An operation and maintenance knowledge intelligent question and answer interaction method, device and equipment based on an LLM large model

Abstract

The application discloses an operation and maintenance knowledge intelligent question and answer interaction method, device and equipment based on an LLM large model. The method receives a question from an operation and maintenance personnel through a natural language interaction interface, analyzes semantics by using a pre-trained LLM to generate a candidate answer, verifies the accuracy of the answer in combination with a dynamic operation and maintenance knowledge graph (containing a device model and fault code mapping, a historical case library and a real-time updating knowledge base), and finally returns a verification result. The device contains an input module, a processing module, a verification module and an output module, wherein the processing module integrates field knowledge fine-tuning and real-time data set functions, and the verification module supports multi-source data fusion and confidence evaluation. The platform adopts a multi-tenant architecture, supports visual diagnosis, automatic device operation and knowledge usage report generation, and ensures that data cannot be tampered with through blockchain technology.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer software and information technology, specifically to an intelligent question-answering system based on natural language processing (NLP) and operation and maintenance knowledge management, and more particularly to an operation and maintenance knowledge intelligent question-answering interaction method, device and equipment based on a large language model (LLM). TECHNICAL BACKGROUND

[0002] With the development of natural language processing (NLP) technology, intelligent question-answering systems based on large models have been applied in multiple fields. For example, patent CN117633179A proposes an intelligent question-answering system for the railway industry, which constructs a vector database by fine-tuning the Qwen-7b model to realize automatic knowledge updating; patent CN117874202A uses anisotropic disambiguation and attention interaction technology to improve the accuracy of question-answering; patent CN118039057B constructs a multi-modal large model to realize intelligent interaction of a family health service robot; patent CN118194993A improves the intelligent level of digital campus question-answering through text vectorization and LLM model; and patent CN118332091A combines a multi-modal vector database and a word meaning matrix to optimize the question-answering of an ancient book knowledge base.

[0003] However, the existing technology still has significant limitations in the field of operation and maintenance knowledge management:

[0004] The knowledge updating mechanism is rigid;

[0005] For example, patent CN117633179A relies on instruction data set fine-tuning, which is difficult to adapt to the dynamic changes of device fault modes; and patent CN118194993A processes campus data through text vectorization, but does not realize real-time data integration, resulting in insufficient knowledge timeliness.

[0006] Multi-modal support is insufficient;

[0007] Patent CN117874202A only processes text data and cannot integrate multi-source information such as device monitoring logs and images; and patent CN118332091A constructs a multi-modal vector database, but does not realize deep collaboration of multi-modal interaction.

[0008] Real-time and automation capabilities are lacking;

[0009] Existing systems are mostly offline optimized (such as the family health service robot of patent CN118039057B), lack real-time data verification mechanisms, and are difficult to cope with sudden failure scenarios.

[0010] Cross-domain knowledge integration capability is weak;

[0011] Patents CN117633179A and CN118194993A are both aimed at specific fields (railway, education), and do not achieve cross-field knowledge transfer, unable to meet the needs of complex operation and maintenance scenarios.

[0012] The decision-making model lacks dynamic adjustment.

[0013] The similarity matching algorithm of patent CN117874202A relies on fixed weights, making it difficult to dynamically optimize the reasoning logic based on real-time data.

[0014] The present invention addresses the above-mentioned pain points and proposes an intelligent operation and maintenance knowledge question-answering system based on LLM. Through dynamic knowledge graph fusion, real-time data integration, and multi-modal interaction, it breaks through the limitations of traditional systems and realizes the intelligentization and automation of operation and maintenance knowledge management. SUMMARY

[0015] The core principle of the present invention is to combine the LLM large model with the operation and maintenance knowledge graph, implement user question understanding and answer generation through natural language processing technology, and verify and optimize the answer using the knowledge graph. At the same time, a dynamic updating mechanism is introduced to ensure the real-time and accuracy of the knowledge graph.

[0016] The following is described in detail in multiple aspects:

[0017] On the one hand, the intelligent question-answering method based on LLM, in an embodiment, the method flow provided by the present invention includes:

[0018] 1. Natural language interaction: Through the input layer, the natural language questions of the operation and maintenance personnel are received. In actual operation and maintenance scenarios, operation and maintenance personnel may describe problems in various ways, such as detailed device failure descriptions, which may mention abnormal sounds, indicator light states, data display abnormalities, etc. during device operation. The target detection model for image input is the YOLO series model (such as YOLOv5), which analyzes the color proportion of the indicator light in the HSV color space (such as 82% red indicator light proportion), and reads the panel error code through OCR technology. The noise reduction processing of voice input adopts VAD noise reduction algorithm, and the speech-to-text adopts CTC algorithm, and the recognition accuracy of noise-reduced voice reaches 98% (such as "NX5500 storage array IOPS sudden drop" voice input to text case).

[0019] 2. Semantic Analysis and Candidate Answer Generation: The processing layer utilizes a pre-trained LLM model to analyze the problem semantics. The LLM model understands each word and sentence structure in the input question from the operations and maintenance personnel, identifies key information such as device model, fault phenomenon, etc. For example, when the operations and maintenance personnel describe "Server A frequently crashes when running large programs", the LLM model can accurately identify the device as "Server A" and the fault phenomenon as "frequent crashes when running large programs". Then, based on these analysis results, the model generates candidate answers using its vast knowledge base. These candidate answers may be based on similar past cases or suggestions inferred from relevant knowledge.

[0020] 3. Dynamic Knowledge Graph Verification: The verification layer calls a dynamic knowledge graph containing device model and fault code mapping, historical case library, and real-time knowledge base. The dynamic knowledge graph is like a real-time updated "knowledge treasure", which integrates multiple aspects of information. Among them, the device model and fault code mapping can help quickly locate possible fault causes, the historical case library records various faults and corresponding solutions that have occurred in the past, and the real-time knowledge base is updated according to the latest device running data and industry knowledge. Through the multi-source data fusion engine, the system verifies the accuracy of the candidate answers. The multi-source data fusion engine comprehensively considers various data sources to evaluate and judge the candidate answers.

[0021] 4. Confidence Assessment and Output: Calculate the answer confidence through the confidence assessment algorithm. The confidence assessment algorithm is like a "credibility evaluator" that quantitatively evaluates the credibility of the answer based on various factors. If the confidence of the answer is higher than the threshold, it is directly output, and the output answer has high reliability and can provide effective solutions for operations and maintenance personnel. If the confidence is lower than the threshold, trigger manual review, which is like a "safety line" to ensure that when the system cannot determine the accuracy of the answer, professional operations and maintenance personnel can make further judgments and handle it.

[0022] Formula 1: Confidence assessment algorithm,

[0023] where (ω i ): data source weight (device model mapping 0.4, historical cases 0.3, real-time knowledge base 0.3), set according to historical accuracy (device model mapping 70%) and data importance. (Score iData Source Matching Score (0-1), calculated based on exact match degree (device model), case similarity (historical cases, Apriori algorithm), real-time data consistency (real-time knowledge base) (e.g. in the core switch case, device model matching score = 0.9, historical case score = 0.75). RealTimeDataConsistency (δ): Real-time data consistency coefficient (0.1-0.5), measured by Pearson correlation coefficient to measure the consistency of real-time data and historical failure patterns (e.g. BGP session oscillation and CPU peak synchronization δ = 0.4).

[0024] For example, device model mapping usually has a more direct correlation with specific failure causes, so the weight is relatively high; while the real-time knowledge base can reflect the latest device status and industry dynamics, also has important reference value. (δ) is the real-time data consistency coefficient (value range 0.1 to 0.5), which is mainly used to measure the matching degree of the answer and the real-time device data. If the answer is highly consistent with the real-time data, the real-time data consistency coefficient will increase, thereby improving the confidence of the answer.

[0025] On the other hand, the intelligent question answering device and platform architecture, in an embodiment, the device module interaction as shown, each module cooperates with each other, and jointly completes the processing task of intelligent question answering.

[0026] Input module: support text, voice, image multi-modal input. In actual application, operation and maintenance personnel can choose different input methods according to their own convenience. For example, when in a noisy environment, voice input may be more convenient; while for some complex device failures, image input may be more intuitive to show the problem. The input module meets the input needs of different users in different scenarios.

[0027] Processing module: integrate field knowledge fine-tuning sub-module and real-time data integration sub-module. The field knowledge fine-tuning sub-module is based on historical dialogue data for supervised learning, enhancing the model's understanding of domain terminology. For example, for some specific operation and maintenance terms, such as "E101 fault code" corresponding to "sensor anomaly", through continuous learning and training, the model can more accurately understand and process these terms. The real-time data integration sub-module is like a "data bridge", which can integrate device monitoring data in real time and dynamically update the reasoning logic. For example, combined with the real-time parameters such as temperature and pressure of the device, the model can correct the fault diagnosis result and improve the accuracy of diagnosis.

[0028] Verification module: verify the answer through multi-source data fusion engine and confidence evaluation algorithm, and trigger manual review if the confidence is lower than the threshold. The verification module strictly verifies and evaluates the answer generated by the processing module to ensure that the output answer has high reliability.

[0029] Output module: return the verification result or the answer after manual review. The output module presents the verified or manually reviewed answer to the operation and maintenance personnel in a clear and understandable manner, facilitating their subsequent fault handling operations.

[0030] Formula 2: LLM field knowledge fine-tuning loss function,

[0031] L = λ·L supervised +(1-λ)·L reward , wherein (L supervised ) is a supervised loss trained based on a field corpus of 12,000 device manuals and 8,000 fault cases, which strengthens the model's accurate analysis of operation and maintenance field terms by labeling the mapping relationship of "device model-fault code" (such as "NX5500→V2.1.3 firmware defect") and "professional term-standard expression" (such as "circuit breaker tripping→overload protection"), and reduces errors such as "device model identification error" and "fault phenomenon misjudgment". (L reward ) dynamically adjusts model parameters according to real-time sensor data (such as UPS battery internal resistance anomaly and temperature sampling frequency fluctuation), and when detecting changes in device fault modes (such as a 20% increase in the failure rate of a certain type of UPS battery), it strengthens the weight of related reasoning paths; if noise data (such as sensor abnormal fluctuations) is detected, the data quality filter automatically reduces the weight of this data source by 10%-20%, preventing invalid data from interfering with reasoning. (λ) is a balance coefficient (default 0.7), with a default value of 0.7, indicating that the model should first focus on accumulating field knowledge during the supervised learning stage (accounting for 70% of the weight) to ensure that it first masters basic operation and maintenance knowledge such as "device model and fault code mapping" and "historical case solutions", and then optimizes the reasoning logic through real-time data (accounting for 30% of the weight).

[0032] Further innovative design:

[0033] 1. Dynamic knowledge graph update strategy;

[0034] Formula 3: Knowledge graph version number management,

[0035] wherein (Δt i ) is the (i)th data update time interval, which records the time elapsed from the last update to the current update. (T update ) is the preset update period (default 24 hours), which is a fixed time period set according to actual conditions, used to measure the frequency of data updates. (Weight i) is to update the importance weight of data (such as fault case weight 0.8, industry standard weight 0.6). Different types of update data have different importance, and by setting the weight, the influence degree of these data on the knowledge graph can be more reasonably reflected.

[0036] 2. Multi-tenant data isolation mechanism,

[0037] Formula 4: Multi-tenant data isolation mechanism, wherein (AccessRisk i ) is the access risk score of the (i)th tenant (calculated based on permission level, operation frequency, etc.). The higher the permission level and the more frequent the operation frequency, the higher the access risk score of the tenant may be. (MaxRisk) is the maximum access risk value, which is a preset upper limit value for measuring the highest degree of tenant access risk. (S) is closer to 1, indicating that the isolation is safer, and through this safety coefficient, the safety of multi-tenant data isolation can be intuitively evaluated.

[0038] Other innovative designs:

[0039] 1. Automatic execution interface;

[0040] Formula 5: Operation instruction parsing accuracy,

[0041] wherein (CorrectCommands) is the number of correctly parsed operation instructions, and (TotalCommands) is the total number of operation instructions. Through this formula, the parsing accuracy of the automatic execution interface for operation instructions can be accurately measured.

[0042] 2. Blockchain storage;

[0043] Formula 6:

[0044] wherein (UpdateDate) is the knowledge update data, which contains the specific content of the knowledge graph update. (Timestamp) is the timestamp, used to record the time of the update. (Nonce) is a random number, used to increase the randomness and security of the hash value. Through SHA256 algorithm, these information is hashed to generate a unique hash digest.

[0045] 3. Abnormality detection module;

[0046] In one embodiment, the anomaly detection module analyzes dialogue history data to identify high-frequency issues and generate knowledge update suggestions. Its core process is as follows: Dialogue history analysis: Extracting key features from the dialogue, including device model, fault symptoms, and solutions; using the TFIDF algorithm to calculate term weights and identify high-frequency fault keywords (such as "power module fault" appearing more frequently than a threshold). High-frequency issue identification: Statistically analyzing the issue distribution per unit time using a sliding window algorithm.

[0047] Formula 7:

[0048] Frequency(q) = TotalConversationsCount(q) × 100%. If the frequency of a problem exceeds a preset threshold (e.g., 20%), it is marked as a high-frequency problem. Generate update suggestion step S43 (see reference). Figure 13 ): Discover the correlation between high-frequency problems and solutions through association rule mining (such as the Apriori algorithm); generate structured suggestions (such as "suggest adding 'power module replacement steps' to the knowledge base"), and push them to the knowledge graph update module.

[0049] Beneficial effects:

[0050] The innovative design of this invention brings about many significant benefits, comprehensively improving the efficiency, accuracy, security and intelligence of operation and maintenance knowledge management, and has great advantages over traditional solutions.

[0051] 1. A dynamic knowledge update mechanism addresses the pain points of manual maintenance;

[0052] Traditional systems rely on manual updates to the knowledge base (e.g., patent CN109800310A requires manual triggering of queries), resulting in data lag. This invention addresses this through a real-time data integration submodule, step S10 (see reference). Figure 3 With a version number management mechanism, the knowledge graph can be updated incrementally in seconds.

[0053] 2. LLM-driven intelligent interaction breaks through the limitations of static rules;

[0054] Traditional rule engines (such as patent CN201810490177.2) rely on keyword matching and cannot handle complex semantics. This invention enables the LLM model to deeply understand operation and maintenance terms (such as "E101 fault code" corresponding to "sensor anomaly") through domain knowledge fine-tuning and context label management, and supports multi-turn dialogue.

[0055] 3. Adaptive decision-making model to address dynamic scenario requirements;

[0056] Static rule system (such as patent CN117610854A) cannot adapt to device state changes. The invention fuses real-time data consistency coefficient δ (0.10.5) through confidence evaluation algorithm to dynamically adjust reasoning logic.

[0057] 4. Deep fusion of multi-source data, breaking data silos;

[0058] Traditional systems only process single text data (such as patent CN109800310A), and cannot integrate multi-modal information such as images and logs. The invention cross- validates the candidate answers generated by LLM with device monitoring data and image recognition results through the multi-source data fusion engine (S23).

[0059] 5. Enhanced security and personalization to meet industry compliance requirements;

[0060] Traditional shared knowledge base (such as patent CN202210010940.3) has data leakage risk. The invention reduces customer data leakage risk through multi-tenant data isolation architecture and blockchain storage, and ensures traceability of knowledge update process.

[0061] Description of the drawings

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0063] Figure 1 is an LLM-based operation and maintenance knowledge intelligent question and answer system architecture process schematic diagram in an embodiment of the present application;

[0064] Figure 2 is a dynamic knowledge graph verification process schematic diagram in an embodiment of the present application;

[0065] Figure 3 is an LLM domain knowledge fine-tuning process schematic diagram in an embodiment of the present application;

[0066] Figure 4 is a dynamic knowledge graph updating process schematic diagram in an embodiment of the present application;

[0067] Figure 5 is a context tag management process schematic diagram in an embodiment of the present application;

[0068] Figure 6 is an intelligent question and answer device module interaction process schematic diagram in an embodiment of the present application;

[0069] Figure 7This is a schematic diagram of the answer confidence evaluation algorithm in one embodiment of the present invention;

[0070] Figure 8 This is a schematic diagram of the intelligent question-and-answer platform process in one embodiment of the present invention;

[0071] Figure 9 This is a schematic diagram of the equipment status heatmap generation process in one embodiment of the present invention;

[0072] Figure 10 This is a schematic diagram of the automated execution interface process in one embodiment of the present invention;

[0073] Figure 11 This is a schematic diagram of a knowledge graph weight adjustment process based on user feedback in one embodiment of the present invention;

[0074] Figure 12 This is a schematic diagram of a multi-tenant data isolation architecture process in one embodiment of the present invention;

[0075] Figure 13 This is a schematic diagram of the data flow of the anomaly detection module in one embodiment of the present invention;

[0076] Figure 14 This is a schematic diagram of a multimodal interaction interface process in one embodiment of the present invention;

[0077] Figure 15 This is a schematic diagram of the blockchain knowledge update record process in one embodiment of the present invention. Detailed Implementation

[0078] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] In one embodiment, such as ( Figure 1 S01-S04, Figure 2 S05-S07, Figure 6 S19-S22, Figure 9 S38-S40 Figure 14 As shown in S32-S34), the system uses a natural language interaction interface in step S32 (see reference). Figure 14 The system receives user input regarding the issue of "NX5500 storage array experiencing a sudden drop in IOPS at 3 AM daily". Figure 1 S01). Input module ( Figure 6 After converting the speech input to text (S19), environmental noise interference is removed through a preprocessing pipeline. Figure 14, S34). The processing module ( Figure 6 , S20) activates the fine-tuned LLM model ( Figure 3 , S09) that has been loaded with a domain dataset containing 15,000 historical work orders ( Figure 3 , S08), and identifies the key features of "IOPS crash" and "timing task", "firmware defect", etc. The LLM generates three candidate answers: "check the RAID card firmware version", "review the timing maintenance task", and "detect the hard disk backplane voltage" ( Figure 2 , S05). At this time, the verification module ( Figure 6 , S21) starts the multi-source verification process ( Figure 2 , S06): first, it matches the known firmware defect library of the device model NX5500 (weight 0.4) and finds that the V2.1.3 version has a memory leak problem; second, it retrieves the historical case library (weight 0.3) and finds three similar cases pointing to the firmware upgrade solution; finally, it accesses the real-time monitoring system ( Figure 9 , S38) to verify that the current firmware version is indeed V2.1.3 and the memory usage rate reached 98% during the fault period.

[0080] The confidence is calculated by formula 1:

[0081] The confidence is calculated by formula 1: Figure 6 , S22) returns the structured answer: "immediately upgrade to V2.1.5 firmware, which needs to be executed in the maintenance window, and backup the configuration before upgrading" ( Figure 8 , S51). Preferably, the context label management module ( Figure 5 , S29) synchronously records the device serial number, fault time pattern, and solution hash value ( Figure 15 , S37), providing traceability for subsequent automated upgrade operations ( Figure 10 , S17). Further, the system triggers the automated execution interface ( Figure 10 , S18) when outputting the answer, pre-generating a work order template containing the firmware download link and operation manual, which is directly submitted to the device management system after confirmation by the operation and maintenance personnel.

[0082] In an embodiment, in combination with the fine-tuning mechanism of Figure 3 (S08-S10), the system faces the challenge of insufficient domain knowledge in the cold start stage. The processing module ( Figure 6 , S20) first loads the basic LLM model ( Figure 3 , S07), and then injects the operation and maintenance corpus containing 6 major domains of power, communication, data center, etc. ( Figure 3, S08), covering 12,000 pages of device manuals and 8,000 fault cases.

[0083] The domain knowledge fine-tuning submodule adopts a two-stage training strategy:

[0084] The first stage uses formula 2: L = λ·L supervised +(1-λ)·L reward , where the supervised loss (L supervised ) is strengthened by contrastive learning for professional term mappings such as "circuit breaker tripping → overload protection"; the second stage introduces a real-time data reward mechanism (L reward ), which dynamically increases the weight of related reasoning paths when monitoring abnormal increases in the failure rate of a certain type of UPS battery. Preferably, during the fine-tuning process, progressive distillation technology is used to encode the device topology relationships in the knowledge graph ( Figure 2 , S06) into attention masks ( Figure 3 , S09), enabling the model to automatically associate the physical relationship between the stack cable specifications (SFP+28AWG) and signal attenuation when processing "core switch stack abnormality" problems. Furthermore, the real-time data integration submodule ( Figure 3 , S10) is designed with a data quality filter ( Figure 4 , S26) that automatically reduces the weight coefficient δ of the data source when detecting abnormal fluctuations in sensor data (such as abnormal temperature sampling frequency), preventing noise data from interfering with the reasoning results.

[0085] On the other hand, in an embodiment, as shown in Figure 5 (S29-S31) and Figure 7 (S24), the system exhibits context awareness when processing multiple rounds of complex conversations. When the user first asks about "A3 computer room temperature control system alarm," the processing module ( Figure 6 , S20) generates an answer "check the precise air conditioner refrigerant pressure" (confidence 0.78). In the subsequent follow-up question "how to determine if the compressor is damaged," the context label management module ( Figure 5 , S29) extracts the device identification label "Liebert PEX4#07" and associates it with the fault mode library (weight 0.6) of this type in the knowledge graph ( Figure 2 , S06). The verification module ( Figure 6 , S21) simultaneously accesses the real-time sensor network ( Figure 9 , S38) to obtain compressor current waveform data ( Figure 14 , S33), and through feature extraction, it discovers abnormal starting current ( Figure 14, S34). At this time, the confidence evaluation algorithm is calculated by formula 1, and the weight distribution is dynamically adjusted: the real-time data weight is increased from 0.3 to 0.5, and the historical case weight is reduced to 0.2, and the final confidence is 0.83. The output module ( Figure 6 , S22) returns a multi-modal answer: text guidance "measure the starting capacitance capacity", schematic diagram showing the comparison of normal / abnormal current waveforms ( Figure 9 , S39), and maintenance video link. Preferably, the system automatically generates a preventive maintenance suggestion ( Figure 8 , S51), prompting the remaining life prediction value of the same batch of compressors (calculated based on the LSTM model as 142 days).

[0086] In still another embodiment, as shown in Figure 6 (S19-S22) and Figure 10 (S17-S18), the intelligent question and answer device realizes a closed-loop control function. When processing the "network port optical module frequent disconnection" problem, the input module ( Figure 6 , S19) receives an image containing the SFP+ module serial number ( Figure 14 , S33), and the optical character recognition sub-module ( Figure 14 , S34) extracts the model information of "AXM763-50-31". The processing module ( Figure 6 , S20) queries the compatibility database ( Figure 2 , S06), and finds that there is a known compatibility problem (weight 0.7) between the module and the FWQ8923 chip of the switch. The verification module ( Figure 6 , S21) combines the CRC error count in the real-time log ( Figure 4 , S26) to confirm that the fault feature matching degree reaches 92%. At this time, the system not only returns the suggestion "replace the compatible optical module" (confidence 0.88), but also generates operation instructions with timing control through the automated execution interface ( Figure 10 , S17): 1) enter maintenance mode at 1am; 2) automatically issue the disable port command (CLI: interface shutdown); 3) push the new module procurement list to the procurement system ( Figure 10 , S18).

[0087] It should also be noted that, in a preferred embodiment, as shown in Figure 2 (S06) and Figure 11 (S54-S55), the system realizes dynamic optimization of the knowledge graph. When processing the "UPS battery pack capacity drops too fast" problem, the initial answer "quarterly deep discharge maintenance" receives multiple user feedbacks ( Figure 11 , S54), 82% of which are marked as "invalid". The user behavior analysis module ( Figure 11, S55) starts the weight adjustment process: 1) reduces the weight of the solution in the historical case base (from 0.3 to 0.15); 2) increases the weight of the battery internal resistance detection index in the real-time monitoring data (from 0.2 to 0.35); 3) adds an environmental temperature correlation rule ( Figure 4 , S27), establishing a relationship edge of "computer room temperature > 25°C → battery aging acceleration 3 times". In the next processing of similar problems, the system preferentially recommends "installing a constant temperature cabinet and using a colloidal battery" (confidence 0.79), and this solution has been verified for 6 months to reduce the failure rate by 67%. Preferably, the version number V2.1.7 is automatically generated during this process, and the metadata (weight adjustment parameters, number of new rules, etc.) are recorded (

[0088] Figure 15 , S37). Further, the multi-tenant isolation mechanism ( Figure 12 , S12) ensures that the data updates of the financial customers (such as battery safety standards) will not be leaked to the manufacturing customers, and the data isolation coefficient S = 0.93 is calculated by formula 4:

[0089] , which meets the ISO27001 security requirements, where S represents the data isolation coefficient, which is used to measure the security of multi-tenant data isolation. The closer the S value is to 1, the safer the isolation, and K represents the number of tenants (i.e. the number of tenants participating in the calculation).

[0090] In an embodiment, as shown in Figure 2 (S06) and Figure 7 (S23-S24), the system realizes cross-validation of complex faults through a multi-source data fusion engine. When processing the problem of "periodic CPU utilization of core switch soaring to 95%", the processing module ( Figure 6 , S20) first generates two candidate answers "check routing protocol oscillation" and "check DDoS attack" ( Figure 2 , S05). The verification module ( Figure 6 , S21) starts a three-level verification process: first, matches the known vulnerability library of the device model Cisco Nexus9508 ( Figure 2 , S06), and finds that this model has a memory leak defect in the OSPF protocol implementation (weight 0.4); second, retrieves the historical case base ( Figure 2 , S06, weight 0.3), and finds that 9 out of 12 similar cases are protocol configuration errors; finally, accesses the NetFlow real-time traffic analysis system ( Figure 9 , S38), and detects that the BGP session oscillation frequency is completely synchronized with the CPU peak (delta = 0.4).

[0091] ​The confidence level is calculated using the following formula 1: 0.4 × 0.9 (model matching) + 0.3 × 0.75 (case matching) + 0.3 × 0.85 (real-time data) × 1.4 (δ enhancement) = 0.89. Preferably, a multi-source data fusion engine ( Figure 7 (S23) During this process, the weight allocation strategy is dynamically adjusted—when the real-time data confidence level exceeds 0.9 for 5 consecutive minutes, its weight is increased from 0.3 to 0.5 to ensure timely response to sudden traffic anomalies. Figure 7 (S24). Furthermore, the system detected frequent BGP neighbor resets ( Figure 4 S26), automatically triggers knowledge graph ( Figure 2 The association rule in S06 corrected the answer to "Upgrade to IOSXR 7.3.2 and enable Graceful Restart" (confidence 0.92). This process involved collaborative analysis of 12 data sources, including SNMPtrap logs ( Figure 14 S33), traffic mirroring data packets ( Figure 14 (S34) and device CLI configuration snapshot ( Figure 10 (S17).

[0092] In one embodiment, combined with Figure 7 The confidence threshold mechanism (S24-S25) enables the system to demonstrate intelligent decision-making capabilities when handling edge scenarios. When maintenance personnel query "standby generator cannot automatically switch," the LLM generates the answer "check the ATS control module" (initial confidence level 0.68). Verification module ( Figure 6 S21) detected the following contradiction: Knowledge Graph ( Figure 2 In the case of generator model S06, there were 32 similar faults (weight 0.3), but the real-time sensor showed normal fuel pressure (δ=0.2). Using formula 1, the final confidence level was calculated to be 0.68×(0.3×1.0+0.3×0.8+0.4×0.6)=0.65.

[0093] A value below the threshold of 0.7 triggers a manual review process. Figure 7 (S25). During the review process, experts found that LLM had not considered the logical changes of the new firmware and manually added a solution to "upgrade ATS firmware to V3.1.7" ( Figure 4 S28). Preferably, the system initiates a dual verification mechanism during this process: 1) Blockchain evidence storage module ( Figure 15 S37) Record the entire audit process log and generate an audit trail formula 6 containing timestamps and digital signatures: 2) Multi-tenant isolation module ( Figure 12 S12) Ensure that the knowledge update is only pushed to the customer group using the same model of equipment (isolation coefficient S = 0.91).

[0094] It should also be noted that, in a preferred embodiment, such as Figure 10 (S17-S18) and Formula 5: As shown, the automated execution interface implements a closed loop for operation and maintenance. When the system suggests "expanding the etcd node of the K8s cluster", the automation engine ( Figure 10 S17) Perform the following chain of operations: 1) Call Terraform to create 3 VMs (specifications: 8C16G); 2) Deploy the etcd 3.5 binary file via Ansible; 3) Configure the SSL certificate ( Figure 10 , S18); 4) Perform cluster health checks (timeout threshold 300 seconds). The entire process is calculated using formula 5, with the following command parsing accuracy: successful VM creation (1 / 1), correct version deployment (1 / 1), and correct certificate configuration (1 / 1), accuracy = 3 / 3 × 100% = 100%. Preferably, the system design includes a rollback mechanism—when an etcd cluster heartbeat anomaly is detected ( Figure 4 S26), automatically triggers the snapshot recovery process ( Figure 10 (S17), keeping the RPO within 15 seconds. Furthermore, the blockchain evidence storage module ( Figure 15 S37) generates verifiable credentials for each operation step (Formula 6), including: 1) VM creation timestamp; 2) Ansible Playbook hash value; 3) SSL certificate fingerprint. These credentials are synchronized to the auditing system via the OAuth 2.0 protocol. Figure 8 (S51), meeting the compliance requirements of the financial industry.

[0095] In one embodiment, such as Figure 12 As shown in (S12) and Formula 4, the multi-tenant data isolation mechanism achieves enterprise-level security control. When financial customer A and manufacturing customer B access the platform simultaneously, the user management module ( Figure 8 S53) Creates an independent knowledge base partition for each tenant. Figure 12 S12). The access control engine calculates tenant risk scores in real time: Customer A, due to its involvement in the core transaction system, has its access level set to P3 (AccessRisk_i = 0.15), while Customer B, being a research and development testing environment, has its access level set to P1 (AccessRisk_i = 0.35). Using formula 4, the isolation coefficient S = (1 - 0.15 / 0.5) × (1 - 0.35 / 0.5) = 0.85 × 0.3 = 0.255, triggering a strict isolation strategy—Customer A's knowledge update records ( Figure 4 S28) uses AES-256 encryption for storage. Figure 15 , S37), while customer B's data is only BASE64 encoded.

[0096] In one embodiment, combined with Figure 11 The feedback mechanism (S54-S55) enables the system to self-optimize the knowledge graph. When 20 users mark the "replace RAID card battery" solution as invalid ( Figure 11 S54), weight adjustment module ( Figure 11 S55) Initiate four-stage optimization: 1) Source tracing analysis and association of 50 similar work orders ( Figure 13 (S41) It was found that the battery model failed faster under high temperature environment; 2) Reduce the weight of this solution in the historical case library (from 0.3→0.18); 3) Increase the weight of environmental temperature monitoring data (from 0.2→0.4); 4) Add a knowledge graph relationship edge "computer room temperature > 28℃ → battery life shortening coefficient 1.8" Figure 2 S06). Preferably, the system synchronously updates the equipment maintenance strategy ( Figure 10 S17), when the temperature sensor ( Figure 9 When S38 detects a sustained high temperature, it automatically pushes a "install a temperature-controlled cabinet" suggestion (confidence level 0.91). Furthermore, the version management module ( Figure 4 The version number V3.2.1 is generated using formula 3, and the record includes metadata such as weight adjustment parameters and the number of new rules. Figure 15 (S37), and ensures the update process is auditable through blockchain notarization (Formula 6). Experimental data shows that the optimized solution adoption rate in this scenario increased from 58% to 89%.

[0097] On the other hand, in one embodiment, such as Figure 13 As shown in (S41-S43), the anomaly detection module enables proactive maintenance. The system analyzes 200,000 dialogue records daily. Figure 13 S41), using an improved TF-IDF algorithm (window size = 50 lines) to identify high-frequency terms. When the "optical module compatibility" problem occurs 47 times within 8 hours (frequency 23.5%), exceeding the threshold of 20% (Formula 7: Frequency(q) = TotalConversationsCount(q) × 100%), a three-level processing flow is triggered: 1) Association rule mining (Apriori algorithm) found that the problem has an 82% correlation with the switch model N9K-C9336TC-FX; 2) Knowledge graph ( Figure 2 (S06) Automatically marks this model's compatibility warning level as critical ( Figure 4 , S28); 3) Push the optical module replacement list to the procurement system ( Figure 8 (S51).

[0098] In other embodiments, such as Figure 14As shown in (S32-S34), the multimodal interactive interface revolutionizes the operation and maintenance process. When an engineer takes a photo of the equipment indicator light... Figure 14 When the image recognition submodule (S33) asks "Is the system status normal?", it will respond. Figure 14 S34) Perform the following processing: 1) Position the 4 LED indicators in the YOLOv5 model; 2) Analyze the HSV color space to show that the red indicator accounts for 82%; 3) Read the panel error code "E207" using OCR.

[0099] It should also be noted that, in a preferred embodiment, such as Figure 15 As shown in (S37) and Formula 6, the blockchain notarization mechanism ensures data credibility. When the knowledge graph updates the "UPS battery replacement cycle" rule ( Figure 4 S28), the evidence storage module executes: 1) Extracting updated content (including 32 parameter changes); 2) Generating a timestamp "2025-03-15T08:23:17Z"; 3) Creating a random number Nonce = 0x7A3F19; 4) Calculating the hash value Hash = SHA256(UpdateData⊕Timestamp⊕Nonce) = a1b2...f9e0. This hash value is written to Hyperledger Fabric via a smart contract. Figure 15 (S37), forming an immutable record containing 6 confirmation nodes. Preferably, the evidence storage system is designed with a cross-chain verification interface ( Figure 8 The system (S51) supports interoperability with public blockchains such as Ethereum and Corda, with a verification latency of <200ms. Furthermore, when auditors examine firmware upgrade records from December 2024, the system uses MerkleProof technology (…). Figure 15 (S37) provides a complete chain of evidence containing 17 related blocks within 3 seconds, meeting GDPR compliance requirements. Figure 12 (S12).

[0100] In one embodiment, such as Figure 8 As shown in (S51) and Formula 3, the system optimizes operational decisions through knowledge usage reports. At the beginning of each month, the report generation module (… Figure 8 S51) automatically analyzes 2.5TB of dialogue data ( Figure 13 S41) extracted three core indicators: 1) Distribution of hot issues (TF-IDF weighted ranking); 2) Solution effectiveness score (user feedback weighted); 3) Impact of knowledge graph updates (Formula 3, version number increment). Regarding the monthly hot issue of "storage latency" (occurring 27% of the time), the report ( Figure 8S51) Displays multi-dimensional analysis: a) Distribution of associated device models (NX5500 accounts for 68%); b) Comparison of solution adoption rates (firmware upgrade 92% vs. hardware replacement 35%); c) Improved resolution rate brought about by knowledge graph updates (from 71% to 89%). Preferably, the system uses association rule mining (Apriori algorithm) to discover a strong correlation between "latency peaks and backup tasks overlapping" (support 0.82, confidence 0.91), and automatically generates a strategy suggestion "adjust the backup window to the off-peak period". Figure 10 (S17). Furthermore, the report integrates a predictive model (LSTM) to warn of a potential 300% increase in NVMe SSD failure risk next quarter. Figure 9 (S38), it is recommended to purchase spare parts in advance. Figure 10 (S18), and stored through blockchain ( Figure 15 (S37) Ensure the procurement process is traceable.

[0101] In one embodiment, combined with Figure 4 The dynamic update mechanism (S26-S28) enables the system to achieve real-time evolution of the knowledge graph. When the sensor network ( Figure 9 (S38) An input voltage fluctuation of more than ±8% was detected in a certain model of UPS. Figure 4 S26), the dynamic update engine starts immediately: 1) Extract relevant fault cases from the past 24 hours ( Figure 13 , S41) 42 examples; 2) Calculate update priority (calculated using formula 3, Δt_i = 2 hours, Weight_i = 0.7); 3) Generate version number V3.4.1 (base version V3.4.0 + ∑(2 / 24 × 0.7) = 0.058). Update content ( Figure 4 S28) includes: adding a new rule for "voltage tolerance threshold ±5%" (confidence level 0.88), adjusting the fault weight of the rectifier module (0.3→0.5), and associating the mains power quality monitoring data source (δ coefficient +0.2).

[0102] Preferably, the update process triggers an automated verification process. Figure 10 S17): Simulate voltage fluctuation scenarios (±12%) in a sandbox environment to verify that the new rules accurately intercept 92% of abnormal cases. Figure 7 S24). Furthermore, the multi-tenant synchronization module ( Figure 12 S12) employs a differentiated push strategy, sending emergency updates only to 19 customer groups using this UPS model (isolation coefficient S>0.85). Figure 15 (S37) To avoid contamination of the knowledge base by irrelevant customers.

[0103] On the other hand, in one embodiment, such as Figure 7 (S25) and Figure 15As shown in (S37), the manual review closed-loop mechanism improves system reliability. When dealing with the "virtual machine host kernel crash" issue, LLM recommends "upgrading the KVM version" (confidence level 0.66), triggering a three-level review process: 1) Primary operations and maintenance checks the kernel logs ( Figure 14 S33), confirmed the existence of the "BUG:softlockup" error; 2) Senior engineers analyzed the core dump ( Figure 14 (S34), locating an NVMe driver defect; 3) The architect approved adding the "disable APST power-saving function" solution to the knowledge graph ( Figure 4 (S28). In other embodiments, such as Figure 9 (S38-S39) and Figure 14 As shown in (S33-S34), 3D visualization and multimodal data fusion improve diagnostic accuracy. When dealing with the "abnormal speed drop of all-flash arrays" problem, the visualization module ( Figure 9 S39) Generate a four-dimensional analysis view: 1) A three-dimensional spatial heat map displays the physical topology of the controller and SSD (XYZ coordinate accuracy ±2mm); 2) Color gradient represents IO latency (0-50ms, RGB(0,0,255)→(255,0,0)); 3) Dynamic particle flow displays data path congestion points; 4) AR overlay displays the SSD chip-level temperature distribution. Figure 14 (S34).

[0104] It should also be noted that, in a preferred embodiment, such as Figure 10 As shown in (S17-S18) and Equation 5, the automated execution interface achieves industrial-grade precision control. When handling a "hyperconverged cluster split-brain" fault, the system triggers a five-stage automated repair process. Figure 10 S17): 1) Isolate the faulty node (execute the CLI command "cluster quorum force"); 2) Create a consistent snapshot (generate a LUKS encrypted image); 3) Rebuild the quorum node (call Terraform to deploy 3 witness VMs); 4) Synchronize data (rsync differential synchronization, bandwidth limited to 10Gbps); 5) Verify cluster health (execute 473 check items). Calculated using Formula 5, the instruction parsing accuracy = 473 successful items / 473 total items × 100% = 100%. Preferably, the system design includes atomic operation rollback ( Figure 10 (S18) When any stage fails, the reverse operation chain is automatically executed (such as deleting the witness VM or restoring the snapshot) to ensure that RTO < 5 minutes.

[0105] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above-mentioned division of each functional unit and module is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0106] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. An LLM-based operation and maintenance knowledge intelligent question and answer method, characterized in that, Comprise the following steps: S1: receiving the question input by the operation and maintenance personnel through a natural language interaction interface supporting text, voice, and image multi-modal input, wherein the image input is recognized by a target detection model, and the voice input is processed by noise reduction and then converted into text; S2: analyzing the semantics of the question using a pre-trained LLM model to generate candidate answers, wherein the LLM model is optimized by: Two-stage fine-tuning based on a domain dataset containing a large number of historical work orders, the first stage uses supervised learning to strengthen professional term mapping, and the second stage introduces a real-time data reward mechanism; Encoding the device topology relationship in the knowledge graph into the model inference process through attention mask technology; S3: calling a dynamic operation and maintenance knowledge graph to verify the accuracy of the candidate answers; S4: returning the verified answers to the interaction interface through a structured output module, and the output content includes: Multi-modal information text, schematic diagram, and maintenance video link; Prediction value based on a device remaining life prediction model; Associated automatic execution instructions; S5: recording the conversation history and generating context labels containing device identification, fault time pattern, and solution hash value, identifying high-frequency fault keywords through a text feature extraction algorithm, and analyzing problem distribution using time window statistical techniques; S6: optimizing the subsequent dialogue process based on the context labels, establishing a cross-system correlation matrix to support complex problem decomposition, dynamically adjusting the weights of each node in the knowledge graph, and ensuring data security through a multi-tenant data isolation mechanism.

2. The LLM-based operation and maintenance knowledge intelligent question-answering method according to claim 1, characterized in that, In step S1, the target detection model for image input is a YOLO series model, the noise reduction processing for voice input uses a VAD noise reduction algorithm, and the voice-to-text conversion uses a CTC algorithm.

3. The LLM-based operation and maintenance knowledge intelligent question-answering method of claim 1, wherein, In step S2, the two-stage fine-tuning domain dataset covers different types and scenarios of operation and maintenance historical work order data, and the first stage of supervised learning is trained through the mapping relationship between professional terms and standard expressions, and the second stage of the real-time data reward mechanism adjusts the model parameters according to the feedback of real-time operation and maintenance data.

4. The LLM-based operation and maintenance knowledge intelligent question-answering method according to claim 1, characterized in that, In step S3, the version number management mechanism of the device model and fault code mapping library will be updated in real time according to the updates of the device and the changes of the fault code; the historical solution case library uses the Apriori algorithm to mine association rules; the real-time data consistency coefficient of the real-time updated knowledge base is used to measure the consistency degree of real-time data, historical data, and model prediction results.

5. The LLM-based operation and maintenance knowledge intelligent question-answering method according to claim 1, characterized in that, In step S3, the multi-level verification of the multi-source data fusion engine includes data layer verification, logic layer verification, and result layer verification, and the accessed device monitoring data includes but is not limited to NetFlow traffic and SMART logs.

6. The LLM-based operation and maintenance knowledge intelligent question-answering method according to claim 1, characterized in that, In step S4, the device remaining life prediction model is an LSTM model, and the automatic execution instructions include but are not limited to firmware upgrade work order templates and device restart instructions.

7. The LLM-based operation and maintenance knowledge intelligent question-answering method according to claim 1, characterized in that, In step S5, the text feature extraction algorithm is TFIDF algorithm, and the time window statistical technique uses sliding window technique to statistically analyze the problem distribution of different devices and different fault types within a unit time.

8. The LLM-based operation and maintenance knowledge intelligent question-answering method according to claim 1, characterized in that, The cross-system association matrix is established according to the service association and data interaction relationship between different operation and maintenance systems, and the basis for dynamically adjusting the node weight of the knowledge graph includes but is not limited to the frequency of occurrence of the question and the accuracy of the answer.

9. An LLM-based operation and maintenance knowledge intelligent question answering device, characterized in that, Comprise: A multi-modal input module for receiving questions input by operation and maintenance personnel through a natural language interaction interface supporting text, voice, and image multi-modal input; A semantic analysis and answer generation module that uses a pre-trained LLM model to analyze the semantics of the question and generate candidate answers, the LLM model being optimized by the method of any one of claims 1 to 3; An answer verification module that calls a dynamic operation and maintenance knowledge graph to verify the candidate answers, the knowledge graph and verification process being as claimed in any one of claims 1, 4 to 5; A structured output module that returns the verified answers to the interaction interface through the structured output module, the output content being as claimed in any one of claims 1, 6; A dialogue record and analysis module that records the dialogue history and generates context labels, and performs high-frequency fault keyword identification and question distribution analysis by the method of any one of claims 1, 7; A process optimization module that optimizes the subsequent dialogue process based on the context labels, establishes a cross-system association matrix, dynamically adjusts the node weight of the knowledge graph, and ensures data security through a multi-tenant data isolation mechanism, the method being as claimed in any one of claims 1, 8.

10. An electronic device, comprising: Comprise a processor and a memory, the memory storing a computer program, the processor executing the computer program to realize the steps of the LLM-based operation and maintenance knowledge intelligent question-answering method of any one of claims 1 to 8.

Images