodes and generate node feature vectors. The adaptive prediction module includes a model pool and a meta-learner. It dynamically selects the optimal prediction model based on the node profile and constructs a feature input set for causal perception based on the causal relationship graph. The transfer learning module is used to build predictive capabilities for new nodes through cross-node transfer learning, and adopts a progressive transfer weight mechanism to achieve a smooth transition. The scoring and output module is used to calculate multidimensional composite performance scores, output prediction results and interpretability reports based on causal paths; The adaptive update module is used to monitor data distribution drift through the ADWIN algorithm and trigger cascaded adaptive updates of the causal relationship graph, prediction model, and node profile.

Images