Operation and maintenance management method and system based on multi-algorithm fusion

Abstract

The operation and maintenance management method and system based on multi-algorithm fusion belong to the technical field of computers and comprise the following steps: S1, receiving operation and maintenance data and algorithm conclusions generated by multiple concurrent algorithms, and identifying a current event context in the operation and maintenance data; S2, for the multiple concurrent algorithms, calculating a context confidence based on the current event context and a historical cumulative performance score of the algorithm; S3, based on all algorithm conclusions, using a preset consensus measurement function to evaluate the consensus divergence of the algorithm group; S4, combining the context confidence and the consensus divergence and taking into account a preset minority identification mechanism, calculating the dynamic weight of each algorithm; S5, according to the dynamic weight, performing weighted fusion on all algorithm conclusions to generate a final decision; S6, in response to the execution result of the final decision, feeding back the historical cumulative performance score of each algorithm for updating, and the present application significantly improves the accuracy and scene adaptability of operation and maintenance decisions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to an operation and maintenance management method and system based on multi-algorithm fusion. BACKGROUND

[0002] In modern automated operation and maintenance systems, multiple analysis algorithms are usually run in parallel to deal with complex operation and maintenance events. The existing technology mostly adopts a static weight fusion strategy, which cannot adapt to different operation and maintenance contexts, such as planned changes and sudden failures, and the dynamic changes of the performance of each algorithm, often leading to delayed decision-making or submerging of key minority alarm information, lacking quantitative management and optimization capabilities for consensus processes. Therefore, there is a need for an advanced technical solution that can dynamically adjust the weight of algorithms according to the event context and balance the consensus efficiency and abnormal sensitivity.

[0003] The above information disclosed in the above BACKGROUND section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0004] The purpose of the present application is to provide an operation and maintenance management method and system based on multi-algorithm fusion to solve the problems raised in the above BACKGROUND.

[0005] The technical solution of the present application is as follows:

[0006] S1, receiving operation and maintenance data and algorithm conclusions generated by multiple concurrent algorithms, and identifying the current event context in the operation and maintenance data, the multiple concurrent algorithms being root cause localization algorithms, anomaly detection algorithms and security scanning algorithms, respectively;

[0007] S2, for the multiple concurrent algorithms, calculating a context confidence based on the current event context and the historical cumulative performance scores of the algorithms;

[0008] S3, based on all algorithm conclusions, using a preset consensus measurement function to evaluate the consensus divergence of the algorithm group;

[0009] S4, combining the context confidence and the consensus divergence, and taking into account a preset minority identification mechanism, calculating the dynamic weight of each algorithm;

[0010] S5, according to the dynamic weight, weighting and fusing all algorithm conclusions to generate a final decision;

[0011] S6, in response to the execution result of the final decision, feeding back to update the historical cumulative performance scores of each algorithm.

[0012] Preferably, S2 specifically includes:

[0013] The confidence decay model is combined with the historical cumulative performance and the time interval since the last conclusion of the algorithm was successfully verified to calculate the context confidence;

[0014] The confidence decay model includes a confidence decay factor associated with the current event context, which is used to adjust the degree of influence of the time interval on the confidence.

[0015] Preferably, S3 specifically includes:

[0016] When the algorithm conclusion is a discrete category label, the Gini impurity is calculated to quantify the inconsistency of the category distribution of all algorithm conclusions as the consensus divergence;

[0017] When the algorithm conclusion is a continuous numerical value, the coefficient of variation is calculated to quantify the numerical dispersion of all algorithm conclusions as the consensus divergence.

[0018] Preferably, S4 specifically includes:

[0019] S41, calculate the basic activation value, which is generated by the context confidence after being punished by the consensus divergence;

[0020] S42, for the algorithm that hits the key minority mode, the basic activation value is gain adjusted to generate the final activation value;

[0021] S43, process the final activation value of all algorithms through the Softmax function to generate dynamic weights.

[0022] Preferably, the hit determination of the key minority mode specifically includes:

[0023] Combine the conclusions of the algorithm with the characteristics of the current event to form a to-be-detected pattern;

[0024] Match the to-be-detected pattern with the preset key minority pattern library;

[0025] The key minority pattern library is used to store high-risk patterns;

[0026] When the to-be-detected pattern hits the high-risk pattern, the gain adjustment of the basic activation value is triggered.

[0027] Preferably, S5 specifically includes:

[0028] When the algorithm conclusion is a category label, the weighted voting method is used to select the category with the highest cumulative weight as the final decision;

[0029] When the algorithm conclusion is a continuous numerical value, the weighted average method is used to calculate the weighted average value of the algorithm conclusion as the final decision.

[0030] Preferably, S6 specifically includes:

[0031] Set a reward value based on the success or failure of the final decision execution result;

[0032] An incremental learning formula is used to update the historical cumulative performance score in combination with the preset learning rate and the reward value.

[0033] A multi-algorithm fusion-based operation and maintenance management system comprises:

[0034] A data and conclusion acquisition module is configured to receive operation and maintenance data and algorithm conclusions of concurrent algorithms and identify event context;

[0035] A dynamic evaluation module is configured to calculate the context confidence of each algorithm and the consensus divergence of the algorithm group.

[0036] A weight generation module is configured to generate dynamic weights in combination with the context confidence, the consensus divergence and the minority identification mechanism.

[0037] A decision fusion module is configured to perform weighted fusion on the algorithm conclusions according to the dynamic weights to generate a final decision.

[0038] A feedback optimization module is configured to update the historical cumulative performance score of each algorithm in response to the execution result of the final decision.

[0039] The present application provides a multi-algorithm fusion-based operation and maintenance management method and system, which has the following improvements and advantages compared with the prior art:

[0040] 1. By calculating the context confidence closely related to the current event context, the present application ensures that in specific scenarios such as database failure, the weight of an algorithm with a long history but poor recent performance will naturally decrease, while the weight of an algorithm that has successfully predicted such failures in recent times will increase; this makes the decision always tend to adopt the most appropriate and most reliable algorithm conclusion in the current scenario, and the decision accuracy is much higher than that of the traditional method using static weights, significantly improving the accuracy and scenario adaptability of operation and maintenance decisions.

[0041] 2. By introducing consensus divergence as a global adjustment factor, when the algorithm group opinions are seriously divergent, the system will automatically enter a cautious mode, and the basic weight of all algorithms will be inhibited; this mechanism effectively prevents the system from making wrong decisions by blindly accepting any conclusion when facing new, unknown or data ambiguous complex events, thereby improving the stability and robustness of the entire operation and maintenance system and greatly enhancing the decision robustness and risk avoidance ability of the system.

[0042] 3. The sensitivity to high-risk is improved; relying on the minority identification mechanism, the scheme can accurately salvage a few minority opinions from a large amount of redundant information, which indicate a major risk; for example, at the initial stage of a cloud platform being attacked by an advanced persistent threat, most performance monitoring algorithms may show normal, and only one traffic behavior analysis algorithm detects a weak abnormal signal; the traditional method ignores this signal, but the scheme can identify it as a high-risk pattern by matching the key minority pattern library, and instantly increase the dynamic weight of the algorithm, thereby triggering an alarm and gaining a valuable time window to deal with a major security event;

[0043] 4. The historical cumulative performance of each algorithm is dynamically adjusted through the feedback update mechanism based on the success or failure of the final decision, and each operation and maintenance event is converted into a learning opportunity; successful experience is solidified, and the lessons of failure are also remembered, and these knowledge is automatically deposited in the quantitative score of each algorithm; this enables the system to realize true self-adaptation and self-optimization, and continuously improve the overall decision-making efficiency without a large amount of manual intervention. BRIEF DESCRIPTION OF DRAWINGS

[0044] The application will be further explained below in conjunction with the accompanying drawings and embodiments:

[0045] Figure 1 is a flow chart of the system of the application. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in conjunction with specific embodiments.

[0047] Embodiment 1:

[0048] Please refer to Figure 1 The application provides a multi-algorithm fusion-based operation and maintenance management method and system, including the following steps: S1, receiving operation and maintenance data and algorithm conclusions generated by multiple concurrent algorithms, and identifying the current event context in the operation and maintenance data, the multiple concurrent algorithms being respectively a root cause positioning algorithm, an anomaly detection algorithm and a security scanning algorithm;

[0049] S2, for the multiple concurrent algorithms, calculating a context confidence based on the current event context and the historical cumulative performance of the algorithm;

[0050] S3, based on all algorithm conclusions, using a preset consensus measurement function to evaluate the consensus divergence of the algorithm group;

[0051] S4, combining the context confidence and the consensus divergence, and taking into account the preset minority identification mechanism, calculating the dynamic weight of each algorithm;

[0052] S5, according to the dynamic weight, all algorithm conclusions are weighted and fused to generate a final decision;

[0053] S6, in response to the execution result of the final decision, feedback is given to update the historical cumulative performance score of each algorithm.

[0054] Here, the operation and maintenance data is non-limiting, and can include but is not limited to: server CPU / memory usage and other indicator data, application output text logs, and request call chain data in a distributed system; and the step of identifying the current event context in the operation and maintenance data can be specifically implemented by preprocessing the original operation and maintenance data; for example, specific error codes in the log are matched through a regular expression, or a pre-trained classification model is used to classify the time series pattern of the indicator data, so that the massive original data is mapped to a predefined and standardized context label set, such as: database slow query, network jitter, middleware exception, security attack;

[0055] The embodiment of the application discloses an operation and maintenance management method based on multi-algorithm fusion; the method aims to solve the technical problems in the prior art that the multi-algorithm fusion strategy usually adopts static weight, and thus cannot adapt to the dynamically changing operation and maintenance context, and it is difficult to effectively balance the group consensus and the key minority opinion; through a closed-loop process composed of six interrelated steps, comprehensive management is realized from data reception, dynamic evaluation, weight generation, decision fusion to feedback optimization; the core of the technical scheme is to abandon fixed weight allocation, and instead calculate the context confidence of each algorithm in a specific event context in real time, and quantify the consensus divergence degree among the algorithm groups, and then combine the minority identification mechanism to generate dynamic weight; through subsequent weighted fusion and feedback update, the method can dynamically identify the most credible root cause localization algorithm in the current fault mode in a complex IT operation and maintenance scene, such as a large-scale cloud service failure, while giving sufficient weight promotion to the abnormal detection algorithm that issues rare but key alarms; in this way, a decision-making system that can quickly respond to mainstream problems and is highly sensitive to potential high-risk risks is formed, which is continuously self-optimized, thereby significantly improving the accuracy, timeliness and security of operation and maintenance decisions.

[0056] Embodiment 2

[0057] According to the foregoing method, S2 specifically includes:

[0058] A confidence decay model is used to calculate the context confidence in combination with the historical cumulative performance score and the time interval since the last successful verification of the algorithm conclusion;

[0059] The confidence decay model includes a confidence decay factor associated with the current event context, which is used to adjust the influence degree of the time interval on the confidence;

[0060] S3 specifically comprises:

[0061] When the algorithm conclusion is a discrete category label, the Gini impurity is calculated to quantify the inconsistency of the category distribution of all algorithm conclusions as the consensus divergence;

[0062] When the algorithm conclusion is a continuous numerical value, the coefficient of variation is calculated to quantify the numerical dispersion degree of all algorithm conclusions as the consensus divergence.

[0063] When the algorithm conclusion is a continuous numerical value, the consensus divergence can be calculated by the coefficient of variation , which is defined as the ratio of the standard deviation of all algorithm conclusion numerical values to the average value , that is ; this value quantifies the relative dispersion degree of the conclusion;

[0064] The embodiment provides a context confidence calculation method and a consensus divergence evaluation method; the design aims to quantify the real-time credibility of a single algorithm and the opinion consistency of an algorithm group, and provides key and mutually independent quantitative inputs for subsequent dynamic weight calculation;

[0065] To calculate the context confidence, the confidence decay model constructed in the embodiment is inspired by the Ebbinghaus forgetting curve, and it is believed that the current credibility of an algorithm depends not only on its long-term historical performance, but also on its recent performance. The mathematical expression is:

[0066]

[0067] Among them, represents the context confidence of the algorithm at time when processing the current event, which is a dimensionless score; is the unique index of the algorithm; is the current processing time; is the context type of the current event, such as database slow query or network jitter; represents the historical cumulative performance score of the algorithm under a specific context , which is continuously updated by the success or failure of historical decisions, and its value range is ; refers to the scalar value of the time elapsed or the event count since the last conclusion of the algorithm under the context was successfully verified; is the context a confidence decay factor associated with the context S, being a positive real number, dimensioned as the inverse of time or the inverse of the number of events, which can be set a priori based on the dynamic characteristics of different contexts, for example, set to a larger value in a rapidly changing network jitter scenario to strengthen the sensitivity to recent performance; : the base of natural logarithm, being a mathematical constant; : the unique identifier of the i-th algorithm;

[0068] To make the setting more operational, the value of may be associated with the average occurrence period of historical events under the context S; for example, an empirical formula can be set as: , where refers to the average time interval or the average event count interval of the occurrence of key events in history under the context S, and is a global adjustment constant; for example ; in this way, for a context with frequent event occurrence ( small), its decay factor will be naturally larger, and vice versa, thereby achieving adaptive setting of the parameter, wherein : the confidence decay factor associated with the context S; : a global adjustment constant;

[0069] The application of this formula enables the system to accurately evaluate the instantaneous value of each algorithm under a specific situation;

[0070] Based on the aforementioned calculation results, to evaluate the consensus consistency of the algorithm group, this embodiment introduces the calculation of consensus divergence; the purpose is to quantify the conflict degree of group opinion, so that the system can adopt a more prudent decision-making posture when the conflict is severe; for discrete category label conclusions, the Gini impurity is used for quantification:

[0071]

[0072] , where represents the consensus divergence of all algorithm conclusions at time , with a value range of , the larger the value, the more serious the divergence; is the set of unique category labels output by all algorithms at time ; is the total number of algorithms participating in the analysis; is the number of algorithms whose conclusions are category label ; the application of this formula provides a global uncertainty adjustment factor for the system; when the divergence ​​The higher, the more significant uncertainty of the current event, and the value will act as a global penalty factor in subsequent calculations, overall suppressing the weight basis of all algorithms, preventing the system from making hasty decisions in the absence of consensus; is a discrete category label of the algorithm conclusion.

[0073] Embodiment 3

[0074] According to the foregoing method, S4 specifically comprises:

[0075] S41, calculating a basic activation value, which is generated by the context confidence being punished by the consensus divergence;

[0076] S42, gain-adjusting the basic activation value for the algorithm that hits the key minority mode, to generate a final activation value;

[0077] S43, processing the final activation values of all algorithms by a Softmax function to generate dynamic weights;

[0078] And the hit determination of the key minority mode specifically comprises:

[0079] Combining the conclusions of the algorithms with the features of the current event to form a to-be-inspected mode;

[0080] Matching the to-be-inspected mode with a preset key minority mode library;

[0081] The key minority mode library is used to store high-risk modes;

[0082] The construction of the library can follow the following semi-automatic process:

[0083] Candidate mode mining: for the historical major failure operation and maintenance data set, adopt association rule mining such as Apriori algorithm or unsupervised clustering algorithm to automatically find the rare coexistence mode of event features and specific algorithm conclusions as candidates; for example, it is found that in the time window when the software version is just released, the specific security scanning algorithm conclusion is high-risk SQL injection risk, this low-frequency but key associated mode;

[0084] Pattern structured definition: store the mined candidate patterns and high-risk patterns directly defined by domain experts according to experience in a structured pattern library. The library can be a key-value database, where the key is composed of the event context S and the event core features such as source IP and error code, and the value is the algorithm conclusion that should be considered as a key minority and matched therewith;

[0085] Expert review and annotation: organize operation and maintenance experts and security experts to review and confirm all patterns in the library, eliminate pseudo-patterns, and set risk levels for the confirmed patterns. This process ensures the accuracy and authority of the pattern library;

[0086] Continuous maintenance and update: during the operation of the system, for each significant event confirmed by manual intervention and not hit the current mode library, a review process should be triggered to refine and supplement new high-risk modes into the library, realizing the continuous iteration of the mode library;

[0087] Through the above process, the construction and maintenance of the key minority mode library can be ensured to be specific, implementable and continuously optimized;

[0088] When the to-be-inspected mode hits the high-risk mode, the gain adjustment of the basic activation value is triggered.

[0089] The embodiment provides a dynamic weight generation method, and the internal logic is that a calculation process in which confidence of a single algorithm, consensus of an algorithm group and sensitivity to a special high-risk scene are fused in a unified mathematical framework is constructed; the process is completed by constructing an intermediate activation value and performing normalization processing;

[0090] An integrated intermediate activation value is constructed for measuring the final influence of the opinions of algorithms:

[0091]

[0092] Wherein, is the final activation value of the algorithm at the moment , which is a dimensionless value; is the context confidence calculated in the foregoing step; is the consensus divergence calculated in the foregoing step; is a dimensionless hyperparameter, which serves as a difference penalty coefficient of cross-algorithm output, and the value thereof can be set according to the risk level of the operation and maintenance scene; for a high-risk scene, a higher value should be set to strengthen the punishment of divergence; is a dimensionless gain coefficient, which serves as a key minority promotion factor; is a minority identification flag, which is a binary value ; The determination of is based on a strict matching process: the system combines the current conclusion of the algorithm with the feature combination of the current event to form a to-be-inspected mode, and matches the to-be-inspected mode with a preset key minority mode library; the library is constructed by domain expert knowledge and historical major fault review data, and stores high-risk modes such as {event feature combination, algorithm conclusion}; if the to-be-inspected mode hits an entry in the library, is set to 1, otherwise, is set to 0; the hyperparameters and are determined according to the following formula: The optimal value can be determined by cross-validation on the historical event dataset; in particular, one can first set a reasonable initial range, e.g., a higher risk scenario corresponds to a higher range, and then find the combination that maximizes the objective function within the range by grid search or other methods, as its optimal value. This way, both expert experience and data-driven optimization of parameters are utilized; : algorithm index and current time, respectively;

[0093] where the system performance should be defined by a clear, quantifiable objective function; for example, in the root cause localization scenario, the objective function can be defined as the combination of the weighted accuracy of all decisions on the validation set and the average decision time , such as maximizing the objective function , where is a weight factor used to balance the accuracy and timeliness; find the combination that maximizes the objective function within the range by grid search or other methods, as its optimal value; : objective function for optimizing hyperparameters; : weighted accuracy of all decisions on the validation set; : weight factor used to balance the accuracy and timeliness; : average decision time;

[0094] This calculation of activation value reflects the balance of multiple factors; the first term reflects the adjustment of basic confidence by group consensus, when the difference is large, the basic activation value of all algorithms is suppressed; the second term is a gain adjustment mechanism aimed at amplifying the influence of key minority opinions; for example, in a software release process, most monitoring algorithms show normal, but a security scanning algorithm detects a behavior that meets the key minority mode library entry of remote code execution exploiting specific API of the new version, at this time, even if the algorithm is not high, its will be set to 1, through the gain adjustment of factor, its final activation value will be significantly improved, ensuring that the key alarm will not be submerged;

[0095] Based on the above calculation results, to generate normalized dynamic weights, the Softmax function is used to process the final activation value of all algorithms:

[0096]

[0097] where,​​​​ is the final dynamic weight of algorithm at time t, is a dimensionless probability value, and the sum of all weights is 1. is the final activation value calculated in the previous step. is the final dynamic weight of algorithm at time t, is a dimensionless probability value, and the sum of all weights is 1. is the final activation value calculated in the previous step. is the total number of algorithms. is the index for traversing all algorithms; this step converts independent activation values into a unified probability distribution, completing the logical chain from multi-dimensional factor evaluation to final weighting. : the base of natural logarithm, which is a mathematical constant.

[0098] Embodiment 4

[0099] S5 specifically includes:

[0100] When the algorithm conclusion is a category label, a weighted voting method is used to select the category with the highest cumulative weight as the final decision.

[0101] When the algorithm conclusion is a continuous numerical value, a weighted average method is used to calculate the weighted average value of the algorithm conclusion as the final decision.

[0102] This embodiment provides a decision fusion method, which aims to effectively apply the dynamic weights calculated in the previous steps to each algorithm conclusion, so as to generate a unique and optimal final decision according to the data type of the conclusion.

[0103] If the algorithm conclusion is a discrete category label, such as fault level, a weighted voting method is used:

[0104]

[0105] wherein, represents the final decision category at time t; represents the final decision category at time t; represents the index of the algorithm whose conclusion is a category ; is the dynamic weight of each algorithm; wherein, represents the final decision category at time t; is the category conclusion output by the algorithm ; represents the category ; : the dynamic weight of each algorithm; : the discrete category label of the algorithm conclusion; If the algorithm conclusion is a continuous numerical value, such as predicted resource usage, a weighted average method is used:

[0106] ​​​

[0107]

[0108] wherein, represent the final decision value at time ; is the dynamic weight of algorithm ; is the continuous numerical conclusion output by algorithm ; is the total number of algorithms; : algorithm index;

[0109] The decision fusion step converts the complex weight calculation result into specific and executable operation and maintenance instructions; for example, if the root cause positioning algorithm A has a conclusion of database lock waiting with a weight , algorithm B has a conclusion of network delay with a weight , through weighted voting, the final decision will clearly point to database lock waiting, thereby guiding the operation and maintenance personnel to immediately concentrate resources to process the problem, improving the efficiency and accuracy of fault processing.

[0110] Embodiment 5

[0111] S6 specifically includes:

[0112] Based on the success or failure result of the execution of the final decision, a reward value is set;

[0113] An incremental learning formula is used to update the historical cumulative performance score of each algorithm in combination with the preset learning rate and reward value.

[0114] The embodiment provides a feedback optimization method, aiming to establish a closed-loop learning mechanism, so that the entire operation and maintenance management system can learn from its decision experience and realize continuous self-optimization.

[0115] The embodiment uses an incremental learning formula to update the historical cumulative performance score of each algorithm in a specific context:

[0116]

[0117] wherein, is the updated historical cumulative performance score of algorithm in context ; is the performance score before updating; is the preset learning rate, which is a hyperparameter between to control the amplitude of this learning;

[0118] The selection of a value such as 0.01 to 0.1 can ensure the historical cumulative performance score The updating process is more stable but slower in convergence; while a larger value such as 0.1 to 0.5 can accelerate learning but may cause dramatic fluctuations in performance scores; in practice, may be set as a fixed value or as a hyperparameter to be tuned together with the aforementioned , through cross-validation;

[0119] is the time point According to the reward value set by the success or failure of the execution of the final decision , for example, when the decision is verified as correct , or incorrect ; : algorithm index, event context, and current time, respectively;

[0120] And for any algorithm When it is first run in the context , the initial value of its historical cumulative performance score should be set; to avoid initial bias, it is recommended to use a neutral initial value, i.e. ; this value represents the system's neutral attitude towards the algorithm's performance in this scenario without any historical basis; with the continuous occurrence of subsequent decisions and feedback, this value will quickly converge to its actual performance level according to the formula;

[0121] This feedback optimization mechanism constitutes the core feedback path of the system's adaptive and long-term performance optimization; after a successful fault handling, the system confirms that the final decision is correct, then ; at this time, the historical cumulative performance score of the algorithm that contributed to this correct decision will increase; after long-term operation, the algorithm that consistently makes correct judgments will have a steadily increasing value in the corresponding scenario, which directly improves its probability of obtaining higher context confidence in future similar events; this mechanism ensures that the system can dynamically give more influence to algorithms that perform better, thus making the long-term performance of the entire multi-algorithm fusion system tend to be optimal.

[0122] Embodiment 6

[0123] A multi-algorithm fusion-based operation and maintenance management system based on any of the foregoing multi-algorithm fusion-based operation and maintenance management methods, comprising: a data and conclusion acquisition module for receiving operation and maintenance data and algorithm conclusions of concurrent algorithms and identifying event context;

[0124] a dynamic evaluation module configured to calculate the contextual confidence of each algorithm and the consensus divergence of the algorithm group;

[0125] a weight generation module configured to generate dynamic weights by combining the contextual confidence, the consensus divergence, and a minority identification mechanism;

[0126] a decision fusion module configured to perform weighted fusion on algorithm conclusions according to the dynamic weights to generate a final decision;

[0127] a feedback optimization module configured to update the historical cumulative performance scores of each algorithm in response to the execution result of the final decision.

[0128] The embodiment provides an operation and maintenance management system based on multi-algorithm fusion, which is a system implementation of the foregoing method and ensures the implementability of the technical solution. The system includes five core functional modules that work together to realize a dynamic consensus optimization decision-making process: a data and conclusion acquisition module, which is an input interface of the system and is responsible for capturing operation and maintenance data, distributing the operation and maintenance data to concurrent algorithms, and identifying event context; a dynamic evaluation module, which is responsible for executing a confidence decay model and consensus divergence calculation and quantitatively evaluating algorithm individual performance and group consistency; a weight generation module, which is a core calculation unit of the system and is responsible for executing activation value calculation and Softmax normalization containing a minority promotion mechanism to generate dynamic weights; a decision fusion module, which is responsible for applying a weighted fusion method to apply the weights to algorithm conclusions to form a unified decision; and a feedback optimization module, which constitutes a learning loop of the system and updates the historical performance scores of each algorithm according to the actual execution effect of the decision. The five modules are functionally coupled and together constitute an intelligent operation and maintenance management system that can autonomously learn, dynamically adapt, and efficiently make decisions.

[0129] It should be noted that the foregoing embodiments are merely used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been particularly described with reference to the preferred embodiments, it should be understood that those of ordinary skill in the art can make modifications or equivalent replacements to the technical solutions of the present application without departing from the spirit and scope of the present application, and all these modifications and equivalent replacements should be encompassed in the scope of the claims of the present application.

Claims

1. An operation and maintenance management method based on multi-algorithm fusion, characterized in that, The method comprises the following steps: S1, receiving operation and maintenance data and algorithm conclusions generated by multiple concurrent algorithms, and identifying a current event context in the operation and maintenance data, the multiple concurrent algorithms being respectively a root cause positioning algorithm, an anomaly detection algorithm and a security scanning algorithm; S2, for the multiple concurrent algorithms, calculating a context confidence based on the current event context and a historical cumulative performance score of the algorithm; S3, based on all algorithm conclusions, using a preset consensus measurement function to evaluate a consensus divergence of the algorithm group; S4, combining the context confidence and the consensus divergence and taking into account a preset minority identification mechanism, calculating a dynamic weight of each algorithm; S5, according to the dynamic weight, performing weighted fusion on all algorithm conclusions to generate a final decision; S6, in response to an execution result of the final decision, feeding back updated historical cumulative performance scores of the algorithms; S3 specifically comprises: When the algorithm conclusion is a discrete category label, Gini impurity is used to quantify the category distribution inconsistency of all algorithm conclusions as the consensus divergence; When the algorithm conclusion is a continuous numerical value, a coefficient of variation is used to quantify the numerical dispersion degree of all algorithm conclusions as the consensus divergence; S4 specifically comprises: S41, calculating a basic activation value, which is generated by the context confidence after being punished by the consensus divergence; S42, for the algorithm that hits the key minority mode, gain-adjusting the basic activation value to generate a final activation value; S43, processing the final activation values of all algorithms through a Softmax function to generate the dynamic weight. 2.The operation and maintenance management method based on multi-algorithm fusion of claim 1, characterized in that, S2 specifically comprises: A confidence decay model is used to calculate the context confidence in combination with the historical cumulative performance score and a time interval since the last conclusion of the algorithm was successfully verified; The confidence decay model contains a confidence decay factor associated with the current event context, which is used to adjust the influence degree of the time interval on the confidence. 3.The operation and maintenance management method based on multi-algorithm fusion of claim 1, characterized in that, The hit determination of the key minority mode specifically comprises: Combining the conclusion of the algorithm and the characteristics of the current event to form a to-be-detected mode; Matching the to-be-detected mode with a preset key minority mode library; The key minority mode library is used to store high-risk modes; When the to-be-detected mode hits a high-risk mode, gain-adjusting the basic activation value is triggered. 4.The operation and maintenance management method based on multi-algorithm fusion of claim 1, characterized in that, S5 specifically comprises: When the algorithm conclusion is a category label, a weighted voting method is used to select the category with the highest cumulative weight as the final decision; When the algorithm conclusion is a continuous numerical value, a weighted average method is used to calculate a weighted average value of the algorithm conclusion as the final decision.

5. The operation and maintenance management method based on multi-algorithm fusion according to claim 1, characterized in that, S6 specifically comprises: Based on the success or failure result of the execution of the final decision, setting a reward value; Using an incremental learning formula in combination with a preset learning rate and the reward value to update the historical cumulative performance score.

6. An operation and maintenance management system based on multi-algorithm fusion, based on the operation and maintenance management method based on multi-algorithm fusion of any one of claims 1 to 5, characterized in that, It comprises: A data and conclusion acquisition module for receiving operation and maintenance data and algorithm conclusions of concurrent algorithms and identifying an event context; A dynamic evaluation module for calculating a context confidence of each algorithm and a consensus divergence of the algorithm group; A weight generation module for generating a dynamic weight in combination with the context confidence, the consensus divergence and a minority identification mechanism; A decision fusion module for performing weighted fusion on algorithm conclusions according to the dynamic weight to generate a final decision; a feedback optimization module for updating the historical cumulative performance scores of the algorithms in response to the execution results of the final decisions.

Images