Intelligent fault-tolerant method for high-trust computer systems based on multi-modal failure prediction

Abstract

This invention relates to the field of electronic digital data processing technology, specifically to an intelligent fault-tolerant method for highly reliable computer systems based on multimodal fault prediction. The method includes: constructing an initial mutual information baseline matrix containing multiple operating condition categories based on the computer system's operational data during historical fault-free periods; dynamically determining baseline update parameters based on statistical analysis of data windows during real-time operation of the computer system; dynamically updating the mutual information reference values ​​in the initial mutual information baseline matrix according to the baseline update parameters; monitoring the computer system's operating status based on the dynamically updated mutual information reference values ​​and generating an early warning signal when an anomaly is detected; and executing preset resource control operations in response to the early warning signal to achieve system fault tolerance. This invention achieves accurate detection of early-stage latent faults in computer systems, enhancing the service continuity, operational reliability, and overall fault tolerance adaptability of highly reliable computer systems.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and specifically to a highly reliable intelligent fault-tolerant method for computer systems based on multimodal fault prediction. Background Technology

[0002] With the deep application of computer systems in key fields such as finance and energy, high reliability of their operation has become a core requirement. Network resource allocation, as a core support link for system operation, is directly related to key performance indicators such as bandwidth usage, server load, and data transmission latency. However, during system operation, various resource allocation-related anomalies are prone to occur due to factors such as load fluctuations and application scenario switching, posing challenges to system reliability and service continuity. Effective fault prediction and intelligent fault-tolerant solutions are urgently needed to provide assurance.

[0003] Current methods for handling network resource allocation failures in computer systems mainly rely on preset fixed thresholds and static rules to achieve fault detection and fault tolerance. These methods generally suffer from insufficient sensitivity in detecting early-stage hidden faults, delayed early warnings, and a lack of precision in fault tolerance responses. They fail to meet the actual needs of highly reliable computer systems for early fault prediction, rapid fault location, and efficient fault tolerance, thus affecting the overall stability and reliability of the system. Summary of the Invention

[0004] To address the technical problem of poor early warning and fault-tolerant response due to the difficulty in accurately detecting early hidden faults in computer systems, the present invention aims to provide a highly reliable intelligent fault-tolerant method for computer systems based on multimodal fault prediction. The specific technical solution adopted is as follows: Firstly, a highly reliable intelligent fault-tolerant method for computer systems based on multimodal fault prediction is provided. This method includes: constructing an initial mutual information baseline matrix containing multiple operating condition categories based on the computer system's historical fault-free operating data; storing mutual information reference values ​​for each key mode pair under different operating condition categories; dynamically determining baseline update parameters based on statistical analysis of data windows during real-time operation of the computer system; dynamically updating the mutual information reference values ​​in the initial mutual information baseline matrix according to the baseline update parameters; monitoring the computer system's operating status based on the dynamically updated mutual information reference values, and generating an early warning signal when an anomaly is detected; and executing preset resource control operations in response to the early warning signal to achieve system fault tolerance.

[0005] In one possible design, an initial mutual information baseline matrix containing multiple operating condition categories is constructed, including: dividing the operational data using a sliding window to generate a continuous data window containing all operating condition categories; discretizing the modal data within each data window, and determining the mutual information value of each key modal pair based on the frequency of occurrence of individual modal data categories and the frequency of occurrence of combinations of modal data categories in each key modal pair; multiple key modal pairs include modal pairs consisting of bandwidth utilization and end-to-end latency, modal pairs consisting of end-to-end latency and packet loss rate, modal pairs consisting of active connection count and response time, and modal pairs consisting of traffic type and queue depth; determining the mutual information value of each data window within each data window. The local variance of the mutual information values ​​of all key modal pairs is used to identify data windows with local variances below a preset variance threshold as candidate windows. Candidate windows are grouped into window blocks according to a preset number and time order. The global variance of each window block is determined based on the mean mutual information value of each candidate window included in each window block, and the window block corresponding to the minimum global variance is identified as the initial healthy window. For each operating condition category, the mean mutual information value of the data windows of each key modal pair within the initial healthy window in that operating condition category is determined as the mutual information reference value of each key modal pair under that operating condition category. The mutual information reference values ​​of each key modal pair under all operating condition categories are integrated to construct an initial mutual information baseline matrix.

[0006] In one possible design, a sliding window is used to divide the operational data to generate a continuous data window containing all operating condition categories. This includes: acquiring operational data for a first preset duration; dividing the operational data into continuous data windows using a sliding window based on a preset window length and a preset window overlap rate; determining the operating condition category corresponding to each data window based on the load characteristics of each data window, where the load characteristics include traffic rate characteristics, application protocol characteristics, and user concurrency characteristics; and updating the continuous data windows based on operational data for the first preset duration and operational data for the second preset duration, until the operating condition categories corresponding to all data windows in the continuous data window contain all operating condition categories.

[0007] In one possible design, the operating condition category corresponding to each data window is determined based on the load characteristics of each data window, including: for each data window, determining the load level of the data window based on the traffic rate characteristics and user concurrency characteristics of the data window; determining the application scenario of the data window based on the application protocol characteristics of the data window; and combining the load level and the application scenario to obtain the operating condition category corresponding to the data window.

[0008] In one possible design, baseline update parameters are dynamically determined based on statistical analysis of the data windows, including: determining the number of times the operating condition category changes within the previous baseline update cycle based on the operating condition category of each data window within the previous baseline update cycle; determining the operating condition switching frequency within the previous baseline update cycle based on the number of changes and the duration of the previous baseline update cycle; determining the current baseline statistical duration based on the operating condition switching frequency and the duration of the previous baseline update cycle; and determining the larger value between the current baseline statistical duration and a preset minimum cycle as the current baseline update cycle. The current baseline update cycle is used to indicate the update of the mutual information reference value in the initial mutual information baseline matrix after the current baseline update cycle ends.

[0009] In one possible design, the baseline update parameters are dynamically determined based on statistical analysis of the data windows. This also includes: determining the operating condition category of the health data windows in the previous baseline update cycle, where the health data windows are data windows without anomalies; and for each operating condition category, determining the smoothing factor corresponding to the operating condition category based on the cumulative number of health data windows corresponding to the operating condition category.

[0010] In one possible design, the mutual information reference values ​​in the initial mutual information baseline matrix are dynamically updated according to the baseline update parameters. This includes: after the previous baseline update cycle ends, for each working condition category, the mutual information reference values ​​of each key mode pair under the working condition category in the initial mutual information baseline matrix are updated and replaced according to the smoothing factor corresponding to the working condition category.

[0011] In one possible design, the operating status of the computer system is monitored based on dynamically updated mutual information reference values, and an early warning signal is generated when an anomaly is detected. This includes: determining the real-time mutual information value of each key modality pair in the current data window; for each key modality pair, determining the relative deviation between the real-time mutual information value of the key modality pair and the corresponding updated mutual information reference value; determining the mean of the relative deviations corresponding to at least one data window containing the current data window as the correlation decay factor corresponding to the current data window, wherein the data windows in at least one data window are continuous and the operating condition category is the current operating condition category, and the current operating condition category is the operating condition category corresponding to the current data window; determining that the current data window is abnormal when the correlation decay factor is greater than the adaptive threshold corresponding to the current operating condition category; and determining that the computer system is operating abnormally when multiple consecutive data windows are abnormal, and generating an early warning signal.

[0012] In one possible design, determining the adaptive threshold includes: for each key mode pair under each working condition category, if the number of determined correlation attenuation factors is less than a preset number, determining a preset fixed value as the adaptive threshold; if the number of determined correlation attenuation factors is greater than a preset number, determining the adaptive threshold based on the preset percentile value of the numerical distribution of correlation attenuation factors less than the preset fixed value, thereby obtaining the adaptive thresholds corresponding to each key mode pair under all working condition categories.

[0013] In one possible design, generating an early warning signal includes: identifying target key modal pairs, where the target key modal pairs are key modal pairs whose correlation decay factors are greater than the corresponding adaptive thresholds in multiple consecutive data windows; when the target key modal pairs are modal pairs consisting of bandwidth utilization and end-to-end latency, the early warning signal is used to indicate abnormal bandwidth allocation; when the target key modal pairs are modal pairs consisting of end-to-end latency and packet loss rate, the early warning signal is used to indicate network congestion; when the target key modal pairs are modal pairs consisting of active connection count and response time, the early warning signal is used to indicate server overload; and when the target key modal pairs are modal pairs consisting of traffic type and queue depth, the early warning signal is used to indicate traffic resource preemption.

[0014] The present invention has the following beneficial effects: In the intelligent fault-tolerant method for high-reliability computer systems based on multimodal fault prediction provided by this invention, a multi-condition mutual information baseline matrix based on historical fault-free data is constructed. Based on this matrix, update parameters are dynamically determined by combining real-time operational condition switching dynamics and the accumulated level of health awareness, thus achieving adaptive dynamic maintenance of the mutual information baseline. This dynamically updated baseline is then used to monitor the system in real time. Early warnings are generated by quantifying the degree of correlation decay and combining adaptive thresholds with a continuous judgment mechanism, ultimately triggering targeted resource regulation. Through a closed loop of "dynamic baseline construction - adaptive update - intelligent monitoring - automatic fault tolerance," the method effectively overcomes the problems of insensitivity to early latent faults and delayed early warning in existing static threshold or rule-based methods. It significantly improves the early warning capability and prediction accuracy for resource allocation faults such as bandwidth congestion, server overload, and misconfiguration. Simultaneously, the system can automatically execute precise resource regulation operations, transforming traditional passive fault response into proactive fault prevention and self-healing. This greatly enhances the service continuity, operational reliability, and overall fault tolerance adaptability of high-reliability computer systems in complex and dynamic network environments. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages 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.

[0016] Figure 1 This is a flowchart illustrating a highly reliable intelligent fault-tolerant computer system method based on multimodal fault prediction, provided in one embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a highly reliable intelligent fault-tolerant computer system based on multimodal fault prediction, provided as an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a highly reliable intelligent fault-tolerant computer system based on multimodal fault prediction proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] In embodiments of the present invention, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0019] In the description of this invention, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" and "more than one" refer to two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The following description, in conjunction with the accompanying drawings, details the specific scheme of the intelligent fault-tolerant method for highly reliable computer systems based on multimodal fault prediction provided by this invention.

[0022] Please see Figure 1 The diagram illustrates a flowchart of a highly reliable intelligent fault-tolerant computer system method based on multimodal fault prediction provided by an embodiment of the present invention, including the following steps S101-S105.

[0023] S101. Based on the operating data of the computer system during historical fault-free periods, construct an initial mutual information baseline matrix containing multiple operating condition categories.

[0024] The initial mutual information baseline matrix is ​​used to store the mutual information reference values ​​of each key mode pair under different operating conditions.

[0025] As one possible approach, a sliding window is used to divide the computer system's operating data during historical fault-free periods, generating a continuous data window.

[0026] In some embodiments, operational data from historical computer system operation data spanning a first preset duration (e.g., not less than 90 days) with no failures is filtered to ensure sufficient statistical representativeness. Then, multiple sets of time-series operational data for key modal pairs are collected from these data at a preset sampling frequency (1-10Hz). The key modal pairs include a first modal pair consisting of bandwidth utilization and end-to-end latency, a second modal pair consisting of end-to-end latency and packet loss rate, a third modal pair consisting of active connection count and response time, and a fourth modal pair consisting of traffic type and queue depth. Subsequently, a sliding window method is used to segment the operational data. Based on a preset window length (e.g., 10 minutes) and a preset window overlap rate (e.g., 30%), the operational data is divided into continuous, temporally overlapping, standardized data windows.

[0027] Secondly, based on the load characteristics of each data window, the corresponding operating condition category for each data window is determined. These load characteristics include traffic rate characteristics (such as the average bandwidth utilization within the data window, denoted as...). Application protocol characteristics and user concurrency characteristics (such as the maximum number of active connections within a data window, denoted as...) The application protocol characteristics can be based on the packet port number or protocol identifier, and can be used to count the dominant (e.g., most frequent) application protocol types within the data window, such as Real-Time Streaming Protocol (RTSP), Hypertext Transfer Protocol (HTTP) / Hypertext Transfer Protocol Secure (HTTPS), File Transfer Protocol (FTP), Transmission Control Protocol (TCP) / User Datagram Protocol (UDP).

[0028] In some embodiments, for each data window, the load level of the data window is determined based on the data window's traffic rate characteristics and user concurrency characteristics. The maximum theoretical bandwidth provided for network hardware, if it meets the following conditions ,and If it meets the requirements, it is determined to be a low load level; if it meets the requirements... ,or If it is, then it is determined to be a medium load level; if in ,or If the load level is low, it is classified as a high load level. Further, based on the application protocol characteristics of the data window, the application scenario is determined. For example, if the application protocol type is RTSP, the application scenario is determined to be video streaming; if the application protocol type is HTTP / HTTPS, the application scenario is determined to be web browsing. Finally, the load level and application scenario are combined to obtain the corresponding operating condition category for the data window, such as low load - video streaming, medium load - web browsing, or high load - file download, etc.

[0029] Furthermore, it is determined whether the operating condition categories corresponding to all data windows in the continuous data window include all operating condition categories. If they do, subsequent processing is performed based on the current continuous data window. If they do not include all operating condition categories, the system further filters the historical operating data of the computer system for fault-free periods with a time span of a second preset duration (e.g., 10 days, 20 days, etc.). This second preset duration can be a completely different period from the first preset duration, and there can also be an overlapping data window (e.g., the overlap ratio between the last data window in the second preset duration and the first data window in the first preset duration is a preset window overlap rate, or the overlap ratio between the last data window in the first preset duration and the first data window in the second preset duration is a preset window overlap rate). Then, based on the operating data of the first preset duration and the operating data of the second preset duration, the continuous data window is updated, and it is determined whether the operating condition categories corresponding to all data windows in the current continuous data window include all operating condition categories. If they do, subsequent processing is performed based on the current continuous data window; otherwise, the continuous data window is updated again until the operating condition categories corresponding to all data windows in the continuous data window include all operating condition categories.

[0030] Furthermore, the mutual information values ​​of each key modality pair within each data window are determined.

[0031] In some embodiments, for any key mode pair, with This represents two related operational metrics, specifically two consecutive operational metrics in a network mode pair. The values ​​are discretized, for example, using the equal-frequency binning method, dividing the range of each operating indicator into segments based on its frequency of occurrence in the historical data of that operating condition. There are several levels, which can be determined based on experience. The value can be 3, 4, or 5, thus converting continuous values ​​into discrete categories. and For any key modality pair within any data window. , recorded as Modality The value falls into a specific discrete category frequency Modality The value falls into a specific discrete category frequency and the combination of the two Frequency of occurrence The total number of sample points in this data window is Based on this, the corresponding probability distribution is calculated: , , Then, substitute the values ​​into the mutual information calculation formula. , That is, in the data window Key mode pairs within The mutual information value is obtained. Based on this, the mutual information value of each key modality pair within each data window can be obtained.

[0032] It should be noted that, if or If the value is 0, then its joint probability with any other class is... The value must also be 0, and the contribution of this combination to the total mutual information is considered zero. In actual calculations, before using the formula for calculating mutual information, the value is... as well as Check if it is 0. or If the value is 0, then the calculation of this type of term is ignored.

[0033] Furthermore, the variance of the mutual information values ​​of all key modal pairs within each data window is determined, denoted as the local variance, and its formula is expressed as follows: In the formula, For data window The corresponding local variance, For data window Key mode pairs within mutual information value, For data window The mean mutual information value of all key mode pairs within the group. To determine the number of key mode pairs, this embodiment of the invention uses four key mode pairs as an example. Local variance is used to characterize the fluctuation of the mutual information value of each key mode pair within a single data window. Data windows with local variance below a preset variance threshold are then identified as candidate windows.

[0034] In some embodiments, after obtaining the local variance of each data window, all local variances are sorted in ascending order, and the 30th percentile is selected as the preset variance threshold. That is, from the continuous data windows, the top 30% of data windows with the lowest local variance are selected as candidate windows.

[0035] Furthermore, candidate windows are grouped into window blocks according to a preset number (20 or 30, empirical values) and time order. The global variance of each window block is determined based on the mean mutual information value of each candidate window included in that block. The formula is expressed as follows: In the formula, For window block The corresponding global variance, For window block Internal Data Window The mean mutual information value of all key mode pairs within the group. For window block The mean of the mutual information values ​​of all data windows within the window. For window block The number of internal data windows and the global variance are used to characterize the overall stability of the modalities within a window block to the associated states. The window block corresponding to the minimum global variance is then determined as the initial healthy window, which corresponds to the most stable healthy runtime segment of the computer system.

[0036] Furthermore, for each operating condition category, the average mutual information value of each key modality pair within the initial health window in the data window of the operating condition category is determined as the mutual information reference value of each key modality pair under the operating condition category.

[0037] In some embodiments, the mutual information values ​​of each key modality pair in all data windows within the initial health window are extracted, these mutual information values ​​are grouped according to the operating condition category, the mean value of the mutual information value of each key modality pair under each operating condition category is calculated, and the mean value is determined as the mutual information reference value of the corresponding key modality pair under that operating condition category.

[0038] Finally, the mutual information reference values ​​of each key modality pair under all operating conditions are integrated to construct an initial mutual information baseline matrix. The row dimension of this matrix is ​​the operating condition category, and the column dimension is the key modality pair. Each cell in the matrix stores the mutual information reference value of the corresponding key modality pair under the corresponding operating condition category, forming a complete initial mutual information baseline covering all operating condition categories.

[0039] S102. During the real-time operation of the computer system, the baseline update parameters are dynamically determined based on the statistical analysis of the data window.

[0040] It should be noted that the method of dividing data windows and determining the corresponding operating condition category of data windows during the real-time operation of the computer system is consistent with the method of processing the operating data of the computer system during the historical fault-free period in step S101 above. Data windows are divided in real time during operation, and the corresponding operating condition category of data windows is determined.

[0041] The baseline update parameters include the baseline update period and the smoothing factor. The baseline update period determines the applicable time range of the baseline, and the smoothing factor determines the accuracy of the mutual information baseline judgment result.

[0042] As one possible approach, in order to accurately predict the baseline and establish a highly reliable computer system, it is first necessary to perform statistical analysis on the data windows during the real-time operation of the computer system to determine a reasonable baseline update cycle and smoothing factor.

[0043] In some embodiments, determining a reasonable baseline update cycle includes: starting from the time corresponding to the initial health window, setting the duration of the first baseline update cycle to a preset initial duration, denoted as . For example, within a 1-hour period, based on the mutual information reference values ​​of each key mode pair under different operating condition categories stored in the initial mutual information baseline matrix, the real-time data window is monitored, and the operating condition category of each data window is statistically analyzed. This allows for the determination of healthy data windows that are not detected as abnormal, and the number of times the operating condition category changes within the baseline update cycle is recorded as follows: The operating condition switching frequency within this baseline update cycle can be expressed by the formula. Obtain the operating condition switching frequency. A higher value indicates a more unstable system operating mode and more frequent changes. After the first baseline update cycle ends, the duration of the second baseline update cycle is determined based on the duration of the first baseline update cycle and the frequency of operating condition switching within the first baseline update cycle, and so on. During the operation of the computer system, the baseline update cycle is dynamically updated, and subsequent baseline update cycles depend on the operating condition stability of the previous baseline update cycle.

[0044] Specifically, for the current baseline update cycle during computer system operation, excluding the first baseline update cycle. The duration of the baseline update cycle It can be determined using the following formula.

[0045] In the formula, A minimum preset period is used to prevent system oscillation caused by a period that is too short. For the current baseline update cycle The duration of the previous baseline update cycle, For the current baseline update cycle The frequency of operating condition switching within the previous baseline update cycle. The current baseline statistical duration is determined based on the operating condition switching frequency and duration of the previous baseline update cycle, and is greater than... In this case, it is determined as the current baseline update cycle. The more frequently the operating conditions are switched ( The larger the value, the more it is calculated. The shorter the interval, the more timely the baseline update, in order to adapt to the rapidly changing environment; conversely, the longer the baseline update cycle, the more time-consuming the unnecessary computational overhead.

[0046] This enables a cyclical recursion of the baseline update cycle, and the calculated... As the duration of the current baseline update cycle, when After completion, repeat the above process and compile statistics. Operating condition switching frequency within a time period And calculate the next baseline update cycle accordingly. This cycle allows the baseline update period to continuously and adaptively track the long-term dynamic characteristics of the computer system's operating mode.

[0047] In some embodiments, within the first baseline update cycle starting at the time corresponding to the initial health window, the real-time data window is monitored based on the mutual information reference values ​​of each key modality pair under different operating condition categories stored in the initial mutual information baseline matrix. In the second baseline update cycle, and any subsequent baseline update cycle, the operating condition category of the health data window in the previous baseline update cycle is determined. For each operating condition category, a smoothing factor corresponding to the operating condition category is determined based on the cumulative number of health data windows corresponding to that category. This smoothing factor is used to update the mutual information reference values ​​of each modality pair under the corresponding operating condition category stored in the initial mutual information baseline matrix used in the previous baseline update cycle.

[0048] The formula for calculating the initial smoothing factor for each health data window is as follows: In the formula, As a dynamic smoothing factor, This is the cumulative number of health data windows preceding the current health data window. To quickly learn the threshold, a value of 30 could be used, for example. This is the threshold for full maturity, for example, it can be set to 100. This is the maximum dynamic smoothing factor, for example, it can take a value of 0.15. This is the minimum dynamic smoothing factor, for example, it can be 0.03.

[0049] Based on this, we can obtain the following for the working condition categories. In the previous baseline update period, a total of The initial smoothing factor for each of the health data windows. to determine The mean of the initial smoothing factor corresponding to each health data window is used as the operating condition category in the previous baseline update cycle. The corresponding smoothing factor is denoted as Thus, by using the same method, the smoothing factor corresponding to each working condition category in the previous baseline update cycle can be obtained.

[0050] It should be noted that if no data window for a certain operating condition category appears in the previous baseline update cycle, the smoothing factor corresponding to that operating condition category will be set to zero.

[0051] Understandably, through the above process, two key statistical analyses can be performed in parallel during the real-time operation of the computer system: first, analyzing the operating condition change characteristics (operating condition switching frequency) reflected by the data window sequence to determine when to update (baseline update cycle); second, analyzing the health awareness accumulation characteristics reflected by the data windows (the number of health windows corresponding to each operating condition category) to determine how to update (smoothing factor). These two dynamically determined parameters together constitute the baseline update parameters, providing a precise control basis for subsequent steps to achieve adaptive and intelligent maintenance of the mutual information baseline.

[0052] S103. Based on the baseline update parameters, dynamically update the mutual information reference values ​​in the initial mutual information baseline matrix.

[0053] As one possible implementation, after the previous baseline update cycle ends, for each working condition category, the mutual information reference values ​​of each key mode pair under the working condition category in the initial mutual information baseline matrix are updated and replaced according to the smoothing factor corresponding to the working condition category.

[0054] In some embodiments, for operating condition categories Key mode pairs Mutual information reference value in the initial mutual information baseline matrix After the previous baseline update cycle ends, it can be updated using the following formula.

[0055] In the formula, For the updated working condition category Key mode pairs The mutual information reference value is calculated, and after this value is obtained, the corresponding mutual information reference value stored in the initial mutual information baseline matrix is ​​replaced and updated. Operating condition categories stored in the initial mutual information baseline matrix Key mode pairs The mutual information reference value, during the first update, is based on the operating condition category determined in step S101 above and stored in the initial mutual information baseline matrix. Key mode pairs The mutual information reference value, in non-first updates, is the operating condition category in the initial mutual information baseline matrix after the last update. Key mode pairs Mutual information reference value; The operating condition category is as determined in step S102 above. The smoothing factor corresponding to the previous baseline update cycle. Operating condition category within the previous baseline update cycle Key modality pairs in each corresponding health data window The mean of the mutual information values.

[0056] Based on this, all mutual information reference values ​​in the initial mutual information baseline matrix are updated. The updated initial mutual information baseline matrix is ​​used to monitor the data window in the current baseline update cycle in real time. At the same time, after the current baseline update cycle ends, a new round of iterative updates is performed on the initial mutual information baseline matrix and the smoothing factor corresponding to each working condition category in the current baseline update cycle.

[0057] Understandably, through the above process, the initial mutual information baseline matrix transforms from a static reference benchmark based on historical data into a dynamic benchmark capable of adaptively predicting future health status. By dynamically adjusting the smoothing factor, it indirectly reflects the long-term stability learning outcomes of the system's operation. This avoids potential anomalies introduced by directly using real-time data while ensuring the baseline adapts to the system's normal performance evolution, thus maintaining its sensitivity and reliability as an early warning benchmark for faults.

[0058] S104. Based on the dynamically updated mutual information reference value, monitor the operating status of the computer system and generate an early warning signal when an anomaly is detected.

[0059] As one possible implementation, firstly, during the real-time operation of the computer system, data windows are divided according to the same acquisition frequency, window length, and overlap rate parameters as those used to construct the initial mutual information baseline matrix. For each acquired real-time data window, the same core load feature extraction method and operating condition category combination rules are used to determine the current operating condition category corresponding to the current data window, ensuring logical consistency between operating condition judgment and baseline construction and updating. Simultaneously, using the same discretization processing method, frequency statistics method, probability calculation method, and mutual information formula derivation as the initial baseline construction, the real-time mutual information values ​​of each key modality pair within the current data window are obtained, denoted as... ,in, For the current operating condition category, This refers to the key mode pair sequence number.

[0060] Secondly, the current operating condition category is extracted from the dynamically updated initial mutual information baseline matrix. The following key mode pairs Corresponding updated mutual information reference value , The previous baseline update cycle number is used as the reference value, and the relative deviation is calculated based on this mutual information reference value and the real-time mutual information value.

[0061] In some embodiments, the formula for calculating the relative deviation is as follows: In the formula, Current operating condition category Next key mode pair The corresponding relative deviation, The current operating condition category in the dynamically updated initial mutual information baseline matrix. Next key mode pair Mutual information reference value, It is a very small positive number, and an empirical value of 0.001 can be taken to prevent the denominator from being zero.

[0062] Subsequently, the correlation decay factor corresponding to the current data window is calculated to filter out instantaneous and accidental data fluctuations and highlight the continuous trend of correlation strength. At least one data window (e.g., 4-6) containing the current data window is selected. Within this selection, the data windows are consecutive and all operate under the current operating condition category. The current data window is the last data window in chronological order within this selection; that is, the selection includes the current data window and several preceding data windows with the same operating condition category. If no preceding data window has the same operating condition category as the current data window, then the selection includes only the current data window.

[0063] Then, the relative deviations of the same key mode pair corresponding to each data window are extracted from at least one data window, and their arithmetic mean is calculated as the correlation decay factor corresponding to the current data window. This indicates that the current data window is in the current operating condition category. Next key mode pair The corresponding associated decay factor.

[0064] Furthermore, the correlation decay factor of each key mode pair corresponding to the current data window is... , and the current operating condition category The corresponding adaptive thresholds are compared.

[0065] The adaptive threshold can be determined in the following way.

[0066] For each key mode pair under each operating condition category, if the number of determined correlation attenuation factors is less than a preset number (e.g., 100), a preset fixed value (denoted as ) will be applied. For example, an empirical value of 0.2 can be used to determine the adaptive threshold. If the number of identified correlation decay factors exceeds a preset number, an adaptive threshold is determined based on a preset percentile (e.g., 95%) of the distribution of correlation decay factors that are less than a preset fixed value. This allows us to obtain the adaptive thresholds for each key mode pair under all operating conditions, ensuring that the adaptive thresholds match the healthy operating characteristics of each key mode pair under the current operating condition, and reducing the false alarm rate.

[0067] Furthermore, in the correlation decay factor Greater than the adaptive threshold corresponding to the current working condition category In this case, it indicates that the correlation strength of the key mode pair is continuously weakening, suggesting a potential system fault and confirming an anomaly in the current data window. If If the current data window is functioning normally, it is recorded as a healthy data window and will continue to be monitored in real time.

[0068] Finally, to avoid false alarms caused by a single, accidental anomaly, the presence of anomalies in the computer system is determined based on the number of consecutive anomaly windows. If multiple consecutive data windows (e.g., three consecutive data windows) show anomalies... If the number of consecutive abnormal data windows does not reach three, the anomaly count is reset, and monitoring of subsequent data windows continues.

[0069] In some embodiments, for anomalies existing in multiple consecutive data windows, the data window can be determined as an anomaly if the correlation decay factor of the same key mode pair under the same operating condition category is greater than its corresponding adaptive threshold, or if the data window is determined as an anomaly if the correlation decay factor of different operating condition categories or different key mode pairs is greater than its corresponding adaptive threshold.

[0070] In some embodiments, when a system malfunction is determined, a targeted early warning signal is generated: First, the target critical modal pair is identified, which is a critical modal pair in which the correlation decay factor in multiple consecutive abnormal data windows is greater than the corresponding adaptive threshold. Based on the type of the target critical modal pair, the direction of the early warning signal is determined. If the target critical modal pair is a modal pair consisting of bandwidth utilization and end-to-end latency, the early warning signal indicates abnormal bandwidth allocation; if it is a modal pair consisting of end-to-end latency and packet loss rate, the early warning signal indicates network congestion; if it is a modal pair consisting of active connection count and response time, the early warning signal indicates server overload; if it is a modal pair consisting of traffic type and queue depth, the early warning signal indicates traffic resource preemption. The abnormal condition category, the target critical modal pair, and the fault type indication information are integrated into the early warning signal to ensure that subsequent fault-tolerant operations are clearly targeted.

[0071] S105. In response to the warning signal, execute the preset resource regulation operation to achieve system fault tolerance.

[0072] As one possible implementation, when a warning signal indicates abnormal bandwidth allocation, dynamic adjustment of bandwidth quotas can be performed. Optionally, this could involve first using traffic identification technology to distinguish between critical applications (such as real-time interactive services) and non-critical applications (such as file download services), reserving a preset proportion of bandwidth resources for critical applications (the preset proportion is based on the importance of the service), while limiting the maximum bandwidth usage of non-critical applications to prevent them from preempting critical resources; during the adjustment process, the correlation strength deviation between bandwidth utilization and end-to-end latency is monitored in real time until the deviation falls below a preset deviation value.

[0073] When a warning signal indicates network congestion, perform rerouting or enable traffic priority scheduling. Optionally, rerouting may involve using a route probing algorithm to select alternative routes with lower latency and less load, automatically switching traffic transmission paths, and updating the routing table to ensure subsequent traffic is transmitted along the optimized path. Traffic priority scheduling may involve assigning priorities to data traffic based on application protocols (e.g., real-time interactive traffic has the highest priority, video streams the next, and file downloads the lowest), prioritizing the transmission of high-priority traffic to alleviate queue congestion.

[0074] When a warning signal indicates server overload, perform load balancing adjustments or connection migrations. Optionally, perform load balancing adjustments, including adjusting the scheduling strategy of the load balancing algorithm (e.g., changing from round-robin scheduling to weighted scheduling based on server load), collecting CPU utilization, memory usage, and active connection counts of each server in real time, and dynamically allocating new connections to servers with lighter loads to balance the cluster load; perform connection migrations, including migrating some non-critical application active connections from the currently overloaded server to a less loaded standby server, maintaining connection session continuity during the migration process to avoid service interruptions.

[0075] When a warning signal indicates that traffic resources are being preempted, traffic control operations based on application protocols are performed. Optionally, these operations may include using data compression techniques to reduce bandwidth requirements for specific types of traffic (such as video streams) that are preempting resources, or enabling latency-tolerant transmission mechanisms to temporarily buffer such traffic data when network resources are scarce and transmit it when resources become available. Simultaneously, an upper limit on the queue depth of such traffic is set to prevent it from consuming excessive network device buffer resources and to ensure the normal transmission of other types of traffic.

[0076] Understandably, in the intelligent fault-tolerant method for high-reliability computer systems based on multimodal fault prediction provided in this invention, a multi-condition mutual information baseline matrix based on historical fault-free data is constructed. Based on this matrix, update parameters are dynamically determined by combining real-time operational condition switching dynamics and the accumulated level of health awareness, thus achieving adaptive dynamic maintenance of the mutual information baseline. This dynamically updated baseline is then used to monitor the system in real time. Early warnings are generated by quantifying the degree of correlation decay and combining adaptive thresholds with a continuous judgment mechanism, ultimately triggering targeted resource regulation. Through the closed loop of "dynamic baseline construction - adaptive update - intelligent monitoring - automatic fault tolerance," the method effectively overcomes the problems of insensitivity to early latent faults and delayed early warning in existing static threshold or rule-based methods. It significantly improves the early warning capability and prediction accuracy for resource allocation faults such as bandwidth congestion, server overload, and misconfiguration. Simultaneously, the system can automatically execute precise resource regulation operations, transforming traditional passive fault response into proactive fault prevention and self-healing. This greatly enhances the service continuity, operational reliability, and overall fault tolerance adaptability of high-reliability computer systems in complex and dynamic network environments.

[0077] Please see Figure 2 It illustrates a flowchart of a highly reliable intelligent fault-tolerant computer system based on multimodal fault prediction, according to an embodiment of the present invention. Figure 2 As shown, the intelligent fault-tolerant system 20 for highly reliable computer systems based on multimodal fault prediction includes a baseline construction unit 21, a baseline update unit 22, a monitoring and early warning unit 23, and a fault-tolerant execution unit 24.

[0078] The baseline construction unit 21 is used to construct an initial mutual information baseline matrix containing multiple operating condition categories based on the operating data of the computer system during historical fault-free periods. Its specific implementation can be referred to the description in step S101 of the above embodiments, and will not be repeated here.

[0079] The baseline update unit 22 is used to dynamically determine the baseline update parameters based on statistical analysis of the data window during the real-time operation of the computer system; and to dynamically update the mutual information reference values ​​in the initial mutual information baseline matrix according to the baseline update parameters. Its specific implementation can be referred to in steps S102 and S103 of the above embodiments, and will not be repeated here.

[0080] The monitoring and early warning unit 23 is used to monitor the operating status of the computer system based on dynamically updated mutual information reference values, and generate an early warning signal when an anomaly is detected. Its specific implementation can be referred to in step S104 of the above embodiments, and will not be repeated here.

[0081] The fault-tolerant execution unit 24 is used to execute preset resource regulation operations in response to the warning signal to achieve system fault tolerance. Its specific implementation can be found in step S105 of the above embodiments, and will not be repeated here.

[0082] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0083] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A highly reliable intelligent fault-tolerant method for computer systems based on multimodal fault prediction, characterized in that, The method includes: Based on the operating data of the computer system during historical fault-free periods, an initial mutual information baseline matrix containing multiple operating condition categories is constructed. The initial mutual information baseline matrix is ​​used to store the mutual information reference values ​​of each key mode pair under different operating condition categories. During the real-time operation of the computer system, baseline update parameters are dynamically determined based on statistical analysis of the data window; The mutual information reference values ​​in the initial mutual information baseline matrix are dynamically updated according to the baseline update parameters. Based on dynamically updated mutual information reference values, the operating status of the computer system is monitored, and an early warning signal is generated when an anomaly is detected. In response to the warning signal, a preset resource regulation operation is executed to achieve system fault tolerance.

2. The intelligent fault-tolerant method for high-reliability computer systems based on multimodal fault prediction according to claim 1, characterized in that, Construct an initial mutual information baseline matrix containing multiple operating condition categories, including: The operating data is divided using a sliding window to generate a continuous data window containing all operating condition categories; The modal data within each data window is discretized. Based on the frequency of occurrence of individual modal data categories and the frequency of occurrence of combinations of modal data categories in each key modal pair, the mutual information value of each key modal pair is determined. The multiple key modal pairs include modal pairs composed of bandwidth utilization and end-to-end latency, modal pairs composed of end-to-end latency and packet loss rate, modal pairs composed of active connection count and response time, and modal pairs composed of traffic type and queue depth. Determine the local variance of the mutual information values ​​of all key mode pairs within each data window, and identify data windows with local variances lower than a preset variance threshold as candidate windows; Candidate windows are grouped into window blocks according to a preset number and time order. The global variance of each window block is determined based on the mean mutual information value of each candidate window included in each window block. The window block corresponding to the minimum global variance is determined as the initial healthy window. For each operating condition category, the average mutual information value of each key modality pair within the initial health window of the operating condition category is determined as the mutual information reference value of each key modality pair under the operating condition category. The mutual information reference values ​​of each key mode pair under all operating conditions are integrated to construct the initial mutual information baseline matrix.

3. The intelligent fault-tolerant method for highly reliable computer systems based on multimodal fault prediction according to claim 2, characterized in that, The operating data is divided using a sliding window to generate a continuous data window containing all operating condition categories, including: The running data of a first preset duration is obtained, and the running data is divided into continuous data windows using a sliding window based on a preset window length and a preset window overlap rate. Based on the load characteristics of each data window, the corresponding operating condition category for each data window is determined. The load characteristics include traffic rate characteristics, application protocol characteristics, and user concurrency characteristics. If the operating condition categories corresponding to all data windows in the continuous data window do not include all operating condition categories, the continuous data window is updated based on the operating data of the first preset duration and the operating data of the second preset duration until the operating condition categories corresponding to all data windows in the continuous data window include all operating condition categories.

4. The intelligent fault-tolerant method for high-reliability computer systems based on multimodal fault prediction according to claim 3, characterized in that, Based on the load characteristics of each data window, determine the operating condition category corresponding to each data window, including: For each data window, the load level of the data window is determined based on the traffic rate characteristics and user concurrency characteristics of the data window; Based on the application protocol characteristics of the data window, determine the application scenario of the data window; By combining the load level with the application scenario, the operating condition category corresponding to the data window is obtained.

5. The intelligent fault-tolerant method for highly reliable computer systems based on multimodal fault prediction according to claim 1, characterized in that, Based on statistical analysis of the data window, baseline update parameters are dynamically determined, including: Based on the operating condition category of each data window in the previous baseline update cycle, determine the number of times the operating condition category changed in the previous baseline update cycle; Based on the number of changes and the duration of the previous baseline update cycle, the operating condition switching frequency within the previous baseline update cycle is determined. The current baseline statistical duration is determined based on the operating condition switching frequency and the duration of the previous baseline update cycle. The larger of the current baseline statistical duration and the preset minimum period is determined as the current baseline update period. The current baseline update period is used to indicate the update of the mutual information reference value in the initial mutual information baseline matrix after the current baseline update period ends.

6. The intelligent fault-tolerant method for high-reliability computer systems based on multimodal fault prediction according to claim 5, characterized in that, Based on statistical analysis of the data window, the baseline update parameters are dynamically determined, which also includes: Determine the operating condition category of the health data window within the previous baseline update cycle, wherein the health data window is a data window without any abnormalities; For each operating condition category, a smoothing factor corresponding to the operating condition category is determined based on the cumulative number of health data windows corresponding to that operating condition category.

7. The intelligent fault-tolerant method for high-reliability computer systems based on multimodal fault prediction according to claim 6, characterized in that, Based on the baseline update parameters, the mutual information reference values ​​in the initial mutual information baseline matrix are dynamically updated, including: After the previous baseline update cycle ends, for each operating condition category, the mutual information reference values ​​of each key mode pair under the operating condition category in the initial mutual information baseline matrix are updated and replaced according to the smoothing factor corresponding to the operating condition category.

8. The intelligent fault-tolerant method for high-reliability computer systems based on multimodal fault prediction according to claim 1, characterized in that, Based on dynamically updated mutual information reference values, the operating status of the computer system is monitored, and an early warning signal is generated when an anomaly is detected, including: Determine the real-time mutual information values ​​of each key modality pair in the current data window; For each key mode pair, determine the relative deviation between the real-time mutual information value of the key mode pair and the corresponding updated mutual information reference value; The mean of the relative deviations corresponding to at least one data window containing the current data window is determined as the correlation decay factor corresponding to the current data window. The data windows in the at least one data window are continuous and the operating condition categories are all the current operating condition category. The current operating condition category is the operating condition category corresponding to the current data window. If the correlation decay factor is greater than the adaptive threshold corresponding to the current operating condition category, it is determined that the current data window is abnormal. If multiple consecutive data windows show abnormalities, the computer system is determined to be malfunctioning, and an early warning signal is generated.

9. The intelligent fault-tolerant method for highly reliable computer systems based on multimodal fault prediction according to claim 8, characterized in that, Determining the adaptive threshold includes: For each key mode pair under each working condition category, if the number of determined correlation decay factors is less than the preset number, the preset fixed value is determined as the adaptive threshold. If the number of known correlation decay factors is greater than the preset number, the adaptive threshold is determined based on the preset percentile value of the numerical distribution of correlation decay factors that are less than the preset fixed value, thereby obtaining the adaptive threshold corresponding to each key mode pair under all working condition categories.

10. The intelligent fault-tolerant method for high-reliability computer systems based on multimodal fault prediction according to claim 8, characterized in that, Generate early warning signals, including: Identify target key mode pairs, which are key mode pairs whose correlation decay factor is greater than the corresponding adaptive threshold in the consecutive multiple data windows; When the target critical mode pair is a mode pair consisting of bandwidth utilization and end-to-end delay, the warning signal is used to indicate abnormal bandwidth allocation; When the target critical mode pair is a mode pair consisting of end-to-end delay and packet loss rate, the warning signal is used to indicate network congestion; When the target key modal pair is a modal pair consisting of the number of active connections and the response time, the warning signal is used to indicate server overload; When the target critical modal pair is a modal pair consisting of traffic type and queue depth, the early warning signal is used to indicate traffic resource preemption.

Images