An operation and maintenance index prediction method, device, system, equipment and medium

Abstract

This disclosure provides a method, apparatus, system, device, and medium for predicting operation and maintenance indicators. The method includes: extracting time series data of preset operation and maintenance indicators of target equipment from a time series database of an intelligent computing center and using it as the original time series; decomposing the original time series to obtain a trend term sequence and a seasonal term sequence; performing adaptive normalization processing on the seasonal term sequence to obtain a normalized seasonal term sequence; inputting the normalized seasonal term sequence into a seasonal term prediction model based on frequency domain processing to obtain a normalized seasonal term prediction sequence; inputting the trend term sequence into the trend term prediction model to obtain a trend term prediction sequence; and merging the normalized seasonal term prediction sequence and the trend term prediction sequence to obtain a merged prediction sequence. Embodiments of this disclosure can effectively address the non-stationary characteristics of operation and maintenance indicators and accurately capture multiple seasonal and nonlinear mutation patterns in operation and maintenance indicators.

Description

Technical Field

[0001] This disclosure relates to the field of IT operations and maintenance technology, and in particular to a method, apparatus, system, equipment and medium for predicting operations and maintenance indicators. Background Technology

[0002] Intelligent Computing Centers (ICCs) undertake critical tasks such as large-scale artificial intelligence training and inference, and their operational status directly impacts business efficiency. These tasks are typically characterized by high susceptibility and drastic dynamic changes in resource demands, resulting in complex time-series fluctuations in the utilization of computing and network resources within the ICC. Traditional operation and maintenance (O&M) systems primarily rely on real-time monitoring and passive alarms based on static thresholds, which are insufficient to effectively address this dynamic nature. When resource bottlenecks or potential failures are imminent, O&M personnel often cannot anticipate them, leading to delayed responses that may affect business continuity or performance. Therefore, introducing intelligent predictive capabilities into the ICC O&M system to predict future trends of key O&M indicators is crucial for achieving proactive O&M and optimizing resource scheduling.

[0003] Time series forecasting technology has been extensively explored in the IT operations and maintenance field, including traditional statistical models, models based on recurrent neural networks, and various variant models based on Transformers. Some commercial or open-source operations and maintenance platforms have also integrated basic time series forecasting capabilities. However, these existing technologies all use general forecasting models, which have many shortcomings when applied to predicting operations and maintenance indicators in intelligent computing centers, as follows:

[0004] (1) Insufficient prediction accuracy: Existing general prediction models are difficult to accurately capture the complex nonlinear patterns, multiple seasonalities, and abrupt changes driven by AI (Artificial Intelligence) tasks in the operation and maintenance indicators of intelligent computing centers. For data containing complex trend items and multiple seasonalities, there is a lack of refined processing methods.

[0005] (2) Poor handling of non-stationary data: The operation and maintenance indicators of the intelligent computing center are often non-stationary. Traditional models or simple normalization methods are difficult to adapt to the dynamic changes in data distribution, which seriously affects the stability of prediction.

[0006] (3) Insufficient processing of complex pattern decomposition: The trend and seasonal terms in the time series were not effectively separated and predicted independently, resulting in mutual interference between different features and further reducing the prediction accuracy.

[0007] (4) Lack of deep integration with operation and maintenance processes. Prediction results are often presented independently and fail to be effectively integrated into operation and maintenance workflows such as visualization and alarm triggering. The value of prediction is not fully realized and cannot provide effective support for proactive operation and maintenance. Summary of the Invention

[0008] This disclosure provides a method, apparatus, system, equipment, and medium for predicting operation and maintenance indicators, in order to solve the problems of insufficient prediction accuracy, poor handling of non-stationary data, insufficient handling of complex pattern decomposition, and lack of deep integration with operation and maintenance processes in existing operation and maintenance indicator prediction methods.

[0009] Firstly, this disclosure provides a method for predicting operational and maintenance indicators, the method comprising:

[0010] Extract time series data of preset operation and maintenance indicators of target devices from the time series database of the intelligent computing center, and use it as the raw time series;

[0011] The original time series is decomposed to obtain a trend sequence and a seasonal sequence;

[0012] The seasonal term sequence is subjected to adaptive normalization to obtain a normalized seasonal term sequence;

[0013] The normalized seasonal term sequence is input into the frequency domain-based seasonal term prediction model to obtain the normalized seasonal term prediction sequence.

[0014] The trend term sequence is input into the trend term prediction model to obtain the trend term prediction sequence;

[0015] The normalized seasonal prediction sequence is merged with the trend prediction sequence to obtain a merged prediction sequence, and the merged prediction sequence is denormalized to obtain the final prediction sequence.

[0016] The final predicted sequence is pushed to the visualization interface for display, and / or an alarm is triggered when the predicted value of the preset operation and maintenance indicator in the final predicted sequence exceeds the preset alarm threshold.

[0017] Furthermore, the step of extracting time-series data of preset operation and maintenance indicators of the target device from the time-series database of the intelligent computing center and using it as the original time series specifically includes:

[0018] Receive a prediction task, the prediction task including the preset operation and maintenance indicators, the device identifier of the target device, the duration of historical data and the time granularity;

[0019] Based on the preset operation and maintenance indicators, device identifier, historical data duration and time granularity, the time series database is queried to obtain the time series data of the preset operation and maintenance indicators of the target device aggregated according to the time granularity within the historical data duration, and this data is used as the original time series.

[0020] Furthermore, the decomposition of the original time series to obtain a trend sequence and a seasonal sequence specifically includes:

[0021] The original time series is subjected to a moving average operation using the moving average method to obtain the trend term series.

[0022] The seasonal sequence is obtained by subtracting the trend sequence from the original time series.

[0023] Further, the adaptive normalization process performed on the seasonal term sequence to obtain the normalized seasonal term sequence specifically includes:

[0024] Extract fixed-length continuous sequence windows sequentially from the seasonal item sequence in chronological order;

[0025] For each sequence window, the data of the sequence window is input into BackNET, the current mean and current variance of the data in the sequence window are calculated by BackNET, and the data in the sequence window are normalized based on the current mean and current variance to obtain the normalized data of the sequence window;

[0026] By concatenating the normalized data from all sequence windows in chronological order, a normalized seasonal sequence is obtained.

[0027] For each sequence window, the data of the sequence window is input into HoriNET, and the future mean and future variance corresponding to the future output window are predicted by HoriNET.

[0028] Further, the step of inputting the normalized seasonal term sequence into the frequency domain-based seasonal term prediction model to obtain the normalized seasonal term prediction sequence specifically includes:

[0029] The normalized seasonal term sequence is input into the FEDformer model. The frequency domain encoder of the FEDformer model performs fast Fourier transform, frequency domain information selection and inverse Fourier transform on the normalized seasonal term sequence. Then, the self-attention and cross-attention mechanisms of the frequency domain decoder are used to calculate and output the normalized seasonal term prediction sequence.

[0030] Further, the step of merging the normalized seasonal prediction sequence with the trend prediction sequence to obtain a merged prediction sequence, and then performing inverse normalization on the merged prediction sequence to obtain the final prediction sequence, specifically includes:

[0031] The normalized seasonal prediction sequence and the trend prediction sequence are added point by point according to the timestamp to obtain the merged prediction sequence;

[0032] Based on the future mean and future variance corresponding to the future output window predicted by HoriNET, the merged prediction sequence is denormalized to obtain the final prediction sequence with real physical dimensions.

[0033] Furthermore, the step of triggering an alarm when the predicted value of the preset operation and maintenance indicator in the final prediction sequence exceeds a preset alarm threshold specifically includes:

[0034] Obtain the predicted value for each future time point in the final prediction sequence;

[0035] The predicted value is compared with a preset alarm threshold to determine whether there are multiple consecutive time points where the predicted value exceeds the alarm threshold.

[0036] If present, a predictive alarm is generated, which includes at least one of the following: the predicted risk occurrence time, the name of the preset operation and maintenance indicator, the target device identifier, and suggested measures. The predictive alarm is then pushed to the operation and maintenance personnel's terminal.

[0037] Secondly, this disclosure provides a device for predicting operation and maintenance indicators, the device comprising:

[0038] The raw time series acquisition module is used to extract time series data of preset operation and maintenance indicators of target devices from the time series database of the intelligent computing center and use it as the raw time series.

[0039] The trend and seasonal item decomposition module is connected to the original time series acquisition module and is used to decompose the original time series to obtain the trend item sequence and the seasonal item sequence.

[0040] An adaptive normalization processing module, connected to the trend seasonal item decomposition module, is used to perform adaptive normalization processing on the seasonal item sequence to obtain a normalized seasonal item sequence.

[0041] The normalized seasonal term prediction module is connected to the adaptive normalization processing module and is used to input the normalized seasonal term sequence into the frequency domain-based seasonal term prediction model to obtain the normalized seasonal term prediction sequence.

[0042] The trend term prediction sequence acquisition module is connected to the normalized seasonal term prediction module and is used to input the trend term sequence into the trend term prediction model to obtain the trend term prediction sequence.

[0043] The final prediction sequence acquisition module is connected to the trend item prediction sequence acquisition module. It is used to merge the normalized seasonal item prediction sequence with the trend item prediction sequence to obtain a merged prediction sequence, and to perform inverse normalization on the merged prediction sequence to obtain the final prediction sequence.

[0044] The operation and maintenance indicator prediction application module is connected to the final prediction sequence acquisition module. It is used to push the final prediction sequence to the visualization interface for display, and / or trigger an alarm when the predicted value of the preset operation and maintenance indicator in the final prediction sequence exceeds the preset alarm threshold.

[0045] Thirdly, this disclosure provides a prediction system for operation and maintenance indicators, including a prediction module, which is used to execute the prediction method for operation and maintenance indicators described in the first aspect above.

[0046] Fourthly, this disclosure provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores one or more computer programs executable by the at least one processor, the one or more computer programs being executed by the at least one processor to enable the at least one processor to perform the operation and maintenance indicator prediction method described in the first aspect above.

[0047] Fifthly, this disclosure provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method for predicting operational indicators as described in the first aspect.

[0048] The prediction method, apparatus, system, equipment, and medium for operation and maintenance (O&M) indicators provided in this disclosure effectively address the non-stationary characteristics of O&M indicators and adapt to dynamic changes in data distribution by decomposing the original time series into trend and seasonal sequences and adaptively normalizing the seasonal sequences. By inputting the normalized seasonal sequences into a frequency-domain-based seasonal prediction model, it accurately captures multiple seasonal and nonlinear abrupt change patterns in O&M indicators, significantly improving prediction accuracy. By predicting the trend and seasonal sequences separately and then merging and inverse-normalizing them, it achieves refined decomposition of complex patterns, avoiding mutual interference between trend and seasonal components and further improving prediction reliability. By pushing the final predicted sequence to a visualization interface and / or triggering alarms, the prediction results are deeply integrated into the O&M process, providing real-time and reliable data support and decision-making basis for proactive O&M. This solves the problems of insufficient prediction accuracy, poor handling of non-stationary data, inadequate decomposition of complex patterns, and lack of deep integration with the O&M process in existing O&M indicator prediction methods. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the present disclosure and form part of the specification. They are used together with the embodiments of the present disclosure to explain the disclosure and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 A flowchart illustrating a method for predicting operational and maintenance metrics provided in this embodiment of the disclosure;

[0051] Figure 2 An architecture diagram of the operation and maintenance indicator prediction system provided in this embodiment of the disclosure;

[0052] Figure 3 A block diagram of a device for predicting operation and maintenance indicators provided in an embodiment of this disclosure;

[0053] Figure 4 This is a block diagram of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation

[0054] To enable those skilled in the art to better understand the technical solutions of this disclosure, exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments of this disclosure to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0055] Where there is no conflict, the various embodiments of this disclosure and the features thereof in the embodiments may be combined with each other.

[0056] As used herein, the term “and / or” includes any and all combinations of one or more related enumerated entries.

[0057] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. As used herein, the singular forms “a” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that when the terms “comprising” and / or “made of” are used in this specification, the presence of the stated feature, integral, step, operation, element, and / or component is specified, but the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof is not excluded. Words such as “connected” or “linked” are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0058] Unless otherwise specified, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and this disclosure, and will not be interpreted as having an idealized or overly formal meaning, unless expressly so defined herein.

[0059] Figure 1 A flowchart illustrating a method for predicting operational metrics provided in this embodiment of the disclosure. (Refer to...) Figure 1 The method includes:

[0060] Step S101: Extract time series data of the target device's preset operation and maintenance indicators from the intelligent computing center's time series database and use it as the original time series.

[0061] Specifically, the time-series database of the intelligent computing center is used to store the full-volume operation and maintenance monitoring data of the intelligent computing center collected by the operation and maintenance platform through the Agent and protocol interface. The target device is the computing node or network device in the intelligent computing center. The preset operation and maintenance indicators are the pre-selected key operation and maintenance parameters of the intelligent computing center, such as GPU (Graphics Processing Unit) utilization, port transmission bandwidth, and the number of PFC (Priority-based Flow Control) pause frames per second. The extraction process needs to retain the original timestamp information of the data to ensure the continuity of the extracted time series data in the time dimension, thereby forming a raw time series that can be directly used for subsequent preprocessing.

[0062] In some embodiments, the step of extracting time-series data of preset operation and maintenance indicators of the target device from the time-series database of the intelligent computing center and using it as the original time series specifically includes:

[0063] Receive a prediction task, the prediction task including the preset operation and maintenance indicators, the device identifier of the target device, the duration of historical data and the time granularity;

[0064] Based on the preset operation and maintenance indicators, device identifier, historical data duration and time granularity, the time series database is queried to obtain the time series data of the preset operation and maintenance indicators of the target device aggregated according to the time granularity within the historical data duration, and this data is used as the original time series.

[0065] Specifically, the prediction task can be manually configured and triggered by the user through the unified user interface of the intelligent computing center operation and maintenance platform, or it can be automatically triggered by the system according to preset operation and maintenance rules. The device identifier is the unique identification information of the device in the intelligent computing center, such as serverA-gpu0, switchX-portY, etc., used to accurately locate the target computing node or network device. The historical data duration is the time range of historical data of operation and maintenance indicators used for subsequent prediction, such as the last 7 days, the last 30 days, etc. The time granularity is the time interval of data aggregation, such as 1 minute, 5 minutes, etc. When querying, the original monitoring data in the time series database is aggregated according to the time granularity, and time alignment and resampling are completed at the same time to ensure that the obtained original time series data points are evenly distributed on the time axis to meet the processing requirements of subsequent sequence decomposition.

[0066] Step S102: Decompose the original time series to obtain the trend sequence and the seasonal sequence.

[0067] Specifically, the original time series includes the trend characteristics of the intelligent computing center's operation and maintenance indicators over time, the seasonal characteristics of multiple periodic fluctuations, and noise components. The trend sequence is used to characterize the long-term overall change trend of the operation and maintenance indicators, while the seasonal sequence is used to characterize the periodic fluctuation characteristics of the operation and maintenance indicators over time, such as daily cycle, weekly cycle, task cycle, etc. By performing feature separation decomposition on the original time series, targeted prediction processing can be performed on sequences with different characteristics, avoiding the problem of decreased prediction accuracy caused by mutual interference between different features, and adapting to the complex data patterns of the intelligent computing center's operation and maintenance indicators.

[0068] In some embodiments, the decomposition of the original time series to obtain a trend sequence and a seasonal sequence specifically includes:

[0069] The original time series is subjected to a moving average operation using the moving average method to obtain the trend term series.

[0070] The seasonal sequence is obtained by subtracting the trend sequence from the original time series.

[0071] Specifically, the moving average method can employ a simple moving average with a fixed window size or a weighted moving average. The window size is set according to the main periodic characteristics of the intelligent computing center's operation and maintenance indicators. For example, for minute-level GPU utilization data, a window size of 1440 points can be set for the daily period. Before sliding, the original time series is first filled with boundary data, and then the moving average calculation is performed point by point to smooth out short-term fluctuations and noise components in the original time series, accurately extracting the overall trend characteristics of the series to obtain the trend term series. Then, the difference between the value of each time point in the original time series and the corresponding time point in the trend term series is calculated. After stripping the trend characteristics, a seasonal term series containing the periodic fluctuations of the operation and maintenance indicators and a small amount of noise is obtained. This series can fully reflect the multiple seasonal variation patterns of the intelligent computing center's operation and maintenance indicators.

[0072] Step S103: Perform adaptive normalization on the seasonal term sequence to obtain the normalized seasonal term sequence.

[0073] Specifically, the adaptive normalization process is a dynamic normalization method designed for the non-stationary characteristics of intelligent computing center operation and maintenance indicators. Unlike traditional global static normalization, it can adaptively adjust the normalization parameters according to the local data distribution characteristics of the seasonal item sequence. This effectively addresses the data distribution drift problem caused by the suddenness of AI tasks and dynamic changes in resource requirements of intelligent computing center operation and maintenance indicators, and avoids the decrease in subsequent prediction model training and prediction accuracy due to differences in data distribution. This makes the normalized seasonal item sequence more suitable for the input requirements of the frequency domain-based seasonal item prediction model. At the same time, this adaptive normalization process can be performed only on the seasonal item sequence, while the trend item sequence, due to its relatively smooth data, does not need to undergo this processing.

[0074] In some embodiments, the adaptive normalization process performed on the seasonal term sequence to obtain a normalized seasonal term sequence specifically includes:

[0075] Extract fixed-length continuous sequence windows sequentially from the seasonal item sequence in chronological order;

[0076] For each sequence window, the data of the sequence window is input into BackNET, the current mean and current variance of the data in the sequence window are calculated by BackNET, and the data in the sequence window are normalized based on the current mean and current variance to obtain the normalized data of the sequence window;

[0077] By concatenating the normalized data from all sequence windows in chronological order, a normalized seasonal sequence is obtained.

[0078] For each sequence window, the data of the sequence window is input into HoriNET, and the future mean and future variance corresponding to the future output window are predicted by HoriNET.

[0079] Specifically, the BackNET is a small, lightweight neural network adapted to the data characteristics of intelligent computing centers, such as a neural network containing multiple linear layers. Its core function is to window the input sequence. The data adaptively learns and calculates the current mean value of the data within the window. and current variance Based on the following formula: The data within the window is normalized by introducing a small constant to prevent division by zero. The splicing operation strictly follows the temporal order of each sequence window in the original seasonal item sequence to ensure the temporal continuity and integrity of the normalized seasonal item sequence after splicing. HoriNET is a small, collaborative neural network adapted to the BackNET structure. It also receives sequence window data as input and infers the future mean corresponding to the future output window by learning the features of the current window. and future variance This provides accurate statistical parameters to support the subsequent inverse normalization of the prediction results.

[0080] Step S104: Input the normalized seasonal term sequence into the seasonal term prediction model based on frequency domain processing to obtain the normalized seasonal term prediction sequence.

[0081] Specifically, the frequency domain-based seasonal prediction model is a prediction model adapted to the multiple seasonal characteristics of intelligent computing center operation and maintenance indicators. This model leverages the advantages of frequency domain analysis to effectively capture the multiple periodic features inherent in the seasonal item sequence. The model performs encoding-decoding processing on the input sequence, explicitly models and selects the periodic components in the frequency domain space, and finally outputs a normalized seasonal item prediction sequence corresponding to the time length of the input sequence. This prediction sequence covers a specified future prediction period (e.g., the next 60 minutes).

[0082] In some embodiments, inputting the normalized seasonal term sequence into a frequency domain-based seasonal term prediction model to obtain a normalized seasonal term prediction sequence specifically includes:

[0083] The normalized seasonal term sequence is input into the FEDformer model. The frequency domain encoder of the FEDformer model performs fast Fourier transform, frequency domain information selection and inverse Fourier transform on the normalized seasonal term sequence. Then, the self-attention and cross-attention mechanisms of the frequency domain decoder are used to calculate and output the normalized seasonal term prediction sequence.

[0084] Specifically, the system inputs the normalized seasonal term sequence into the frequency domain encoder of the FEDformer model. The encoder first performs a Fast Fourier Transform (FFT) on the input sequence, converting the time-domain signal to the frequency domain to obtain the spectral representation of the sequence. Subsequently, information selection is performed in the frequency domain, such as selecting low-frequency components or components with larger amplitudes. After completing the frequency domain selection, the processed frequency-domain signal is converted back to the time domain through an Inverse Fourier Transform (IFFT) to obtain an enhanced time-domain feature representation. These features are then fed into the frequency domain decoder, whose structure is similar to that of the encoder. The decoder uses self-attention and cross-attention mechanisms to dynamically focus on key periodic patterns in the input sequence and their correlation with the target prediction period during the generation of the prediction sequence, ultimately generating the normalized seasonal term prediction sequence point by point.

[0085] Step S105: Input the trend term sequence into the trend term prediction model to obtain the trend term prediction sequence.

[0086] Specifically, the trend prediction model is a lightweight prediction model adapted to the smoothness and gradual change characteristics of the trend sequence. It can be flexibly selected according to the actual operation and maintenance needs of the intelligent computing center. For example, it can be a simple linear extrapolation model, an ARIMA (Autoregressive Integrated Moving Average Model) model, a lightweight RNN (Recurrent Neural Network) / LSTM (Long Short-Term Memory) network, or a simplified FEDformer model. This model predicts the long-term overall change trend of the operation and maintenance indicators reflected by the trend sequence. It does not require complex adaptive normalization processing of the trend sequence. It only needs to use simple global normalization or direct prediction as needed. The timestamps and prediction durations of the output trend prediction sequence are completely aligned with those of the seasonal prediction sequence.

[0087] Step S106: Merge the normalized seasonal prediction sequence with the trend prediction sequence to obtain a merged prediction sequence, and then perform inverse normalization on the merged prediction sequence to obtain the final prediction sequence.

[0088] Specifically, the merging operation aligns the normalized seasonal and trend prediction sequences based on timestamps to ensure accurate fusion of the predicted values ​​at the same time points. The merged prediction sequence retains the trend and seasonal fluctuation characteristics of the operation and maintenance indicators. The denormalization operation is performed based on the future statistical parameters obtained in the previous adaptive normalization process. It restores the normalized values ​​in the merged prediction sequence to operation and maintenance indicator values ​​with real physical meaning, eliminating the influence of normalization on the numerical transformation. The final prediction sequence has the same dimensions as the original operation and maintenance indicators and can directly reflect the true value level of the intelligent computing center's operation and maintenance indicators at future time points.

[0089] In some embodiments, merging the normalized seasonal forecast sequence with the trend forecast sequence to obtain a merged forecast sequence, and then inversely normalizing the merged forecast sequence to obtain the final forecast sequence, specifically includes:

[0090] The normalized seasonal prediction sequence and the trend prediction sequence are added point by point according to the timestamp to obtain the merged prediction sequence;

[0091] Based on the future mean and future variance corresponding to the future output window predicted by HoriNET, the merged prediction sequence is denormalized to obtain the final prediction sequence with real physical dimensions.

[0092] Specifically, the system first aligns the timestamps of the normalized seasonal forecast sequence and the trend forecast sequence to ensure that the forecast periods of the two sequences are completely consistent. Then, for each future time point, the corresponding normalized seasonal forecast value is added to the trend forecast value to generate a combined forecast value for that time point. The combined forecast values ​​of all time points constitute a combined forecast sequence. Next, based on the future time period covered by the current forecast task, the system uses the future mean corresponding to the future output window predicted by HoriNET. and future variance It performs inverse normalization to accurately restore the true physical dimensions of operation and maintenance indicators, so that the final prediction sequence can directly reflect the actual values ​​of operation and maintenance indicators such as GPU utilization and port bandwidth.

[0093] Step S107: Push the final prediction sequence to the visualization interface for display, and / or trigger an alarm when the predicted value of the preset operation and maintenance indicator in the final prediction sequence exceeds the preset alarm threshold.

[0094] Specifically, the final predicted sequence can be displayed in an intuitive form such as curves and charts, while historical monitoring data of operation and maintenance indicators can be superimposed, making it convenient for operation and maintenance personnel to compare and view the historical trends and future predicted trends of the indicators; the preset alarm threshold is a threshold parameter configured separately for different preset operation and maintenance indicators, which is adapted to the operation threshold requirements of different operation and maintenance indicators of the intelligent computing center.

[0095] In some embodiments, triggering an alarm when the predicted value of the preset operation and maintenance indicator in the final prediction sequence exceeds a preset alarm threshold specifically includes:

[0096] Obtain the predicted value for each future time point in the final prediction sequence;

[0097] The predicted value is compared with a preset alarm threshold to determine whether there are multiple consecutive time points where the predicted value exceeds the alarm threshold.

[0098] If present, a predictive alarm is generated, which includes at least one of the following: the predicted risk occurrence time, the name of the preset operation and maintenance indicator, the target device identifier, and suggested measures. The predictive alarm is then pushed to the operation and maintenance personnel's terminal.

[0099] Specifically, a full analysis is performed on the final predicted sequence to extract the predicted values ​​of preset operation and maintenance indicators and their associated timestamps for each future time point. To avoid false alarms caused by fluctuations in single-point predicted values, a continuous interval verification method is adopted. The number of consecutive time points can be flexibly configured according to the characteristics of different operation and maintenance indicators of the intelligent computing center. For example, GPU utilization can be configured for 10 consecutive minutes, and port bandwidth for 5 consecutive minutes. If the alarm triggering conditions are met, the system will generate a predictive alarm message. This predictive alarm message is in a standardized operation and maintenance alarm format. The risk occurrence time is the predicted start time of the indicator exceeding the threshold. The suggested measures are general or customized handling suggestions for the operation and maintenance indicator exceeding the threshold, such as "optimize task scheduling and release computing resources" for GPU utilization exceeding the threshold. The push notifications to operation and maintenance personnel terminals include the alarm display module of the intelligent computing center operation and maintenance platform and instant messaging tools to ensure that operation and maintenance personnel can receive and process alarm information in a timely manner.

[0100] In one specific embodiment, the prediction method for operation and maintenance indicators is applied to an operation and maintenance indicator prediction system integrated into an intelligent computing center operation and maintenance platform (such as the "Zhiyu" platform). This system utilizes the rich historical data collected by the platform to intelligently predict the future trends of key operation and maintenance indicators (i.e., preset operation and maintenance indicators) and provides proactive operation and maintenance support. The corresponding method draws upon and improves upon a prediction framework that combines sequence decomposition, adaptive normalization, and frequency domain attention mechanisms.

[0101] Figure 2 The diagram shows the architecture of the prediction system for this operation and maintenance metric. This system is integrated into the core logic layer of the intelligent computing center's operation and maintenance platform. The specific descriptions of each part are as follows:

[0102] 1. Data source (time series database):

[0103] This system relies on the existing "unified monitoring and data storage" capabilities of the operation and maintenance platform. The platform collects massive amounts of high-frequency historical monitoring data from computing nodes (CPU, memory, GPU utilization / video memory, network card traffic / counters, etc.) and network devices (port bandwidth, PFC / ECN counts, DCQCN status parameters, optical module indicators, etc.) through the Agent and protocol interfaces (SNMP, Telemetry, etc.) of the network data access layer, and stores it in a time-series database (such as InfluxDB).

[0104] 2. Core Logic Layer:

[0105] As the core processing unit of the system, it includes the following functional modules:

[0106] (1) Data access module: provides the ability to interact with the time series database to extract historical time series data of the preset operation and maintenance indicators of the target equipment.

[0107] (2) Agent Management Module: Responsible for managing and configuring the Agent of the network data access layer to ensure the stability and reliability of data collection.

[0108] (3) Prediction module: As a core functional unit, it further includes:

[0109] (a) Task Management Submodule: Receives prediction tasks (specifying indicators, prediction duration, and prediction frequency) automatically triggered by users or the system.

[0110] (b) Data extraction and preprocessing submodule: Extracts historical data of specified indicators from the time series database and performs preprocessing;

[0111] (c) Model Training and Execution Submodule: Loads pre-trained models or trains / fine-tunes models as needed, and performs prediction calculations;

[0112] (d) Result storage and output submodule: Store the prediction results (values ​​and / or confidence intervals at future time points) and provide them to other modules of the platform (such as alarm module and visualization module) through API.

[0113] (4) Alarm triggering module: Receives the final prediction sequence output by the prediction module, determines whether the predicted value of the preset operation and maintenance index exceeds the preset threshold, and generates and triggers a predictive alarm.

[0114] (5) API service module: Provides standardized interfaces to enable data exchange between prediction results, alarm information, etc., and the upper UI layer and other operation and maintenance modules.

[0115] 3. UI layer (upper-layer application):

[0116] (1) Visualization module: Presents historical data curves and predicted trend curves of operation and maintenance indicators to users, and supports prediction task configuration and result query.

[0117] (2) Alarm display module: Displays predictive alarm information generated by the alarm triggering module, which makes it easier for maintenance personnel to perceive potential risks in a timely manner.

[0118] It should be noted that the UI layer can also be extended to include a resource scheduling / optimization module, which can make intelligent resource planning and optimization decisions based on prediction results, further improving the initiative and efficiency of intelligent computing center operation and maintenance.

[0119] Based on the above architecture diagram, the prediction method for this operation and maintenance metric can include the following steps:

[0120] S1. Data Extraction and Preprocessing:

[0121] (1) Extraction: Based on the operation and maintenance indicators defined in the task (such as GPU utilization percentage, port transmission bandwidth, PFC pause frames per second), target devices (such as GPU 0 of server A, port Y of switch X), historical data duration (such as the past 7 days) and time granularity (such as aggregated to 1 minute), the raw time series Xraw is obtained by querying InfluxDB.

[0122] (2) Pretreatment:

[0123] Missing value handling: Use interpolation (linear, spline) or forward / backward filling methods to handle missing data points.

[0124] Outlier handling: Use statistical methods (such as 3-sigma) or isolated forests to identify and smooth outliers.

[0125] Time alignment and resampling: Ensure that data points are evenly distributed along the time axis.

[0126] S2, Time Series Decomposition:

[0127] To handle long-term trends and periodic fluctuations in the data separately, the preprocessed sequence X is decomposed into trend terms. and seasonal items .

[0128] Decomposition using the moving average method:

[0129]

[0130] Here, AvgPool represents the moving average operation (the window size is set according to the main period, such as 1440 points (minute level) for a daily period), and Padding is used to handle sequence boundaries.

[0131]

[0132] The seasonal term includes periodic fluctuations and noise from the original sequence.

[0133] S3, Adaptive Normalization (based on BackNET / HoriNET concepts):

[0134] Considering that the indicators of the intelligent computing center may have distribution drift (non-stationarity), adaptive normalization is adopted to improve the stability of the model.

[0135] BackNET: A small neural network (e.g., containing multiple linear layers) whose input is a window of the original (or decomposed) sequence. Output the estimated mean of the data in this window. and variance .

[0136] Input normalization: ,in It is a small constant used to prevent division by zero.

[0137] HoriNET: Its structure is similar to BackNET, and it also receives... As input (or sharing a portion of the encoder with the prediction model), predict the mean corresponding to the future output window. and variance .

[0138] Output denormalization: The prediction model outputs the normalized predicted value. The final prediction result is ,in, It is a preset micro-smoothing item.

[0139] It should be noted that this adaptive normalization process is applied to the input and output stages of the subsequent seasonal forecast model (FEDformer). The trend term, being relatively smooth, can be normalized using simple global normalization or no normalization at all.

[0140] S4, Seasonal Forecast:

[0141] The seasonal term sequence obtained by decomposition is processed using the FEDformer (Frequency Enhanced Decomposed Transformer) model. (After BackNET normalization) prediction is performed to obtain... The core advantage of FEDformer lies in its ability to effectively capture multiple periodicities by using Fourier transform to process sequences in the frequency domain.

[0142] 1) Model Structure

[0143] (a) Frequency-Enhanced Encoder:

[0144] Input a normalized sequence of seasonal terms.

[0145] Frequency Enhanced Block (FEB): A Fast Fourier Transform (FFT) is applied to the input sequence to select information in the frequency domain (e.g., selecting low-frequency components or components with larger amplitudes), and then the sequence is returned to the time domain via an Inverse Fourier Transform (IFFT). This helps extract key periodic features and reduce noise.

[0146] Frequency Enhanced Attention (FEA): Similar to the self-attention mechanism of Transformer, but it calculates attention weights in the frequency domain, which can more effectively associate patterns of different cycles.

[0147] The encoder consists of multiple layers of stacked FEB and FEA, and includes a feedforward network and a normalization layer.

[0148] (b) Frequency-Enhanced Decoder:

[0149] The structure is similar to that of an encoder, and it also contains multiple layers of FEB and FEA.

[0150] It receives the encoder output and a known portion (or placeholder) of the target prediction sequence, and generates prediction results through cross-attention and self-attention mechanisms.

[0151] 2) Output: The decoder outputs the normalized seasonal prediction sequence. .

[0152] S5. Trend Forecast:

[0153] The trend term sequence obtained from the decomposition Make a prediction and obtain .

[0154] Since the trend term is relatively smooth, a simpler model can be chosen, such as:

[0155] Simple linear extrapolation.

[0156] ARIMA model.

[0157] Or a lightweight RNN / LSTM network.

[0158] You can even use another (potentially simpler to configure) FEDformer instance for prediction.

[0159] S6. Result merging and denormalization:

[0160] (1) Combining: Add the predicted seasonal and trend terms together: (Note: If the trend term has also been normalized, this should be the trend term after denormalization.) Here, it is assumed that Ytrend has not undergone complex normalization or has been denormalized.

[0161] (2) Inverse normalization: using the HoriNET predictions from step S3 and The merged results (mainly the seasonal component) are inversely normalized to obtain the final predicted sequence Y.

[0162] (This is a simplified illustration; precise implementation requires matching the specific scope of BackNET / HoriNET.)

[0163] Or more precisely, if BackNET / HoriNET is applied to sequence X before decomposition, then for the final merged sequence... Perform inverse normalization.

[0164] S7. Output and Application:

[0165] (1) Storage: Store the final predicted sequence Y (containing the predicted values ​​at each future time point) in the database and associate it with the original prediction task information.

[0166] (2) Visualization: On the UI interface, while displaying historical indicator curves, a prediction curve for future time periods (which may include confidence intervals) is drawn to help users intuitively understand the trend.

[0167] (3) Active alarm: The alarm module reads the prediction results. If the predicted value will continue to exceed the threshold P within the next T time period, a "predictive alarm" will be generated, such as: "Alarm: The GPU0 utilization of server Svr101 is expected to reach 98% within the next 15 minutes".

[0168] (4) Reporting and analysis: Resource usage forecast reports can be generated based on the forecast results to assist capacity planning.

[0169] The following example illustrates the implementation process of this embodiment: predicting GPU utilization and generating proactive alerts.

[0170] 1. Task Configuration: Users configure the task on the operations and maintenance platform UI: predict the utilization rate (gpu_utilization_percent) of GPU0 on server Svr101, predicting the next 60 minutes, with predictions every 5 minutes, using 1-minute granular data from the past 7 days as input. Configure predictive alert rules: if the predicted utilization rate exceeds 95% for 10 consecutive minutes, trigger a "Warning" level alert.

[0171] 2. Data preparation: The prediction module extracts the minute-level utilization data of Svr101-GPU0 from InfluxDB over the past 7 days and performs interpolation to fill in the gaps.

[0172] 3. Decomposition: Use a moving average with a 1440-point (1-day) window to decompose the series into trend components. and seasonal items .

[0173] 4. Normalization: BackNET processes the input seasonal item window and obtains... and Normalization is performed. HoriNET predicts the future output window. and .

[0174] 5. Prediction:

[0175] Seasonal Term: The normalized seasonal term is input into the pre-trained FEDformer model, and the output is the normalized seasonal term prediction for the next 60 points (1 hour, minute level). .

[0176] Trend Term: Predict the trend term for the next 60 points using a simple linear regression model. .

[0177] 6. Merging and Denormalization: Merging and Then use HoriNET to get and Perform inverse normalization to obtain the final predicted value Y (utilization forecast for the next 60 minutes).

[0178] 7. Application of Results:

[0179] Storage: Store the predicted sequence Y in the database.

[0180] Visualization: The UI plots a forecast curve for the next hour on the Svr101-GPU0 utilization chart.

[0181] Alarm check: The alarm module checked the predicted sequence Y. It was found that the predicted values ​​from the 25th minute to the 35th minute all exceeded 95%.

[0182] Alarm Triggered: Due to meeting the condition of "exceeding 95% for 10 consecutive minutes", the system generated a predictive alarm: "Warning: Svr101-GPU0 utilization is expected to continue to exceed 95% in 25 minutes". This alarm was sent to the operations team via Lark robot.

[0183] It should be noted that the prediction method for operation and maintenance indicators provided in this disclosure has the following beneficial effects:

[0184] a) The prediction accuracy and robustness are significantly improved. By integrating advanced processing and prediction models that combine time series decomposition, adaptive normalization, and frequency domain attention mechanisms, the model can accurately capture the complex changing patterns of key operation and maintenance indicators such as GPU utilization, network bandwidth, and RDMA congestion count in intelligent computing centers, and achieve high-precision prediction of the future trends of these indicators.

[0185] (b) It enhances the adaptability to complex data patterns. By relying on time series decomposition technology to separate the trend and seasonal components of indicators and combining adaptive normalization technology to dynamically adapt to data distribution, it can effectively handle the non-stationarity, trend and multiple seasonal characteristics of intelligent computing center operation and maintenance data, and improve the stability and environmental adaptability of model operation.

[0186] c) Achieve deep integration of predictive capabilities with the unified operation and maintenance platform of the intelligent computing center, seamlessly integrate the intelligent prediction module into the core logic layer of the operation and maintenance platform, make full use of the massive and high-frequency monitoring data collected by the platform to carry out model training and prediction, and directly implement the prediction results into the platform's functional modules such as proactive alarms, resource planning, and visualization, so as to achieve the organic integration of predictive capabilities and operation and maintenance processes.

[0187] d) Provides precise decision support for proactive operation and maintenance of intelligent computing centers. Based on the high-precision operation and maintenance indicator prediction results, it can predict potential operation and maintenance risks such as resource exhaustion and network congestion in advance. It provides reliable data basis for operation and maintenance personnel to adjust resource allocation, optimize task scheduling, and carry out preventive maintenance of equipment, and promotes the transformation of operation and maintenance mode from passive response to proactive prediction.

[0188] The operation and maintenance (O&M) indicator prediction method provided in this disclosure decomposes the original time series into trend and seasonal sequences and performs adaptive normalization on the seasonal sequence. This effectively addresses the non-stationary characteristics of O&M indicators and adapts to dynamic changes in data distribution. By inputting the normalized seasonal sequence into a frequency domain-based seasonal prediction model, it accurately captures multiple seasonal and nonlinear abrupt change patterns in O&M indicators, significantly improving prediction accuracy. By predicting the trend and seasonal sequences separately and then merging and inverse normalizing them, it achieves refined decomposition of complex patterns, avoiding mutual interference between trend and seasonal components and further improving prediction reliability. By pushing the final predicted sequence to a visualization interface and / or triggering alarms, the prediction results are deeply integrated into the O&M process, providing real-time and reliable data support and decision-making basis for proactive O&M. This solves the problems of insufficient prediction accuracy, poor handling of non-stationary data, insufficient decomposition of complex patterns, and lack of deep integration with the O&M process in existing O&M indicator prediction methods.

[0189] It is understood that the various method embodiments mentioned above in this disclosure can be combined with each other to form combined embodiments without violating the principle and logic. Due to space limitations, this disclosure will not elaborate further. Those skilled in the art will understand that in the above methods of specific implementation, the specific execution order of each step should be determined by its function and possible internal logic.

[0190] Figure 3 A block diagram of a device for predicting operation and maintenance indicators provided in an embodiment of this disclosure.

[0191] Reference Figure 3 This disclosure provides an apparatus for predicting operational and maintenance indicators, used to execute the aforementioned method for predicting operational and maintenance indicators. The apparatus includes:

[0192] The raw time series acquisition module 11 is used to extract the time series data of the preset operation and maintenance indicators of the target device from the time series database of the intelligent computing center, and use it as the raw time series.

[0193] The trend and seasonal item decomposition module 12 is connected to the original time series acquisition module 11 and is used to decompose the original time series to obtain the trend item sequence and the seasonal item sequence.

[0194] The adaptive normalization processing module 13 is connected to the trend seasonal item decomposition module 12 and is used to perform adaptive normalization processing on the seasonal item sequence to obtain the normalized seasonal item sequence.

[0195] The normalized seasonal term prediction module 14 is connected to the adaptive normalization processing module 13 and is used to input the normalized seasonal term sequence into the frequency domain-based seasonal term prediction model to obtain the normalized seasonal term prediction sequence.

[0196] The trend term prediction sequence acquisition module 15 is connected to the normalized seasonal term prediction module 14 and is used to input the trend term sequence into the trend term prediction model to obtain the trend term prediction sequence.

[0197] The final prediction sequence acquisition module 16 is connected to the trend item prediction sequence acquisition module 15. It is used to merge the normalized seasonal item prediction sequence with the trend item prediction sequence to obtain a merged prediction sequence, and to perform inverse normalization on the merged prediction sequence to obtain the final prediction sequence.

[0198] The operation and maintenance indicator prediction application module 17 is connected to the final prediction sequence acquisition module 16 and is used to push the final prediction sequence to the visualization interface for display, and / or trigger an alarm when the predicted value of the preset operation and maintenance indicator in the final prediction sequence exceeds the preset alarm threshold.

[0199] Optionally, the original time series acquisition module 11 includes:

[0200] A prediction task receiving unit is used to receive a prediction task, which includes the preset operation and maintenance indicators, the device identifier of the target device, the duration of historical data and the time granularity.

[0201] The sequence data query unit is used to query the time series database according to the preset operation and maintenance indicators, device identifier, historical data duration and time granularity, to obtain the time series data of the preset operation and maintenance indicators of the target device aggregated according to the time granularity within the historical data duration, and use it as the original time series.

[0202] Optionally, the trend seasonal item decomposition module 12 includes:

[0203] The trend term sequence acquisition unit is used to perform a moving average operation on the original time series using the moving average method to obtain the trend term sequence.

[0204] The seasonal item sequence acquisition unit is used to subtract the trend item sequence from the original time series to obtain the seasonal item sequence.

[0205] Optionally, the adaptive normalization processing module 13 includes:

[0206] The sequence window extraction unit is used to extract a continuous sequence window of fixed length from the seasonal item sequence in chronological order.

[0207] The normalization unit is used to input the data of each sequence window into BackNET, calculate the current mean and current variance of the data in the sequence window through BackNET, and normalize the data in the sequence window based on the current mean and current variance to obtain the normalized data of the sequence window.

[0208] The data splicing unit is used to splice the normalized data from all sequence windows in chronological order to obtain the normalized seasonal item sequence;

[0209] The future mean and variance prediction unit is used to input the data of each sequence window into HoriNET, and predict the future mean and future variance of the corresponding future output window through HoriNET.

[0210] Optionally, the normalized seasonal term prediction module 14 is specifically used for:

[0211] The normalized seasonal term sequence is input into the FEDformer model. The frequency domain encoder of the FEDformer model performs fast Fourier transform, frequency domain information selection and inverse Fourier transform on the normalized seasonal term sequence. Then, the self-attention and cross-attention mechanisms of the frequency domain decoder are used to calculate and output the normalized seasonal term prediction sequence.

[0212] Optionally, the final predicted sequence acquisition module 16 includes:

[0213] The prediction sequence merging unit is used to add the normalized seasonal prediction sequence and the trend prediction sequence point by point according to the timestamp to obtain the merged prediction sequence;

[0214] The inverse normalization unit is used to perform inverse normalization calculation on the merged prediction sequence based on the future mean and future variance corresponding to the future output window predicted by HoriNET, so as to restore the final prediction sequence with real physical dimensions.

[0215] Optionally, the operation and maintenance indicator prediction application module 17 includes:

[0216] A future prediction value acquisition unit is used to acquire the prediction value for each future time point in the final prediction sequence;

[0217] An alarm threshold comparison unit is used to compare the predicted value with a preset alarm threshold to determine whether there are multiple consecutive time points where the predicted value exceeds the alarm threshold.

[0218] A predictive alarm unit is used to generate predictive alarm information if it exists. The predictive alarm information includes at least one of the predicted risk occurrence time, the name of the preset operation and maintenance indicator, the target device identifier, and the suggested measures, and pushes the predictive alarm information to the operation and maintenance personnel terminal.

[0219] This disclosure also provides a prediction system for operation and maintenance indicators, including a prediction module, which is used to execute the above-described prediction method for operation and maintenance indicators.

[0220] Figure 4 This is a block diagram of an electronic device provided in an embodiment of the present disclosure.

[0221] Reference Figure 4 This disclosure provides an electronic device, which includes: at least one processor 701; at least one memory 702; and one or more I / O interfaces 703 connected between the processor 701 and the memory 702; wherein the memory 702 stores one or more computer programs that can be executed by at least one processor 701, and the one or more computer programs are executed by at least one processor 701 to enable at least one processor 701 to execute the above-mentioned method for predicting operation and maintenance indicators.

[0222] This disclosure also provides a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the aforementioned method for predicting operational indicators. The computer-readable storage medium may be volatile or non-volatile.

[0223] In summary, the method, apparatus, system, equipment, and medium for predicting operation and maintenance (O&M) indicators provided in this disclosure effectively address the non-stationary characteristics of O&M indicators and adapt to dynamic changes in data distribution by decomposing the original time series into trend and seasonal sequences and adaptively normalizing the seasonal sequences. By inputting the normalized seasonal sequences into a frequency-domain-based seasonal prediction model, multiple seasonal and nonlinear abrupt change patterns in O&M indicators can be accurately captured, significantly improving prediction accuracy. By predicting trend and seasonal indicators separately and then merging and inverse normalizing them, a refined decomposition of complex patterns is achieved, avoiding mutual interference between trend and seasonal components and further improving prediction reliability. By pushing the final predicted sequence to a visualization interface and / or triggering alarms, the prediction results are deeply integrated into the O&M process, providing real-time and reliable data support and decision-making basis for proactive O&M. This solves the problems of insufficient prediction accuracy, poor handling of non-stationary data, insufficient decomposition of complex patterns, and lack of deep integration with the O&M process in existing O&M indicator prediction methods.

[0224] Those skilled in the art will understand that all or some of the steps, systems, and apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software can be distributed on a computer-readable storage medium, which may include computer storage media (or non-transitory media) and communication media (or transient media).

[0225] As is known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable program instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), static random access memory (SRAM), flash memory or other memory technologies, portable compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, it is known to those skilled in the art that communication media typically contain computer-readable program instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0226] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0227] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0228] This disclosure has disclosed exemplary embodiments, and although specific terminology has been used, it is for general illustrative purposes only and should not be construed as limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of this disclosure as set forth by the appended claims.

Claims

1. A method for predicting operation and maintenance indicators, characterized in that, The method includes: Extract time series data of preset operation and maintenance indicators of target devices from the time series database of the intelligent computing center, and use it as the raw time series; The original time series is decomposed to obtain a trend sequence and a seasonal sequence; The seasonal term sequence is subjected to adaptive normalization to obtain a normalized seasonal term sequence; The normalized seasonal term sequence is input into the frequency domain-based seasonal term prediction model to obtain the normalized seasonal term prediction sequence. The trend term sequence is input into the trend term prediction model to obtain the trend term prediction sequence; The normalized seasonal prediction sequence is merged with the trend prediction sequence to obtain a merged prediction sequence, and the merged prediction sequence is denormalized to obtain the final prediction sequence. The final predicted sequence is pushed to the visualization interface for display, and / or an alarm is triggered when the predicted value of the preset operation and maintenance indicator in the final predicted sequence exceeds the preset alarm threshold.

2. The method according to claim 1, characterized in that, The extraction of time-series data of preset operation and maintenance indicators of target devices from the time-series database of the intelligent computing center, and using it as the original time series, specifically includes: Receive a prediction task, the prediction task including the preset operation and maintenance indicators, the device identifier of the target device, the duration of historical data and the time granularity; Based on the preset operation and maintenance indicators, device identifier, historical data duration and time granularity, the time series database is queried to obtain the time series data of the preset operation and maintenance indicators of the target device aggregated according to the time granularity within the historical data duration, and this data is used as the original time series.

3. The method according to claim 1, characterized in that, The decomposition of the original time series to obtain a trend sequence and a seasonal sequence specifically includes: The original time series is subjected to a moving average operation using the moving average method to obtain the trend term series. The seasonal sequence is obtained by subtracting the trend sequence from the original time series.

4. The method according to claim 1, characterized in that, The adaptive normalization process performed on the seasonal term sequence to obtain the normalized seasonal term sequence specifically includes: Extract fixed-length continuous sequence windows sequentially from the seasonal item sequence in chronological order; For each sequence window, the data of the sequence window is input into BackNET, the current mean and current variance of the data in the sequence window are calculated by BackNET, and the data in the sequence window are normalized based on the current mean and current variance to obtain the normalized data of the sequence window; By concatenating the normalized data from all sequence windows in chronological order, a normalized seasonal sequence is obtained. For each sequence window, the data of the sequence window is input into HoriNET, and the future mean and future variance corresponding to the future output window are predicted by HoriNET.

5. The method according to claim 1, characterized in that, The step of inputting the normalized seasonal term sequence into the frequency domain-based seasonal term prediction model to obtain the normalized seasonal term prediction sequence specifically includes: The normalized seasonal term sequence is input into the FEDformer model. The frequency domain encoder of the FEDformer model performs fast Fourier transform, frequency domain information selection and inverse Fourier transform on the normalized seasonal term sequence. Then, the self-attention and cross-attention mechanisms of the frequency domain decoder are used to calculate and output the normalized seasonal term prediction sequence.

6. The method according to claim 4, characterized in that, The step of merging the normalized seasonal forecast sequence with the trend forecast sequence to obtain a merged forecast sequence, and then performing inverse normalization on the merged forecast sequence to obtain the final forecast sequence, specifically includes: The normalized seasonal prediction sequence and the trend prediction sequence are added point by point according to the timestamp to obtain the merged prediction sequence; Based on the future mean and future variance corresponding to the future output window predicted by HoriNET, the merged prediction sequence is denormalized to obtain the final prediction sequence with real physical dimensions.

7. The method according to claim 1, characterized in that, The alarm is triggered when the predicted value of the preset operation and maintenance indicator in the final prediction sequence exceeds the preset alarm threshold, specifically including: Obtain the predicted value for each future time point in the final prediction sequence; The predicted value is compared with a preset alarm threshold to determine whether there are multiple consecutive time points where the predicted value exceeds the alarm threshold. If present, a predictive alarm is generated, which includes at least one of the predicted risk occurrence time, the name of the preset operation and maintenance indicator, the target device identifier, and suggested measures, and the predictive alarm is pushed to the operation and maintenance personnel's terminal.

8. A device for predicting operation and maintenance indicators, characterized in that, The device includes: The raw time series acquisition module is used to extract time series data of preset operation and maintenance indicators of target devices from the time series database of the intelligent computing center and use it as the raw time series. The trend and seasonal item decomposition module is connected to the original time series acquisition module and is used to decompose the original time series to obtain the trend item sequence and the seasonal item sequence. An adaptive normalization processing module, connected to the trend seasonal item decomposition module, is used to perform adaptive normalization processing on the seasonal item sequence to obtain a normalized seasonal item sequence. The normalized seasonal term prediction module is connected to the adaptive normalization processing module and is used to input the normalized seasonal term sequence into the frequency domain-based seasonal term prediction model to obtain the normalized seasonal term prediction sequence. The trend term prediction sequence acquisition module is connected to the normalized seasonal term prediction module and is used to input the trend term sequence into the trend term prediction model to obtain the trend term prediction sequence. The final prediction sequence acquisition module is connected to the trend item prediction sequence acquisition module. It is used to merge the normalized seasonal item prediction sequence with the trend item prediction sequence to obtain a merged prediction sequence, and to perform inverse normalization on the merged prediction sequence to obtain the final prediction sequence. The operation and maintenance indicator prediction application module is connected to the final prediction sequence acquisition module. It is used to push the final prediction sequence to the visualization interface for display, and / or trigger an alarm when the predicted value of the preset operation and maintenance indicator in the final prediction sequence exceeds the preset alarm threshold.

9. A prediction system for operation and maintenance indicators, characterized in that, It includes a prediction module, which is used to perform the prediction method for operation and maintenance indicators as described in any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores one or more computer programs that can be executed by the at least one processor, and the one or more computer programs are executed by the at least one processor to enable the at least one processor to perform the prediction method of operation and maintenance indicators as described in any one of claims 1-7.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for predicting operation and maintenance indicators as described in any one of claims 1-7.

Images