An intelligent operation and maintenance and fault disposal method based on an AI large model

Abstract

The application discloses an intelligent operation and maintenance and fault disposal method based on an AI large model, and relates to the field of artificial intelligence and information technology.The intelligent operation and maintenance method comprises the following steps: multi-source heterogeneous data fusion, cross-level fault analysis, disposal strategy generation and verification, and resource optimization scheduling.The application has the advantages that multi-source heterogeneous data of equipment layers, software layers and business layers are synchronously collected through a perception layer, and a time-space aligned feature matrix is constructed, so that the limitation of traditional operation and maintenance which only relies on single-level data is solved.Combining a multi-modal AI large model of an analysis layer, time dependence and hierarchical correlation of data can be simultaneously mined, fault root cause probability distribution and influence levels can be accurately output, misjudgment or missed judgment caused by single-level data deviation is avoided, and the accuracy of complex system fault positioning is significantly improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and information technology, specifically to an intelligent operation and maintenance and fault handling method based on an AI large model. Background Technology

[0002] With the deepening of digital transformation, the information systems of enterprises and institutions are exhibiting complex characteristics of multi-source heterogeneity and cross-level connections. They cover the collaborative operation of the equipment layer (such as servers and communication base stations), software layer (such as databases and control programs), and business layer (such as e-commerce platforms and industrial production lines). When dealing with such complex systems, traditional intelligent operation and maintenance methods face problems such as low efficiency of multi-source data fusion, reliance on human experience for fault root cause location, and lack of dynamic optimization of handling strategies and resource scheduling, making it difficult to meet the operation and maintenance requirements of high reliability and low cost.

[0003] In fault handling scenarios, traditional methods often suffer from insufficient cross-level impact analysis capabilities, leading to misjudgment or omission of root causes. At the same time, the verification of handling effects and strategy iteration rely on manual summarization, which cannot achieve model self-optimization, resulting in operation and maintenance response delays, resource waste, and business interruption risks. To address this, we propose an intelligent operation and maintenance and fault handling method based on an AI large model. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent operation and maintenance and fault handling method based on an AI large model.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent operation and maintenance method based on an AI large model, comprising the following steps:

[0006] S1. Multi-source heterogeneous data fusion: Collect multi-source heterogeneous data from the device layer, software layer and business layer through the perception layer, and construct a spatiotemporally aligned feature matrix.

[0007] S2. Cross-level fault analysis: The feature matrix is ​​processed by the multimodal AI big model of the analysis layer to output the probability distribution of fault root causes and the level of influence, and the fault judgment threshold is updated in real time through a dynamic confidence mechanism.

[0008] S3. Generation and verification of handling strategies: Based on the probability distribution of the root causes of the fault, a handling strategy is generated. After execution, the system status data is collected by the perception layer to verify the handling effect and the feedback is sent to the analysis layer to drive the model self-optimization.

[0009] S4. Resource optimization scheduling: Based on fault level and cost constraints, select suitable operation and maintenance resources through a multi-objective optimization function.

[0010] As a further aspect of the present invention: the multi-source heterogeneous data in step S1 includes:

[0011] Device layer data: Temperature, power consumption, and IOPS (Input / Output Operations) collected by device sensors;

[0012] Software layer data: error codes, call chains, and memory usage in software logs;

[0013] Business layer data: QPS (queries per second), response latency, and user conversion rate collected by the business monitoring system;

[0014] The spatiotemporally aligned feature matrix is ​​constructed using the following formula:

[0015] ;

[0016] In the formula, This is a vector concatenation operation. Raw data, The mean of the data. Standard deviation This is a tensor splicing operation. Encode the timestamp. Encode the device location.

[0017] As a further aspect of the present invention: In the hybrid architecture, the output vector of the Transformer and the node embedding vector of the GNN are fused through a concatenation layer and then input into the fully connected layer. In step S2, the multimodal AI large model adopts a hybrid architecture of Transformer + Graph Neural Network (GNN). The Transformer is used to capture the long-term dependency features of time series data, and the GNN is used to model the topological relationship between equipment, software, and business. The probability distribution of the root cause of the fault is the probability value of each fault cause in the equipment layer, software layer, and business layer. The influence level is at least one of the equipment layer, software layer, and business layer.

[0018] As a further aspect of the present invention: the dynamic confidence mechanism in step S2 updates the fault determination threshold in real time using the following formula:

[0019] ;

[0020] In the formula, for The dynamic threshold at any given time. for Threshold of time, Forgetting factor (value range 0.7 ≤ ≤0.9 (determined by convergence experiments based on historical fault data). Number of recent samples (range 50 ≤ ≤200), For the first The actual fault error of a sample This represents the model prediction error.

[0021] As a further aspect of the present invention: the cross-level fault analysis in step S2 further includes calculating the cross-level impact index of equipment-software-business, as shown in the following formula:

[0022] ;

[0023] In the formula, Hierarchical weights (device layer) Values ​​range from 0.4 to 0.6, software layer. Values ​​range from 0.2 to 0.4, business layer. (Values ​​range from 0.1 to 0.3) hierarchical The index offset (i.e., the difference between the current value and the baseline value). hierarchical The maximum allowable offset of the indicator is determined by comparing the values ​​at each level. The value determines the level of the primary cause of the failure.

[0024] As a further aspect of the present invention: the multi-objective optimization function in step S4 is:

[0025] ;

[0026] In the formula, Weighting of fault recovery time (value range 0.5≤ ≤0.8), Resource cost weights (value range 0.2 ≤ ≤0.5), For fault recovery time, To reduce operational resource costs, the adapted operational resources include edge node resources and central node resources in the cloud-edge collaborative architecture. This is achieved by comparing the costs of different resources. and Select the resource that satisfies the SLA and has the smallest objective function value.

[0027] This invention also provides a fault handling method for intelligent operation and maintenance based on an AI large model, the fault handling method comprising the following steps:

[0028] T1. Receive the probability distribution of root causes of failures and the impact level output by the intelligent operation and maintenance method;

[0029] T2. Use the knowledge graph to generate a sequence of actions to handle the situation;

[0030] T3. After executing the sequence of actions, the system status data (including service latency, equipment indicators, and software logs) is collected through the perception layer to evaluate the effectiveness of the actions;

[0031] T4. If the effect is not met (i.e. the Service Level Agreement (SLA) requirements are not met), the handling result will be fed back to the analysis layer of the intelligent operation and maintenance method, the confidence of the corresponding fault-action in the knowledge graph will be updated, and the handling strategy will be regenerated.

[0032] As a further aspect of the present invention: the mapping rules between the knowledge graph storage fault types and the handling actions include, but are not limited to, disk I / O overload → migration to high-performance disk, cache miss → enabling memory caching, and the handling action sequence optimizes the action order through reinforcement learning, giving priority to actions with low cost and quick results.

[0033] As a further aspect of the present invention: the specific method for verifying the treatment effect in steps S3 and T3 is as follows: collect the response latency of the service layer, the IOPS and / or temperature of the device layer, and the error codes and / or call chain data of the software layer, and compare them with the Service Level Agreement (SLA) requirements (such as response latency ≤100ms, device temperature ≤80℃);

[0034] If the data meets the SLA requirements, the action is marked as valid and the confidence level of the corresponding fault-action in the knowledge graph is updated (confidence level increases by 0.05-0.2).

[0035] If the conditions are not met, the action is marked as invalid and the corresponding confidence level is reduced (confidence level decreases by 0.05-0.2).

[0036] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0037] 1. This invention synchronously collects multi-source heterogeneous data (such as equipment temperature, software logs, business response latency, etc.) from the device layer, software layer, and business layer through the perception layer, and constructs a spatiotemporally aligned feature matrix. This solves the limitation of traditional operation and maintenance relying only on data from a single level. Combined with the multimodal AI large model of the analysis layer (such as Transformer capturing temporal features and GNN modeling topological relationships), it can simultaneously mine the temporal dependence and hierarchical relationship of the data, accurately output the probability distribution of the root cause of the fault and the level of influence, avoid misjudgment or omission caused by the bias of data from a single level, and significantly improve the accuracy of fault location in complex systems.

[0038] 2. After generating the handling strategy, this invention verifies the effect by collecting system status data (such as business indicators and equipment parameters) in real time through the perception layer, and feeds the results back to the analysis layer to drive the model self-optimization. For example, if the business response delay after handling does not meet the standard, the system will update the confidence of the corresponding fault-action in the knowledge graph and regenerate a more suitable strategy. This mechanism enables the operation and maintenance system to continuously learn from the fault handling experience in different scenarios, gradually optimize the strategy library, effectively cope with the operation and maintenance challenges brought about by dynamic changes such as sudden increase in business traffic and equipment aging, and improve the long-term effectiveness of the handling strategy.

[0039] 3. Based on fault level and cost constraints, this invention comprehensively considers fault recovery time and resource costs (such as edge / central node resources in a cloud-edge collaborative architecture) through a multi-objective optimization function. It selects the resource that meets the Service Level Agreement (SLA) and has the smallest objective function value. This method avoids the extreme problems of "prioritizing high-cost resources to ensure timeliness" or "over-reliance on low-cost resources leading to recovery delays" in traditional operation and maintenance. It can ensure business continuity (such as e-commerce promotions and industrial production line operation) and reduce resource redundancy consumption, achieve a dynamic balance between operation and maintenance efficiency and economic costs, and improve the overall resource utilization rate. Attached Figure Description

[0040] Figure 1 This is an overall flowchart of the intelligent operation and maintenance method of the present invention;

[0041] Figure 2 This is a flowchart of the fault handling method of the present invention;

[0042] Figure 3 This is a flowchart illustrating the implementation process of handling e-commerce promotion failures on the cloud platform according to the present invention. Detailed Implementation

[0043] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0044] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0045] Please see the appendix Figure 1 - Appendix Figure 3 This invention discloses an intelligent operation and maintenance method based on an AI large model, comprising the following steps:

[0046] S1. Multi-source heterogeneous data fusion: Collect multi-source heterogeneous data from the device layer, software layer and business layer through the perception layer, and construct a spatiotemporally aligned feature matrix.

[0047] S2. Cross-level fault analysis: The feature matrix is ​​processed by the multimodal AI big model of the analysis layer to output the probability distribution of fault root causes and the level of influence, and the fault judgment threshold is updated in real time through a dynamic confidence mechanism.

[0048] S3. Generation and verification of handling strategies: Based on the probability distribution of root causes of faults, handling strategies are generated. After execution, the system status data collected by the perception layer is used to verify the handling effect, and the feedback is sent to the analysis layer to drive the model self-optimization.

[0049] S4. Resource optimization scheduling: Based on fault level and cost constraints, select suitable operation and maintenance resources through a multi-objective optimization function.

[0050] In one embodiment of the present invention, the multi-source heterogeneous data in step S1 includes:

[0051] Device layer data: Temperature, power consumption, and IOPS (Input / Output Operations) collected by device sensors;

[0052] Software layer data: error codes, call chains, and memory usage in software logs;

[0053] Business layer data: QPS (queries per second), response latency, and user conversion rate collected by the business monitoring system;

[0054] The spatiotemporally aligned feature matrix is ​​constructed using the following formula:

[0055] ;

[0056] In the formula, This is a vector concatenation operation. Raw data, The mean of the data. Standard deviation This is a tensor splicing operation. Encode the timestamp. Encode the device location.

[0057] In one embodiment of the present invention, in the hybrid architecture, the output vector of the Transformer and the node embedding vector of the GNN are fused by the concatenation layer and then input into the fully connected layer. In step S2, the multimodal AI large model adopts a hybrid architecture of Transformer + Graph Neural Network (GNN). The Transformer is used to capture the long-term dependency features of time series data, and the GNN is used to model the topological relationship between equipment, software and business. The probability distribution of fault causes is the probability value of each fault cause in the equipment layer, software layer and business layer, and the influence level is at least one of the equipment layer, software layer and business layer.

[0058] In one embodiment of the present invention, the dynamic confidence mechanism in step S2 updates the fault determination threshold in real time using the following formula:

[0059] ;

[0060] In the formula, for The dynamic threshold at any given time. for Threshold of time, Forgetting factor (value range 0.7 ≤ ≤0.9 (determined by convergence experiments based on historical fault data). Number of recent samples (range 50 ≤ ≤200), For the first The actual fault error of a sample This represents the model prediction error.

[0061] In one embodiment of the present invention, the cross-level fault analysis in step S2 further includes calculating the cross-level impact index of equipment-software-business, as shown in the following formula:

[0062] ;

[0063] In the formula, Hierarchical weights (device layer) Values ​​range from 0.4 to 0.6, software layer. Values ​​range from 0.2 to 0.4, business layer. (Values ​​range from 0.1 to 0.3) hierarchical The index offset (i.e., the difference between the current value and the baseline value). hierarchical The maximum allowable offset of the indicator is determined by comparing the values ​​at each level. The value determines the level of the primary cause of the failure.

[0064] In one embodiment of the present invention, the multi-objective optimization function in step S4 is:

[0065] ;

[0066] In the formula, Weighting of fault recovery time (value range 0.5≤ ≤0.8), Resource cost weights (value range 0.2 ≤ ≤0.5), For fault recovery time, To reduce operational resource costs, the adapted operational resources include edge node resources and central node resources in the cloud-edge collaborative architecture. This is achieved by comparing the costs of different resources. and Select the resource that satisfies the SLA and has the smallest objective function value.

[0067] This invention also provides a fault handling method for intelligent operation and maintenance based on an AI large model, the fault handling method including the following steps:

[0068] T1. Receive the probability distribution of root causes of failures and the impact level output by the intelligent operation and maintenance method;

[0069] T2. Use the knowledge graph to generate a sequence of actions to handle the situation;

[0070] T3. After executing the sequence of actions, the system status data (including service latency, equipment indicators, and software logs) is collected through the perception layer to evaluate the effectiveness of the actions;

[0071] T4. If the effect is not met (i.e. the Service Level Agreement (SLA) requirements are not met), the handling result will be fed back to the analysis layer of the intelligent operation and maintenance method, the confidence of the corresponding fault-action in the knowledge graph will be updated, and the handling strategy will be regenerated.

[0072] In one embodiment of the present invention, the knowledge graph stores the mapping rules between fault types and handling actions, including but not limited to disk I / O overload → migration to high-performance disk, cache miss → enabling memory caching, and the handling action sequence optimizes the action order through reinforcement learning, giving priority to actions with low cost and quick results.

[0073] In one embodiment of the present invention, the specific method for verifying the treatment effect in steps S3 and T3 is as follows: collect the response latency of the service layer, the IOPS / temperature of the device layer, and the error code / call chain data of the software layer, and compare them with the Service Level Agreement (SLA) requirements (such as response latency ≤100ms, device temperature ≤80℃);

[0074] If the data meets the SLA requirements, the action is marked as valid and the confidence level of the corresponding fault-action in the knowledge graph is updated (confidence level increases by 0.05-0.2).

[0075] If the conditions are not met, the action is marked as invalid and the corresponding confidence level is reduced (confidence level decreases by 0.05-0.2).

[0076] Example 1: Handling Database Service Failures During E-commerce Promotions on Cloud Platforms

[0077] I. Scene Background

[0078] An e-commerce platform operated by a certain company carries online retail business nationwide. During the annual promotional period, the number of concurrent user visits surges (peak QPS exceeds 50,000 times / second), which places extremely high demands on the stability of the cloud platform's backend database service. At 10:00 on December 12, 2024, the platform monitoring system detected that the user-end API response time suddenly increased from the baseline value of 80ms to 600ms, and the business layer indicator (user order success rate) dropped from 99.5% to 85%, triggering the fault warning process of the intelligent operation and maintenance system.

[0079] II. Implementation Steps

[0080] (a) Multi-source heterogeneous data fusion (S1)

[0081] The perception layer of the intelligent operation and maintenance system collects data in the following ways:

[0082] Device layer: Disk IOPS of the database server cluster increased from 12,000 times / second to 28,000 times / second, CPU utilization increased from 70% to 95%, and memory usage increased from 65% to 88%.

[0083] Software layer: Error logs of Redis caching service (the number of "cache misses" increased from 500 times / minute to 8000 times / minute), and slow query logs of MySQL database (the proportion of SQL statements with execution time > 1 second increased from 3% to 25%).

[0084] Business layer: Client-side API response time (average 600ms), QPS (52,000 times / second), order success rate (85%).

[0085] After data collection, the system uses formulas Perform standardization and spacetime alignment:

[0086] For raw data (e.g., disk IOPS = 28000);

[0087] The baseline average for the past 7 days (baseline average disk IOPS = 12000).

[0088] Standard deviation (disk IOPS standard deviation = 2000);

[0089] Encode the timestamp (2024-12-12 10:00:00 is converted into a numerical vector);

[0090] Encode the server location (e.g., convert “East China-03-DB-01” into a unique identifier vector).

[0091] The final generated matrix contains 128-dimensional features. This is used for subsequent analysis.

[0092] (ii) Cross-level fault analysis (S2)

[0093] The analysis layer calls a "Transformer + GNN" multimodal AI large model for processing. :

[0094] Transformer module: Extracts time series features, identifies sudden surge patterns in business traffic (QPS increases by 400% within 30 minutes), and correlates with historical promotional data to confirm that the traffic surge is an abnormal event rather than a normal fluctuation;

[0095] GNN module: Constructs a topology graph of "database server - cache service - user API" and analyzes the dependencies between nodes (such as API calls needing to access the cache first, and accessing the database only when the cache is not hit).

[0096] Model outputs the root cause probability distribution of the fault:

[0097] Disk I / O overload (device layer): ;

[0098] Cache configuration invalid (software layer): ;

[0099] Business traffic anomaly (business layer): .

[0100] At the same time, through the cross-level influence index formula Calculate the influence (weight) of each level , , );

[0101] Equipment layer: , , ;

[0102] Software layer: , , ;

[0103] Business layer: , , .

[0104] The overall assessment concluded that the primary cause was device-level (disk I / O overload). maximum)

[0105] Furthermore, the dynamic confidence mechanism uses a formula Update threshold:

[0106] previous threshold ;

[0107] The mean absolute error of the most recent 100 samples is 0.15;

[0108] Calculated The impact level is below 0.4 at the equipment level, confirming the fault trigger.

[0109] (III) Generation and Verification of Disposal Strategies (S3)

[0110] Based on the knowledge graph (which stores rules such as "disk I / O overload → migrate to high-performance disk" and "cache miss → enable memory caching"), the system generates a sequence of actions to handle the situation.

[0111] 1. Temporary measure: Migrate frequently queried data (such as product details and prices) from the database to a Redis memory cache to reduce disk I / O pressure;

[0112] 2. Fundamental solution: Migrate the primary database instance from a regular cloud disk (IOPS capped at 20,000) to a high-performance SSD cloud disk (IOPS capped at 40,000).

[0113] After the strategy is executed, the perception layer continuously collects system status data:

[0114] Business layer: API response time decreased to 120ms after 10 minutes (SLA requirement ≤200ms), and order success rate rebounded to 98%;

[0115] Device layer: Disk IOPS dropped to 18,000 times / second (below the SSD cloud disk limit), and CPU utilization dropped to 80%.

[0116] Software layer: Redis cache misses reduced to 1000 times / minute, and slow query rate reduced to 5%.

[0117] The verification results met the SLA requirements. The system increased the confidence level of "disk I / O overload → migration to SSD cloud disk" from 0.75 to 0.88 and updated the knowledge graph.

[0118] (iv) Resource optimization scheduling (S4)

[0119] Through multi-objective optimization function ( Select resources:

[0120] Option A: Use the SSD cloud disk in the local availability zone. ;

[0121] Option B: Transfer high-configuration cloud disks across availability zones .

[0122] Calculate the objective function value:

[0123] Option A: ;

[0124] Option B: .

[0125] However, option B, which involves cross-regional migration, may result in temporary service interruptions (violating SLA ≥ 99.9%). Therefore, option A was ultimately chosen to ensure business continuity.

[0126] III. Implementation Results

[0127] In this fault handling, the system accurately located the root cause through cross-level analysis (location time < 2 min), avoided false alarms with dynamic thresholds (false alarm rate of 30% in traditional solutions, 0 false alarms in this embodiment), achieved an effectiveness rate of 92% in handling strategies (< 50% in traditional solutions), had a fault recovery time of only 8 min (average 15 min in traditional solutions), and improved the utilization rate of operation and maintenance resources to 85% (40% in traditional solutions), ensuring the stable operation of the platform during the promotion period.

[0128] Example 2: Troubleshooting PLC Controllers in Industrial IoT Production Lines

[0129] I. Scene Background

[0130] A smart manufacturing company's automotive parts production line relies on PLCs (Programmable Logic Controllers) for automated control. It includes 20 PLC controllers responsible for the coordinated operation of robotic arms, conveyor belts, and quality inspection equipment. At 14:30 on September 5, 2024, the production line monitoring system detected an abnormal shutdown of the quality inspection equipment (business layer: production line throughput dropped from 120 pieces / hour to 30 pieces / hour, and the yield rate dropped from 98% to 75%), triggering the fault handling process of the intelligent operation and maintenance system.

[0131] II. Implementation Steps

[0132] (a) Multi-source heterogeneous data fusion (S1)

[0133] The perception layer collects data through industrial IoT gateways:

[0134] Equipment layer: PLC controller temperature (increases from baseline 45℃ to 75℃), voltage (decreases from 24V to 22V), communication interface bandwidth (increases from 10Mbps to 25Mbps);

[0135] Software layer: PLC control program error codes ("ERR-03: Communication timeout"), ladder logic execution delay (increased from 5ms to 20ms);

[0136] Business layer: Robotic arm action completion time (increased from 2s to 5s), conveyor belt speed (decreased from 1.5m / s to 0.8m / s), and image recognition latency of quality inspection equipment (increased from 100ms to 500ms).

[0137] Data standardization and spatiotemporal alignment are used to generate a feature matrix. The timestamp code is "2024-09-05 14:30:00", and the equipment location code is "Production Line-02-PLC-07" (corresponding to the 7th PLC controller).

[0138] (ii) Cross-level fault analysis (S2)

[0139] The analysis layer calls the "Transformer+GNN" model for processing. :

[0140] Transformer module: Identifies abnormal temperature rise trends in PLCs (30°C increase within 30 minutes) and correlates them with historical data (similar PLCs have an 80% higher failure rate when the temperature is >60°C).

[0141] GNN module: Constructs a topology diagram of "PLC-robotic arm-conveyor belt-quality inspection equipment" and finds that the PLC communication delay (20ms) causes the robotic arm's action command to be sent late, which in turn causes the conveyor belt speed to decrease and the quality inspection equipment to stop.

[0142] Model outputs the root cause probability distribution of the fault:

[0143] PLC power module aging (equipment level): ;

[0144] Controlling program logic conflicts (software layer): ;

[0145] Production line load surge (business level): .

[0146] Cross-level influence index calculation (weight) , , );

[0147] Equipment layer: (The maximum allowable temperature deviation of the PLC is 30℃). ;

[0148] Software layer: , (The maximum allowed offset for logical execution delay is 20ms). ;

[0149] Business layer: 100 (Maximum allowable offset of 100 items / hour for throughput). .

[0150] The primary cause was determined to be aging of the equipment layer (PLC power module). ).

[0151] Dynamic confidence mechanism updates threshold :

[0152] previous threshold ;

[0153] The mean absolute error of the recent 80 samples is 0.12;

[0154] The impact level is below 0.6 at the equipment level, confirming the fault trigger.

[0155] (III) Generation and Verification of Disposal Strategies (S3)

[0156] The knowledge graph invokes rules such as "PLC temperature abnormality → check power module" and "communication delay → replace communication interface" to generate a sequence of actions to handle the situation.

[0157] 1. Temporary measures: Switch to the backup PLC controller (same model, pre-configured control program) and restore basic operation of the production line;

[0158] 2. Fundamental solution: Disassemble the faulty PLC, inspect the power supply module (found bulging capacitors), replace with a power supply module of the same specification, and calibrate the voltage output (restoring it to 24V±0.5V).

[0159] After the strategy is executed, the perception layer verifies the effect:

[0160] Business level: Production line throughput has rebounded to 115 units / hour (SLA≥100 units / hour), and yield rate has recovered to 97%;

[0161] Equipment level: PLC temperature dropped to 48℃, voltage stabilized at 24V, and communication interface bandwidth dropped to 12Mbps;

[0162] Software layer: Control program logic execution latency reduced to 6ms, with no error code output.

[0163] The verification results met the criteria, and the system increased the confidence level of "PLC temperature abnormal → power module replacement" from 0.7 to 0.85 and updated the knowledge graph.

[0164] (iv) Resource optimization scheduling (S4)

[0165] Through multi-objective optimization function ( Select resources:

[0166] Option A: Use the backup power modules in the production line warehouse. ;

[0167] Option B: Urgently procure new modules from the supplier .

[0168] Calculate the objective function value:

[0169] Option A: ;

[0170] Option B: .

[0171] However, Option B has a long procurement time (exceeding the SLA-required fault recovery time of ≤60min), so Option A is selected to ensure the production line can be restored quickly.

[0172] III. Implementation Results

[0173] In this failure, the system quickly located the aging problem of the PLC power module through cross-level analysis (location time < 3 min), and the dynamic threshold avoided false alarms caused by temporary load increases on the production line (the false alarm rate of the traditional solution was 40%, while that of this embodiment was 0 false alarms). The effectiveness of the handling strategy reached 92% (the traditional solution was < 50%), the fault recovery time was 25 min (the traditional solution averaged 60 min), and the utilization rate of operation and maintenance resources was increased to 82% (the traditional solution was 38%), ensuring the continuous production of the production line and reducing direct economic losses by 450,000 to 550,000 yuan.

[0174] Example 3: Handling a sudden drop in the number of user connections at a 5G communication base station

[0175] I. Scene Background

[0176] On July 15, 2024, at 19:00 (evening peak hour), a base station (BS-07) in a core business district within the 5G network coverage area of ​​a certain city experienced an abnormal drop in user connections: from a baseline of 3200 users to 1200 users, and the service layer indicators (user uplink and downlink speeds) dropped from 300Mbps to 80Mbps, triggering the fault warning process of the intelligent operation and maintenance system. This base station supports the 5G communication needs of surrounding office buildings, shopping malls, and residential areas, and the drop in user connections directly affects the experience of real-time services such as video calls and mobile payments.

[0177] II. Implementation Steps

[0178] (a) Multi-source heterogeneous data fusion (S1)

[0179] The perception layer of the intelligent operation and maintenance system collects data through base station monitoring modules, network management systems, and user-side terminals:

[0180] Equipment layer: Temperature of base station baseband processing unit (BBU) (increased from baseline 55°C to 85°C), transmit power of radio frequency unit (RRU) (decreased from 43dBm to 38dBm), power module voltage (decreased from -48V to -45V).

[0181] Software layer: Base station controller (CU / DU) logs (including error code "RRC-CONN-FAIL-007: Radio resource control connection establishment failed"), protocol stack processing latency (increased from 10ms to 30ms).

[0182] Business layer: Number of user connections (1200), uplink and downlink speeds (80Mbps), connection establishment success rate (decreased from 99% to 75%).

[0183] After data collection, the system uses formulas Perform standardization and spacetime alignment:

[0184] This is the raw data (e.g., BBU temperature = 85℃);

[0185] The baseline average for the past 30 days (average BBU temperature = 55℃);

[0186] Standard deviation (BBU temperature standard deviation = 8℃);

[0187] Encode the timestamp (convert 2024-07-15 19:00:00 into a time vector);

[0188] Encode the base station location ("City Center-Business District-07" is converted into a geographic identifier vector).

[0189] The final generated matrix contains 256-dimensional features. This is used for subsequent analysis.

[0190] (ii) Cross-level fault analysis (S2)

[0191] The analysis layer calls a "Transformer + GNN" multimodal AI large model for processing. :

[0192] Transformer module: Extracts time series features, identifies abnormal upward trends in BBU temperature (30°C increase in temperature within 1 hour), and correlates with historical data (when the BBU temperature of similar base stations is >70°C, the probability of a decrease in user connections increases by 60%).

[0193] GNN module: Constructs a topology diagram of "BBU-RRU-user terminal" and analyzes the dependencies between nodes (such as the delay in radio frequency signal processing caused by the rise in BBU temperature, which in turn affects the establishment of user connections).

[0194] Model outputs the root cause probability distribution of the fault:

[0195] BBU heat dissipation failure (device level): ;

[0196] Base station controller software malfunction (software layer): ;

[0197] A sudden surge in user traffic (business layer): .

[0198] At the same time, through the cross-level influence index formula Calculate the influence (weight) of each level , , );

[0199] Equipment layer: , (The maximum allowable deviation of BBU temperature is 30℃). ;

[0200] Software layer: (The maximum allowable offset for protocol stack processing delay is 25ms). ;

[0201] Business layer: (Maximum allowed offset for user connections is 2500 households). .

[0202] The overall assessment concluded that the primary cause was equipment-level (BBU heat dissipation failure). maximum).

[0203] Dynamic confidence mechanism through formula Update threshold:

[0204] previous threshold ;

[0205] The mean absolute error of the recent 150 samples is 0.1;

[0206] Calculated If the impact level is below 0.5 at the device level, the fault is confirmed to have been triggered.

[0207] (III) Generation and Verification of Disposal Strategies (S3)

[0208] Based on a knowledge graph (which stores rules such as "BBU temperature abnormality → check the heat dissipation system" and "connection failure → restart the controller"), the system generates a sequence of actions to handle the situation.

[0209] 1. Temporary measures: Switch some user traffic to adjacent base stations (BS-06, BS-08) to alleviate the current base station load;

[0210] 2. Fundamental measures: Inspect the cooling system of the BS-07 base station and find that the fan speed has dropped (from 3000 rpm to 1500 rpm) due to dust accumulation. Clean the fan and replace the damaged bearing.

[0211] After the strategy is executed, the perception layer continuously collects system status data:

[0212] Service layer: The number of user connections rebounded to 2,800 after 30 minutes (SLA≥2,500), and the uplink and downlink speeds recovered to 280Mbps;

[0213] Equipment layer: BBU temperature dropped to 60℃, fan speed returned to 3000 rpm, and RF unit transmit power stabilized at 43dBm;

[0214] Software layer: The processing latency of the base station controller protocol stack has been reduced to 12ms, and the error code "RRC-CONN-FAIL-007" has disappeared.

[0215] The verification results met the SLA requirements. The system increased the confidence level of "BBU temperature abnormal → clean cooling fan" from 0.72 to 0.86 and updated the knowledge graph.

[0216] (iv) Resource optimization scheduling (S4)

[0217] Through multi-objective optimization function ( Select resources:

[0218] Option A: Activate the base station maintenance team's local backup fan. ;

[0219] Option B: Transfer new fans from the regional warehouse .

[0220] Calculate the objective function value:

[0221] Option A: ;

[0222] Option B: .

[0223] Option A has a smaller objective function value and meets the SLA requirement (recovery time ≤ 60 min), so the local backup fan is selected to ensure rapid recovery of base station performance.

[0224] III. Implementation Results

[0225] In this fault handling, the system accurately located the BBU heat dissipation failure through cross-level analysis (location time < 2.5 min), and the dynamic threshold avoided false alarms caused by temporary electromagnetic interference due to surrounding construction (the false alarm rate of the traditional solution was 35%, while this embodiment had 0 false alarms). The handling strategy achieved an effectiveness rate of 92% (the traditional solution was < 50%), the fault recovery time was 45 min (the traditional solution averaged 90 min), and the utilization rate of operation and maintenance resources was increased to 83% (the traditional solution was 39%). This ensured the 5G communication experience for users in the business district and avoided an operating revenue loss of approximately 200,000 yuan / hour due to user churn.

[0226] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any variations and modifications can be made by those skilled in the art without departing from the spirit and scope of the invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the invention, fall within the protection scope defined by the claims of the present invention.

Claims

1. An intelligent operation and maintenance method based on an AI large model, characterized in that: The intelligent operation and maintenance method includes the following steps: S1. Multi-source heterogeneous data fusion: Collect multi-source heterogeneous data from the device layer, software layer and business layer through the perception layer, and construct a spatiotemporally aligned feature matrix. The multi-source heterogeneous data includes: Device layer data: Temperature, power consumption, and IOPS collected by device sensors; Software layer data: error codes, call chains, and memory usage in software logs; Business layer data: QPS, response latency, and user conversion rate collected by the business monitoring system; The spatiotemporally aligned feature matrix is ​​constructed using the following formula: ； In the formula, This is a vector concatenation operation. Raw data, The mean of the data. Standard deviation This is a tensor splicing operation. Encode the timestamp. Encode the device location; S2. Cross-level fault analysis: The feature matrix is ​​processed by the multimodal AI big model of the analysis layer to output the probability distribution of fault root causes and the level of influence, and the fault judgment threshold is updated in real time through a dynamic confidence mechanism. The multimodal AI large model adopts a hybrid architecture of Transformer and Graph Neural Network (GNN). The output vector of Transformer and the node embedding vector of GNN are fused by a concatenation layer and then input into a fully connected layer. Transformer is used to capture the long-term dependency features of time series data, and Graph Neural Network (GNN) is used to model the topological relationship between devices, software and business. The probability distribution of the root cause of the fault is the probability value of each fault cause in the device layer, software layer and business layer, and the influence level is at least one of the device layer, software layer and business layer. The dynamic confidence mechanism updates the fault determination threshold in real time using the following formula: ； In the formula, for The dynamic threshold at any given time. for Threshold of time, Forgetting factor, This represents the recent sample size. For the first The actual fault error of a sample This refers to the model prediction error; The cross-level fault analysis also includes calculating the cross-level impact index of equipment-software-business, as shown in the following formula: ； In the formula, For hierarchical weights, For the device layer, For the software layer, For the business layer, hierarchical The index offset. hierarchical The maximum allowable offset of the indicator is determined by comparing the values ​​at each level. The value determines the level of the primary cause of the failure; S3. Generation and verification of handling strategies: Based on the probability distribution of the root causes of the fault, a handling strategy is generated. After execution, the system status data is collected by the perception layer to verify the handling effect and the feedback is sent to the analysis layer to drive the model self-optimization. S4. Resource optimization scheduling: Based on fault level and cost constraints, select suitable operation and maintenance resources through a multi-objective optimization function; The multi-objective optimization function is: ； In the formula, As a weight for fault recovery time, As a resource cost weight, For fault recovery time, To reduce operational resource costs, the adapted operational resources include edge node resources and central node resources in the cloud-edge collaborative architecture. This is achieved by comparing the costs of different resources. and Select the resource that satisfies the SLA and has the smallest objective function value.

2. A method for intelligent operation and maintenance fault handling based on the method described in claim 1, characterized in that: The fault handling method includes the following steps: T1. Receive the probability distribution of root causes of failures and the impact level output by the intelligent operation and maintenance method; T2. Use the knowledge graph to generate a sequence of actions to handle the situation; The knowledge graph stores the mapping rules between fault types and handling actions, and the sequence of handling actions is optimized through reinforcement learning to prioritize actions that are low-cost and quick to take effect. T3. After executing the sequence of actions, the system status data is collected by the perception layer to evaluate the effectiveness of the actions; The specific method for verifying the effectiveness of the treatment is as follows: collect response latency at the business layer, IOPS and / or temperature at the device layer, and error codes and / or call chain data at the software layer, and compare them with the service level agreement requirements; T4. If the results are not satisfactory, the handling results will be fed back to the analysis layer of the intelligent operation and maintenance method to update the confidence of the corresponding fault-action in the knowledge graph and regenerate the handling strategy. If the data meets the SLA requirements, mark the action as valid and update the confidence level of the corresponding fault-action in the knowledge graph; If the conditions are not met, the marking process is invalidated and the corresponding confidence level is reduced.

Images