Data center energy consumption anomaly detection method and system based on time-frequency domain fusion

Abstract

The application provides a data center energy consumption anomaly detection method and system based on time-frequency domain fusion, relates to the field of data center detection, and comprises the following steps: acquiring multi-source data continuously monitored by a data center; obtaining analysis window detection samples based on the multi-source data and performing feature extraction; analyzing typical spectral structure features through a normal spectrum template; calculating a prediction residual based on a node normal energy consumption baseline model; calculating a spectral consistency index and a main frequency offset based on the typical spectral structure features, and calculating a gating factor and an adaptive anomaly threshold; calculating a gated enhanced residual anomaly degree based on the prediction residual and the gating factor, and generating an adaptive anomaly labeling result; and finally inputting the encoded multi-source data into a model to obtain an anomaly detection score of the data center energy consumption. The application can comprehensively analyze the time domain and frequency domain features of the node energy consumption, thereby improving the identification ability of the node energy consumption anomaly under complex operation scenarios.

Description

Technical Field

[0001] This invention relates to the field of data center detection, and specifically to a method and system for detecting abnormal energy consumption in data centers based on time-frequency domain fusion. Background Technology

[0002] As the core infrastructure of modern information society, data centers undertake massive computing and storage tasks. They contain numerous computing nodes, whose energy consumption constitutes a major part of the total energy consumption of the data center. Real-time monitoring and analysis of node energy consumption data is crucial for ensuring stable data center operation and improving energy efficiency management. Energy consumption data is typically collected and stored in time-series format, and analyzing this data to detect anomalies is a key aspect of intelligent data center operation and maintenance.

[0003] In existing technologies, there are two approaches to anomaly detection in data center energy consumption: one is to establish an energy consumption prediction model, compare the actual energy consumption value with the baseline value predicted by the model, and identify anomalies by analyzing the residuals between the two; the other is to use statistical methods to calculate the statistical confidence intervals of historical energy consumption data at different time periods and determine whether the current energy consumption value exceeds the normal range to detect anomalies.

[0004] However, in complex operating environments, the energy consumption sequence of data center nodes often contains periodic changes, short-term fluctuations, and local structural changes at multiple time scales. The above-mentioned anomaly monitoring methods have relatively limited ability to analyze the distribution pattern of energy consumption sequences in the frequency domain and the time-varying spectral characteristics, making it difficult to accurately identify anomalies in complex operating scenarios. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a data center energy consumption anomaly detection method and system based on time-frequency domain fusion, which solves the problem that existing methods cannot accurately identify anomalies in complex operating scenarios due to incomplete feature analysis.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a data center energy consumption anomaly detection method based on time-frequency domain fusion, comprising: Acquire multi-source data from data center computing nodes; The multi-source data is sampled by a fixed-length sliding window to obtain analysis window detection samples; the analysis window detection samples include the node total power consumption sequence and operating condition sample data. Based on the total power consumption sequence of the nodes, time-domain and frequency-domain features are extracted to obtain time-domain feature data and frequency-domain feature data; prediction residuals are calculated based on the total power consumption sequence of the nodes, operating condition sample data and the node normal energy consumption baseline model. Based on the frequency domain feature data, typical spectral structure features of the node's total power consumption sequence are extracted using a normal spectral template, and spectral consistency index and main frequency offset are calculated based on the typical spectral structure features; the prediction residual is then standardized to obtain the standardized residual. Calculate the gating factor and adaptive anomaly threshold based on the aforementioned spectrum consistency index and dominant frequency offset; The gated enhancement residual anomaly is calculated based on the standardized residual and the gated factor, and an adaptive anomaly labeling result is generated based on the gated enhancement residual anomaly and the adaptive anomaly threshold. The operating condition sample data, sliding window information, computing node identifiers, time-domain feature data, frequency-domain feature data, prediction residuals, spectral consistency index, main frequency offset, gating factor, standardized residuals, gating enhancement residual abnormality, and adaptive anomaly labeling results are uniformly encoded into prompt words and then input into a large language model based on LoRA parameter fine-tuning to obtain an energy consumption anomaly detection score.

[0007] Preferably, the step of calculating the prediction residual based on the node's total power consumption sequence, operating condition sample data, and the node's normal energy consumption baseline model includes: The operating condition sample data is processed by the node normal energy consumption baseline model to obtain the predicted energy consumption value; the node total power consumption sequence is averaged to obtain the actual energy consumption value of the node. The predicted residual is obtained by calculating the residual based on the predicted energy consumption value and the actual energy consumption value of the node.

[0008] Preferably, after acquiring the multi-source data from the data center computing nodes, the method further includes: The multi-source data is preprocessed; the preprocessing includes data alignment based on a unified time axis, missing value imputation based on linear interpolation, and normalization.

[0009] Preferably, the normal spectrum template is constructed as follows: Obtain normal operation samples from the historical operation data of data center computing nodes; Calculate the normalized spectrum of the normally operating sample; A normal spectrum template is constructed based on the normalized spectrum; the expression for the normal spectrum template is: in, Indicates the normal spectral template in the 1st... Average spectral value at each frequency point For the first The normal sample at the ... Normalized spectral values ​​at each frequency point This represents the sample set consisting of all normal samples.

[0010] Preferably, the expression for the main frequency offset is: in, The dominant frequency of the current sample. This is the main frequency of the normal spectrum template. Represents extremely small positive numbers. This represents the maximum dominant frequency of all samples. This represents the minimum dominant frequency among all samples.

[0011] Preferably, the formula for calculating the gated enhancement residual abnormality is: in, For the Sigmoid function, The weights of the spectral inconsistency terms are represented. This indicates the weight of the main frequency offset term. For bias terms, To represent the main frequency offset; As the gating factor, Indicates the spectrum consistency index. Indicates the abnormality of gated enhancement residuals. Represents the standardized residual. This represents the weighting parameter.

[0012] Preferably, the formula for calculating the adaptive anomaly threshold is: in, Indicates the first i Adaptive anomaly threshold for each sample Indicates the basic threshold. Indicates the spectrum consistency index. This indicates the main frequency offset. This represents the adjustment coefficient.

[0013] Secondly, the present invention also provides a data center energy consumption anomaly detection system based on time-frequency domain fusion, comprising: The data acquisition module acquires multi-source data from the data center's computing nodes; The sample construction module constructs samples from the multi-source data using a fixed-length sliding window to obtain analysis window detection samples; the analysis window detection samples include the node total power consumption sequence and operating condition sample data. The first processing module performs time-domain and frequency-domain feature extraction based on the total power consumption sequence of the nodes to obtain time-domain feature data and frequency-domain feature data; and calculates the prediction residual based on the total power consumption sequence of the nodes, operating condition sample data and the node normal energy consumption baseline model. The second processing module extracts typical spectral structure features of the node's total power consumption sequence from the frequency domain feature data using a normal spectral template, and calculates the spectral consistency index and main frequency offset based on the typical spectral structure features; and standardizes the prediction residual to obtain the standardized residual. The third processing module calculates the gating factor and adaptive anomaly threshold based on the spectrum consistency index and the main frequency offset. The fourth processing module calculates the gated enhancement residual anomaly based on the standardized residual and the gate factor, and generates adaptive anomaly labeling results based on the gated enhancement residual anomaly anomaly anomaly threshold. The result prediction module encodes the operating condition sample data, sliding window information, computing node identifiers, time-domain feature data, frequency-domain feature data, prediction residuals, spectrum consistency index, main frequency offset, gating factor, standardized residuals, gating enhancement residual abnormality, and adaptive anomaly labeling results into prompt words, and then inputs them into a large language model based on LoRA parameter fine-tuning to obtain an energy consumption anomaly detection score.

[0014] Thirdly, this application also provides a computer-readable storage medium storing a computer program for a data center energy consumption anomaly detection method based on time-frequency domain fusion, wherein the computer program causes a computer to execute the data center energy consumption anomaly detection method based on time-frequency domain fusion as described above.

[0015] Fourthly, this application also provides an electronic device, comprising: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the data center energy consumption anomaly detection method based on time-frequency domain fusion as described above.

[0016] (III) Beneficial Effects This invention provides a method and system for detecting anomalies in data center energy consumption based on time-frequency domain fusion. Compared with existing technologies, it has the following advantages: 1. This invention first constructs time-series detection samples through a standardized fixed-length sliding window to unify the analysis scale of time-series data; it simultaneously extracts time-domain and frequency-domain features of the samples, and extracts typical spectral structure features based on normal spectrum templates, quantitatively calculates the main frequency offset and spectral consistency index, and can fully capture the dual time-domain and frequency-domain features of energy consumption sequences, such as multi-time-scale periodic changes, short-term fluctuations, and local spectral distortions. It makes full use of the frequency distribution law of energy consumption signals, solves the core defect of the lack of frequency domain feature analysis in existing methods, and provides more comprehensive data basis for anomaly identification under complex working conditions.

[0017] 2. This invention calculates a consistency gating factor based on the spectral consistency index and the main frequency offset, and performs multiplicative enhancement correction on the standardized time-domain prediction residual: when the current sample spectrum has a high matching degree with the normal template (no frequency domain anomaly), the gating factor does not amplify the residual, avoiding false alarms caused by normal periodic fluctuations; when the sample spectrum shows obvious offset or structural distortion, the gating factor amplifies the standardized residual, highlighting the fault signal of frequency domain synchronization anomaly; finally, it generates a gating enhancement residual anomaly degree that fuses time and frequency information, realizing the organic combination of time-domain energy consumption deviation and frequency-domain spectral distortion information, significantly reducing the false alarm rate and false alarm rate of anomaly detection in complex scenarios.

[0018] 3. This invention relies on real-time calculated spectrum consistency index and main frequency offset to dynamically generate adaptive anomaly threshold. The threshold can be adjusted synchronously with the current spectrum distortion degree and energy consumption fluctuation state of the node, which can adapt to the complex operating conditions of the node and make the anomaly judgment boundary fit the real-time operating rules of the equipment, further improving the stability of anomaly identification in complex scenarios. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating the data center energy consumption anomaly detection method provided in an embodiment of the present invention; Figure 2 A schematic diagram of the modules for the data center energy consumption anomaly detection method provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0022] This application provides a data center energy consumption anomaly detection method and system based on time-frequency domain fusion, which solves the problem of incomplete feature analysis in existing methods, achieves accurate identification of abnormal energy consumption and early hidden faults, and significantly reduces the false alarm rate caused by normal business fluctuations.

[0023] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: In actual data center operation, abnormal node energy consumption is often an early sign of decreased cooling system efficiency, uneven resource scheduling, or abnormal local operating conditions. If these anomalies are not effectively identified in the early stages, they may further lead to a decline in system energy efficiency and even affect the stable operation of data center services. Therefore, establishing a detection method that can accurately identify abnormal node energy consumption is of great significance for ensuring the safe and stable operation of data centers and improving energy efficiency management.

[0024] Under complex load conditions in data centers, node energy consumption often simultaneously exhibits periodic variations, sudden fluctuations, and energy consumption structure characteristics at different time scales. Relying solely on time domain features is insufficient to fully characterize the dynamic changes in energy consumption sequences across different time scales. When energy consumption anomalies manifest as local frequency variations, short-term energy consumption mutations, or multi-scale coupled fluctuations, existing methods still fall short in terms of accuracy and stability in anomaly identification. Secondly, in complex operating environments, data center node energy consumption sequences often simultaneously contain periodic variations, short-term fluctuations, and local structural changes across multiple time scales. Relying solely on time domain features is insufficient to fully characterize their frequency structure characteristics. Due to the lack of joint modeling of spectral structure features and temporal local structure features, existing methods still have limitations in their ability to identify abnormal patterns in complex operating scenarios.

[0025] This invention provides a data center energy consumption anomaly detection method and system based on time-frequency domain fusion. By introducing a sliding time window, an energy consumption regression prediction baseline under normal operating conditions, robust residual standardization based on median absolute deviation, and a gating factor based on frequency domain structural stability, an anomaly representation that is coordinated between the time and frequency domains is constructed. This enables dynamic correction and adaptive discrimination of the degree of node energy consumption anomalies, thereby improving the ability to identify node energy consumption anomalies in complex operating scenarios, reducing false alarms caused by normal business fluctuations, and improving the detection sensitivity of early hidden anomalies such as equipment degradation and reduced heat dissipation capacity.

[0026] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0027] like Figure 1 As shown, this embodiment of the invention provides a data center energy consumption anomaly detection method based on time-frequency domain fusion, including: Step S110: Obtain multi-source data from the data center computing nodes; Step S120: Samples are constructed from the multi-source data using a fixed-length sliding window to obtain analysis window detection samples; the analysis window detection samples include the node total power consumption sequence and operating condition sample data; Step S130: Based on the total power consumption sequence of the node, perform time-domain and frequency-domain feature extraction to obtain time-domain feature data and frequency-domain feature data; calculate the prediction residual based on the total power consumption sequence of the node, the operating condition sample data and the node normal energy consumption baseline model; Step S140: Based on the frequency domain feature data, extract the typical spectral structure features of the node total power consumption sequence through a normal spectral template, and calculate the spectral consistency index and main frequency offset based on the typical spectral structure features; standardize the prediction residual to obtain the standardized residual; Step S150: Calculate the gating factor and adaptive anomaly threshold based on the spectrum consistency index and the main frequency offset; Step S160: Calculate the gated enhancement residual anomaly based on the standardized residual and the gate factor, and generate adaptive anomaly labeling results based on the gated enhancement residual anomaly anomaly anomaly threshold; Step S170: The operating condition sample data, sliding window information, computing node identifier, time domain feature data, frequency domain feature data, prediction residual, spectrum consistency index, main frequency offset, gating factor, standardized residual, gating enhancement residual abnormality, and adaptive anomaly labeling results are uniformly encoded into prompt words and then input into the large language model based on LoRA parameter fine-tuning to obtain the energy consumption anomaly detection score.

[0028] This invention first collects time-series energy consumption data of data center nodes, along with environmental and system parameters related to their operational status. A sliding time window mechanism is used to construct samples from the raw data. Based on this, a regression model is used to establish an energy consumption prediction model for nodes under normal operating conditions, thereby obtaining the corresponding normal energy consumption baseline. Subsequently, the residual sequence between actual observed energy consumption and predicted energy consumption is calculated, and robust standardization is performed on the residuals using median absolute deviation to obtain an initial anomaly index for node energy consumption, thereby reducing the impact of outliers and non-Gaussian noise on the detection results. Further, based on historical normal operating samples, time-domain and frequency-domain feature analysis is performed on the node energy consumption sequence to extract statistical indicators that reflect the cyclical structure characteristics and frequency distribution features of energy consumption. Based on these, a reference model is constructed to describe the frequency structure of normal node energy consumption. On this basis, a gating factor is calculated according to the degree of difference between the current energy consumption sequence and the normal frequency structure to characterize the degree of deviation of the energy consumption sequence at the frequency structure level. Subsequently, the gating factor is fused with the initial anomaly score and organized into structured prompts, which are then input into a finely tuned large language model for inference analysis to obtain the final anomaly score for node energy consumption. Based on this anomaly score, the identification and prioritization of data center node energy consumption anomalies can be achieved.

[0029] In one embodiment, step S110 involves acquiring multi-source data from data center computing nodes.

[0030] It should be clarified that the multi-source data in this embodiment includes computing node identifiers, total power consumption of data center computing nodes, and operating condition data. It should also be clarified that the operating condition data is constructed from environmental parameters and equipment operating parameters, and the equipment operating parameters include task characteristics and load characteristics. Preferably, the multi-source data acquisition method is the monitoring data of the data center computing nodes during 15 minutes of continuous operation.

[0031] After acquiring multi-source data from data center computing nodes, the process also includes preprocessing the multi-source data; the preprocessing includes data alignment based on a unified time axis, missing value imputation based on linear interpolation, and normalization.

[0032] Specifically, after data acquisition, multi-source data from different devices are aligned to a unified timeline to ensure that subsequent feature calculations are based on time-consistent data. Then, to address potential data gaps during the acquisition process, linear interpolation is used to fill in missing values, ensuring the continuity and completeness of feature extraction within the window. Finally, to eliminate dimensional differences between different metrics and improve model training stability, a Min... Max normalization method standardizes the data.

[0033] In one embodiment, step S120 involves constructing samples from the multi-source data using a fixed-length sliding window to obtain analysis window detection samples. The analysis window detection samples include the node total power consumption sequence and operating condition sample data. It should be noted that the analysis window detection samples also include computing node identifiers and sliding window data.

[0034] Specifically, after multi-source data is preprocessed, in order to transform continuous time-series data into sample units suitable for modeling and analysis, a fixed-length sliding window is used to construct samples for node time-series data. Through this processing, the original continuous monitoring stream can be transformed into an analysis window with clear local time-series semantics, providing an input basis for the unified extraction of subsequent time-domain statistical features and frequency-domain statistical features.

[0035] In one embodiment, step S130 involves extracting time-domain and frequency-domain features based on the total power consumption sequence of the node to obtain time-domain feature data and frequency-domain feature data; and calculating the prediction residual based on the total power consumption sequence of the node, the operating condition sample data, and the node normal energy consumption baseline model.

[0036] In this embodiment, after constructing the sliding window, statistical features are extracted for each window from both the time and frequency domains. It's important to clarify that the data extracted in the time-frequency domain is the total power consumption sequence of the data center's computing nodes within a fixed time window. The time-domain features describe the amplitude variation, fluctuation intensity, and distribution pattern of the energy consumption signal over time, while the frequency-domain features characterize the periodic structure, spectral concentration, and frequency distribution characteristics of the energy consumption sequence over frequency. These two features complement each other, jointly improving the ability to express the differences between normal and abnormal modes.

[0037] The specific formula for extracting time-domain statistical features from a window is as follows: in, The total number of samples within the analysis window. This represents the total power consumption sample value of the node. The time-domain statistical characteristics are defined as follows: The mean, Standard deviation The square root amplitude, The absolute average. Peak value, For skewness, For kurtosis, For variance, Kuness index Peak indicator For waveform indicators, This is a pulse indicator.

[0038] Further spectral analysis is performed on each analysis window, and frequency domain statistical features are extracted. The specific formula is as follows: in, This indicates the total number of frequency points within the analysis window. For the first spectral amplitude at each frequency point For the first The frequency value corresponding to each frequency point. Specifically, Represents the frequency mean. Represents frequency variance. Represents frequency skewness. Represents frequency kurtosis. Represents the center of gravity frequency. Represents the standard deviation of frequency. Represents the root mean square frequency. Represents the average frequency. Represents the degree of regularity. Represents the changing parameters. Represents the eighth moment. It represents the sixteenth-order moment.

[0039] The anomaly detection method in this embodiment extracts features in the time and frequency domains based on the detected samples in the analysis window, providing input for subsequent residual analysis; and extracts typical spectral structure features of the detected samples in the analysis window through a normal spectrum template, providing input for spectral structure consistency analysis.

[0040] Further calculation of the prediction residuals, prediction residuals This value is used to characterize the deviation between actual energy consumption and the normal energy consumption baseline. The larger the value, the more significant the deviation of node energy consumption from the normal operating mode. The formula for calculating the prediction residual is: in, For the first The prediction residuals of each sample are used to characterize the degree of energy consumption anomaly of the sample. The larger the value, the more obvious the deviation of the actual energy consumption from the normal baseline. For the first The observed energy consumption value of each sample, that is, the energy consumption detection data actually collected by the data center computing nodes; For the first The predicted energy consumption value of each sample is the normal energy consumption baseline output obtained by Huber regression learning.

[0041] The specific implementation of predicting residual construction in this embodiment includes: Step S141: Process the operating condition data through the node normal energy consumption baseline model to obtain the predicted energy consumption value; calculate the average value of the node total power consumption sequence in the analysis window detection sample to obtain the actual energy consumption value of the node. Step S142: Calculate the prediction residual based on the predicted energy consumption value and the actual energy consumption value of the node.

[0042] Specifically, a Huber regression model is trained beforehand using calculated time-domain and frequency-domain feature data to obtain a baseline model for normal node energy consumption. This trained baseline model is then used to process the sample data operating conditions constructed within the current analysis window, outputting predicted energy consumption values. Finally, the average value of the total power consumption sequence of nodes in the samples detected by the analysis window is calculated to obtain the actual energy consumption value of the nodes. The use of the Huber regression model mitigates the interference of abnormal samples and extreme fluctuations on baseline learning.

[0043] In one embodiment, step S140 involves extracting typical spectral structure features of the node's total power consumption sequence from the frequency domain feature data using a normal spectral template, and calculating the spectral consistency index and main frequency offset based on the typical spectral structure features; and standardizing the prediction residual to obtain the standardized residual.

[0044] In this embodiment, the normal spectrum template is constructed as follows: Step S131: Obtain normal operation samples of historical operating data of data center computing nodes; Step S132: Calculate the normalized spectrum of the normally operating sample; Step S133: Construct a normal spectrum template based on the normalized spectrum.

[0045] Specifically, firstly, a normal operating sample set is selected from the historical operating data of the data center computing nodes, and its normalized spectrum is calculated; then, a spectrum template of the node under normal operating conditions is constructed, and the normal spectrum template is defined as: in, Indicates the normal spectral template in the 1st... The average spectral value at each frequency point corresponds to the typical spectral structure characteristics obtained in this step. For the first The normal sample at the ... Normalized spectral values ​​at each frequency point This represents the sample set consisting of all normal samples. This spectral template is used to characterize the typical spectral structure features of a node under normal operating conditions.

[0046] In this embodiment, frequency domain data features in step S130 of the embodiment are extracted by using a normal spectrum template. Normalized spectrum values ​​are first obtained, and then average values ​​are calculated to obtain typical spectrum structure features, which provides a basis for further analysis in the frequency domain.

[0047] It is important to note that in actual data center operation, short-term fluctuations in node energy consumption may originate from changes in business load rather than equipment malfunctions. Therefore, relying solely on residuals for anomaly identification may lead to false alarms. To improve the accuracy of anomaly identification, spectral structure consistency analysis is introduced based on the obtained standardized residuals. This analysis, combined with the spectral structure characteristics of the energy consumption sequence, assists in the identification of node operating modes, thereby improving the reliability of anomaly identification. By comparing the differences between the current energy consumption sequence spectrum and the normal operating spectrum template, residual anomalies are adaptively adjusted, resulting in more reliable anomaly labeling.

[0048] Specifically, after obtaining typical spectral structure features, a weighted cosine similarity is defined to measure the consistency between the current window's spectral structure and the normal spectral template. As an indicator of spectral consistency. When A larger value indicates that the current typical spectral structure is closer to the normal operating mode. To highlight the representativeness of long-term stable frequency points to the normal mode, frequency point weights are generated based on the fluctuation degree of historical normal samples at each frequency point, thereby improving the discriminative effectiveness of the similarity index.

[0049] Spectral similarity (weighted cosine similarity) alone may not be sufficient to fully reflect changes in the spectral center position. Therefore, a dominant frequency offset is further introduced, and its calculation formula is as follows: in, The dominant frequency of the current sample. This is the main frequency of the normal spectrum template. Represents extremely small positive numbers (smoothing / regularization term). This represents the maximum dominant frequency of all samples. This represents the minimum dominant frequency among all samples.

[0050] The dominant frequency offset calculation formula is a normalized index used to quantify the deviation between the dominant frequency of the current sample and the dominant frequency of the normal spectral template. It quantifies the degree of offset between the dominant frequency of the current sample and the dominant frequency of the normal template, supplementing the weighted cosine similarity and more comprehensively reflecting changes in the spectral structure. A larger offset indicates that the dominant frequency of the current sample deviates further from the normal pattern, and the higher the probability of an anomaly.

[0051] In this embodiment, after obtaining the prediction residual, the prediction residual is standardized to obtain the standardized residual.

[0052] Specifically, in actual operating environments, node energy consumption data may experience short-term fluctuations or extreme values, which affect the residual distribution. Therefore, to further improve the robustness of anomaly measurement, this embodiment of the invention performs standardization processing on the residual sequence based on the median absolute deviation, which is defined as: The standardized residuals are: in, Indicates the first i Robust standardized residuals of a sample This represents the median of all residuals. The median absolute deviation represents the residuals. The constant 1.4826 is a scaling factor used to keep the median absolute deviation (MAD) consistent with the standard deviation under the assumption of normality.

[0053] Standardized residuals describe the degree of deviation of the current sample energy consumption from the normal baseline and can be used to characterize the anomalous changes in node energy consumption over the time domain.

[0054] In one embodiment, step S150 involves calculating a gating factor and an adaptive anomaly threshold based on the spectrum consistency index and the dominant frequency offset.

[0055] After obtaining the spectral similarity and dominant frequency offset, a spectral consistency gating factor is defined. The spectral similarity and dominant frequency offset are incorporated into the residual discrimination process. The formula for calculating the spectral consistency gating factor is as follows: in, For the Sigmoid function, The weights of the spectral inconsistency terms are represented. This indicates the weight of the main frequency offset term. For bias terms, To represent the main frequency offset; As the gating factor, This represents the spectral consistency index (weighted cosine similarity).

[0056] This embodiment introduces a gating factor. When the sample residual is large but the spectral structure is still highly consistent with the normal template, the gating factor can suppress false alarms. When the sample not only has a large residual, but also the spectral structure deviates significantly from the normal pattern, the gating factor will increase the intensity of abnormal response.

[0057] In this embodiment, to further reduce the false alarm rate of normal high fluctuations, an adaptive threshold after gating is defined, and the expression for the adaptive abnormal threshold is: in, Indicates the first i Adaptive anomaly threshold for each sample Indicates the basic threshold. Indicates the spectrum consistency index. This indicates the main frequency offset. This represents the adjustment coefficient.

[0058] This embodiment sets an adaptive anomaly threshold so that the final annotation result no longer depends on a fixed threshold, but is determined by the degree of regression residual anomaly and the degree of deviation of the spectral structure from the normal template, thus forming an adaptive anomaly annotation result that is more in line with the actual operating environment.

[0059] In one embodiment, step S160 involves calculating the gated augmented residual anomaly based on the standardized residual and the gated factor, and generating an adaptive anomaly labeling result based on the gated augmented residual anomaly anomaly and the adaptive anomaly threshold.

[0060] In this embodiment, after obtaining the gating factor, the anomaly degree after gating enhancement is defined; wherein, the formula for calculating the anomaly degree after gating enhancement is: in, Indicates the abnormality of gated enhancement residuals. Represents the standardized residual. Represents the weight parameters. This is the consistency gating factor.

[0061] When the spectral structure deviates from the normal pattern, the gating factor increases, thereby amplifying the residual abnormality.

[0062] The abnormality of the gated enhancement residual is compared with the adaptive anomaly threshold to classify the normal and abnormal operating states. This step uses an adaptive anomaly threshold that is dynamically updated with real-time spectral characteristics. The threshold can synchronously adapt to the main frequency offset and spectral distortion of the energy consumption signal within the window, dynamically widening / tightening the judgment boundary, and significantly reducing the number of false alarms under complex operating conditions.

[0063] In one embodiment, step S170 involves uniformly encoding the operating condition sample data, computing node identifiers, sliding window data, time-domain feature data, frequency-domain feature data, prediction residuals, spectral consistency index, dominant frequency offset, gating factor, standardized residuals, gating enhancement residual anomaly, and adaptive anomaly labeling results into a prompt, which is then input into a large language model fine-tuned based on LoRA parameters to obtain an energy consumption anomaly detection score. It is understood that the analysis window detection sample in this embodiment includes computing node identifiers extracted from multi-source data, operating condition sample data (constructed from environmental parameters and equipment operating parameters), and sliding time window information.

[0064] After completing the adaptive anomaly labeling, the node identifier, sliding time window information, and operational sample data, along with time-domain statistical features, frequency-domain statistical features, regression residuals, standardized residuals, spectral consistency index, dominant frequency offset, gating factor, and gating-enhanced residual anomaly degree, are encoded into a prompt. This method maps information originally scattered across different dimensions and statistical spaces into natural language, facilitating cross-feature association learning by the model. Then, a large language model is selected as the base model, and LoRA is used for efficient parameter fine-tuning. The fine-tuned large language model outputs a final anomaly score to the input prompt. This anomaly score represents the overall anomaly risk level of a node or time window. A higher score indicates a higher anomaly risk for that node or time window, requiring priority for operational checks and interventions.

[0065] This embodiment uses a fine-tuned large language model based on the Transformer architecture, which relies entirely on a deep learning network structure and is abbreviated as TF-WLLM (Time-Frequency Window-based Large Language Model). This embodiment selects the Tongyi Qianwen 2.5-15B instruction fine-tuning large model for fine-tuning, and the hyperparameter settings are shown in the table below: Table 1 Monitoring and Fine-tuning Configuration Table 2 Low-rank adaptive fine-tuning like Figure 2 As shown, this embodiment of the invention also provides a data center energy consumption anomaly detection system based on time-frequency domain fusion, comprising: The data acquisition module acquires multi-source data from the data center's computing nodes; The sample construction module constructs samples from the multi-source data using a fixed-length sliding window to obtain analysis window detection samples; the analysis window detection samples include the node total power consumption sequence and operating condition sample data. The first processing module performs time-domain and frequency-domain feature extraction based on the total power consumption sequence of the nodes to obtain time-domain feature data and frequency-domain feature data; and calculates the prediction residual based on the total power consumption sequence of the nodes, operating condition sample data and the node normal energy consumption baseline model. The second processing module extracts typical spectral structure features of the node's total power consumption sequence from the frequency domain feature data using a normal spectral template, and calculates the spectral consistency index and main frequency offset based on the typical spectral structure features; and standardizes the prediction residual to obtain the standardized residual. The third processing module calculates the gating factor and adaptive anomaly threshold based on the spectrum consistency index and the main frequency offset. The fourth processing module calculates the gated enhancement residual anomaly based on the standardized residual and the gate factor, and generates adaptive anomaly labeling results based on the gated enhancement residual anomaly anomaly anomaly threshold. The result prediction module encodes the operating condition sample data, sliding window information, computing node identifiers, time-domain feature data, frequency-domain feature data, prediction residuals, spectrum consistency index, main frequency offset, gating factor, standardized residuals, gating enhancement residual abnormality, and adaptive anomaly labeling results into prompt words, and then inputs them into a large language model based on LoRA parameter fine-tuning to obtain an energy consumption anomaly detection score.

[0066] It is understood that the data center energy consumption anomaly detection system based on time-frequency domain fusion provided in this embodiment of the invention corresponds to the above-mentioned data center energy consumption anomaly detection method based on time-frequency domain fusion. The explanation, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the data center energy consumption anomaly detection method based on time-frequency domain fusion, and will not be repeated here.

[0067] This invention also provides a computer-readable storage medium storing a computer program for a data center energy consumption anomaly detection method based on time-frequency domain fusion, wherein the computer program causes a computer to execute the data center energy consumption anomaly detection method based on time-frequency domain fusion as described above.

[0068] This application also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the data center energy consumption anomaly detection method based on time-frequency domain fusion as described above.

[0069] In summary, compared with existing technologies, it has the following beneficial effects: 1. This invention introduces joint analysis of time-domain deviation information and frequency-domain structure information. The time-domain residual characterizes the magnitude deviation of the actual energy consumption of a node relative to the normal prediction baseline, while the frequency-domain features characterize the structural deviation of the current energy consumption sequence relative to the frequency structure template under historical normal operating conditions. This simultaneously depicts both the intensity and structural changes of energy consumption anomalies, overcoming the shortcomings of existing technologies that rely solely on time-domain deviations and struggle to distinguish between normal service fluctuations and abnormal disturbances. By identifying abnormal states characterized by frequency structure changes, comprehensive detection of multiple types of energy consumption anomalies is achieved, enhancing the anomaly identification capability under complex operating environments and improving the accuracy and stability of detection results.

[0070] 2. The present invention introduces a spectrum gating method, which uses whether the current energy consumption change conforms to the normal frequency structure as the basis for adjusting the anomaly confidence level. On the one hand, this method can reduce misjudgments caused by normal service fluctuations and avoid misidentifying energy consumption changes driven by normal services as anomalies. On the other hand, it can enhance the detection sensitivity of early hidden anomalies, thereby reducing the false alarm rate and improving the effectiveness of anomaly detection.

[0071] 3. This invention organizes time-domain deviation information, frequency-domain structure deviation information, gating results, and relevant operational context information into a structured input for anomaly detection. This input is then fed into a finely tuned large language model for inference analysis to output the final anomaly score for node energy consumption. By fusing the dependencies between different indicators and the impact of operational context, the final anomaly score is no longer limited to the judgment of a single indicator exceeding its limits. Instead, it is based on multi-dimensional information collaborative analysis, thus exhibiting better robustness, consistency, and ranking capabilities. This allows for more accurate application to data center node-level anomaly screening, anomaly priority ranking, and operational support decision-making.

[0072] 4. In this embodiment of the invention, based on the consistency between the current time window energy consumption sequence and the historical normal spectrum template, a gating factor reflecting the stability of the frequency domain structure is calculated, and the initial anomaly degree is corrected by weighting the gating factor. At the same time, the anomaly judgment threshold is dynamically adjusted.

[0073] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0074] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting anomalies in data center energy consumption based on time-frequency domain fusion, characterized in that, include: Acquire multi-source data from data center computing nodes; Samples are constructed from the multi-source data using a fixed-length sliding window to obtain samples for analysis window detection. The analysis window detects samples including the node total power consumption sequence and operating condition sample data; Based on the total power consumption sequence of the nodes, time-domain and frequency-domain features are extracted to obtain time-domain feature data and frequency-domain feature data. The prediction residual is calculated based on the node total power consumption sequence, operating condition sample data and node normal energy consumption baseline model. Based on the frequency domain feature data, typical spectral structure features of the node's total power consumption sequence are extracted using a normal spectral template, and spectral consistency index and main frequency offset are calculated based on the typical spectral structure features; the prediction residual is then standardized to obtain the standardized residual. Calculate the gating factor and adaptive anomaly threshold based on the aforementioned spectrum consistency index and dominant frequency offset; The gated enhancement residual anomaly is calculated based on the standardized residual and the gated factor, and an adaptive anomaly labeling result is generated based on the gated enhancement residual anomaly and the adaptive anomaly threshold. The operating condition sample data, sliding window information, computing node identifiers, time domain feature data, frequency domain feature data, prediction residuals, spectrum consistency index, main frequency offset, gating factor, standardized residuals, gating enhancement residual abnormality, and adaptive anomaly labeling results are uniformly encoded into prompt words and then input into a large language model based on LoRA parameter fine-tuning to obtain an energy consumption anomaly detection score. The formula for calculating the gated enhancement residual abnormality is as follows: in, For the Sigmoid function, The weights of the spectral inconsistency terms are represented. This indicates the weight of the main frequency offset term. For bias terms, To represent the main frequency offset; As the gating factor, Indicates the spectrum consistency index. Indicates the abnormality of gated enhancement residuals. Represents the standardized residual. Indicates the weighting parameter; The formula for calculating the adaptive anomaly threshold is: in, Indicates the first i Adaptive anomaly threshold for each sample Indicates the basic threshold. Indicates the spectrum consistency index. This indicates the main frequency offset. This represents the adjustment coefficient.

2. The data center energy consumption anomaly detection method according to claim 1, characterized in that, The calculation of the prediction residual based on the node's total power consumption sequence, operating condition sample data, and the node's normal energy consumption baseline model includes: The operating condition sample data is processed by the node normal energy consumption baseline model to obtain the predicted energy consumption value; the node total power consumption sequence is averaged to obtain the actual energy consumption value of the node. The predicted residual is obtained by calculating the residual based on the predicted energy consumption value and the actual energy consumption value of the node.

3. The data center energy consumption anomaly detection method according to claim 1, characterized in that, After acquiring multi-source data from the data center computing nodes, the process also includes: The multi-source data is preprocessed; the preprocessing includes data alignment based on a unified time axis, missing value imputation based on linear interpolation, and normalization.

4. The data center energy consumption anomaly detection method according to claim 1, characterized in that, The normal spectrum template is constructed as follows: Obtain normal operation samples from the historical operation data of data center computing nodes; Calculate the normalized spectrum of the normally operating sample; Construct a normal spectrum template based on the normalized spectrum; The expression for the normal spectrum template is: in, Indicates the normal spectral template in the 1st... Average spectral value at each frequency point For the first The normal sample at the ... Normalized spectral values ​​at each frequency point This represents the sample set consisting of all normal samples.

5. The data center energy consumption anomaly detection method according to claim 1, characterized in that, The expression for the main frequency offset is: in, The dominant frequency of the current sample. This is the main frequency of the normal spectrum template. Represents extremely small positive numbers. This represents the maximum dominant frequency of all samples. This represents the minimum dominant frequency among all samples.

6. A data center energy consumption anomaly detection system based on time-frequency domain fusion, characterized in that, include: The data acquisition module acquires multi-source data from the data center's computing nodes; The sample construction module constructs samples from the multi-source data using a fixed-length sliding window to obtain analysis window detection samples; the analysis window detection samples include the node total power consumption sequence and operating condition sample data. The first processing module performs time-domain and frequency-domain feature extraction based on the total power consumption sequence of the node to obtain time-domain feature data and frequency-domain feature data. The prediction residual is calculated based on the node total power consumption sequence, operating condition sample data and node normal energy consumption baseline model. The second processing module extracts typical spectral structure features of the node's total power consumption sequence from the frequency domain feature data using a normal spectral template, and calculates the spectral consistency index and main frequency offset based on the typical spectral structure features; and standardizes the prediction residual to obtain the standardized residual. The third processing module calculates the gating factor and adaptive anomaly threshold based on the spectrum consistency index and the main frequency offset. The fourth processing module calculates the gated enhancement residual anomaly based on the standardized residual and the gate factor, and generates adaptive anomaly labeling results based on the gated enhancement residual anomaly anomaly anomaly threshold. The result prediction module encodes the running condition sample data, sliding window information, computing node identifiers, time domain feature data, frequency domain feature data, prediction residuals, spectrum consistency index, main frequency offset, gating factor, standardized residuals, gating enhancement residual abnormality, and adaptive anomaly labeling results into prompt words and inputs them into a large language model based on LoRA parameter fine-tuning to obtain an energy consumption anomaly detection score. The formula for calculating the gated enhancement residual abnormality is as follows: in, For the Sigmoid function, The weights of the spectral inconsistency terms are represented. This indicates the weight of the main frequency offset term. For bias terms, To represent the main frequency offset; As the gating factor, Indicates the spectrum consistency index. Indicates the abnormality of gated enhancement residuals. Represents the standardized residual. Indicates the weighting parameter; The formula for calculating the adaptive anomaly threshold is: in, Indicates the first i Adaptive anomaly threshold for each sample Indicates the basic threshold. Indicates the spectrum consistency index. This indicates the main frequency offset. This represents the adjustment coefficient.

7. A computer-readable storage medium, characterized in that, It stores a computer program for detecting anomalies in data center energy consumption based on time-frequency domain fusion, wherein the computer program causes a computer to execute the data center energy consumption anomaly detection method based on time-frequency domain fusion as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the data center energy consumption anomaly detection method based on time-frequency domain fusion as described in any one of claims 1 to 5.

Images