An intelligent operation and maintenance fault diagnosis method based on problem reconstruction and consensus reasoning

Abstract

This invention provides an intelligent operation and maintenance fault diagnosis method based on problem reconstruction and consensus reasoning. The method includes: constructing a fault description rewriting model based on Direct Preference Optimization (DPO) by leveraging the generation capabilities and self-evaluation feedback of a large language model; reconstructing the original fault description in diverse ways based on the fault description rewriting model, and filtering non-standard fault descriptions into a set of fault query statements through a rejection sampling mechanism based on perplexity level (PPL); performing multi-perspective hybrid consensus diagnostic decision-making based on the real-time collected fault descriptions and fault query statement set, and outputting intelligent operation and maintenance fault diagnosis results corresponding to the real-time collected fault descriptions, including root cause localization and remediation suggestions. This invention uses the PPL rejection sampling mechanism to filter high-quality candidate rewrites that conform to operation and maintenance specifications, and utilizes a hybrid majority voting strategy to eliminate comprehension biases from a single perspective, achieving highly robust automated fault localization and diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology, and in particular to an intelligent operation and maintenance fault diagnosis method based on problem reconstruction and consensus reasoning. Background Technology

[0002] With the rapid development of artificial intelligence technology, large language models (LLMs), represented by the GPT (Generative Pre-trained Transformer) series, have achieved significant breakthroughs in the field of natural language processing. Leveraging massive amounts of pre-trained data and superior language understanding capabilities, large language models are gradually penetrating vertical industry applications, particularly demonstrating enormous potential in intelligent operations and maintenance (AIOps) and automated fault diagnosis. Enterprises expect to utilize large models as intelligent assistants to help operations and maintenance personnel quickly locate root causes from massive amounts of logs and fault phenomena, thereby improving the stability and maintenance efficiency of IT systems.

[0003] However, in practical machine reasoning and fault diagnosis tasks, although academia has proposed augmented reasoning methods such as Chain-of-Thought (CoT) and Self-Consistency to improve reliability by guiding the model to generate intermediate steps or multi-path voting, these methods still face severe challenges when dealing with input data from real-world operational scenarios. Existing augmentation techniques mainly focus on optimizing the "reasoning process" within the model, but neglect the fundamental problem at the "input end": there is a serious "semantic misalignment" between the user's (or front-line operations personnel's) natural language fault description and the standard operational knowledge system (cognitive framework) stored within the model.

[0004] Specifically, in fault reporting or consultation scenarios, users' original questions often contain colloquial expressions, non-technical terms, vague references, or even irrelevant noise (e.g., a user describes "the system just froze, spinning for a long time," rather than accurately stating "interface response timed out" or "database connection pool full"). Existing inference methods typically rely directly on these low-quality original descriptions for retrieval or reasoning. Due to the sensitivity of large models to professional operational terminology, such non-standard input can prevent the model from correctly activating its internal fault diagnosis knowledge, resulting in illusions or providing seemingly plausible but ultimately inaccurate general suggestions, failing to achieve precise fault localization. Even if the model possesses strong logical reasoning capabilities, it can still deviate from the correct diagnostic direction due to "incorrect input guidance."

[0005] Large Language Models (LLMs) are deep neural network models based on the Transformer architecture, obtained through large-scale pre-training on massive amounts of text data. Representative models include the GPT series, LLaMA, and Qwen. Inference enhancement techniques for LLMs aim to improve the model's performance in complex logical tasks by optimizing its input prompting or decoding strategies, without requiring large-scale retraining of the model parameters. Among these, chain-of-thought and self-consistency methods are the most typical.

[0006] One existing large-scale model reasoning scheme includes a knowledge graph-based large-scale model reasoning enhancement system. This system aims to address the reasoning bias problem caused by knowledge scarcity or logical complexity when large models handle domain-specific problems. The system mainly consists of three parts: a knowledge graph construction module, a knowledge retrieval module, and a reasoning enhancement module. Its specific workflow is as follows: First, the entity relationship analysis unit in the knowledge graph construction module parses the input data, extracting the entity network topology and preset semantic association marker nodes. Then, the knowledge diffusion evaluation unit performs modal decomposition (e.g., empirical modal decomposition) and noise reduction on the connection strength signals of nodes during continuous reasoning, calculating "reasoning state parameters" reflecting the activity and stability of the knowledge graph. Second, the knowledge retrieval module determines whether to trigger a retrieval based on these parameters. When necessary, it performs cross-dimensional supplementary knowledge retrieval from multi-source heterogeneous knowledge bases (e.g., medical ontology, electronic medical record databases, etc.) and establishes a mapping relationship between knowledge anchors and entity paths. Finally, the reasoning enhancement module generates the optimal reasoning path based on the retrieved knowledge and calculates the incremental parameters of the decision logic using an integral separation control algorithm, dynamically adjusting the real-time decision logic of the large model. This invention attempts to improve the adaptability of reasoning by introducing external knowledge graphs and combining them with complex signal processing mechanisms to dynamically intervene in the reasoning process.

[0007] The drawbacks of the aforementioned knowledge graph-based large-scale model reasoning enhancement system include: while the system utilizes external knowledge graphs to supplement the background information required for reasoning, the construction and maintenance of high-quality knowledge graphs are extremely costly, and it is difficult to cover rapidly changing open-domain scenarios in real time. Furthermore, this method only focuses on the injection of external knowledge, ignoring the mismatch (i.e., cognitive misalignment) between the original problem statement and the model's cognitive framework, and cannot resolve reasoning biases caused by ambiguity or inappropriate expression of the problem itself.

[0008] Another existing large-scale model inference approach includes a thought chain-based question-answering method. This method aims to improve the compliance and accuracy of large-scale models' outputs when handling subjective questions. The core of this method lies in extending the large-scale model's inference process from traditional linear output to a "structured inference chain." Specifically, this approach constructs a "target tree" structure containing inference stage types, content, and confidence levels by adding structured tags to the vocabulary. Simultaneously, the method imposes "budget constraints" (such as the number of inference steps and the number of words generated per step) on the structured inference chain and sets inference rewards to control the depth and breadth of generation. During the training phase, the method utilizes historical subjective question texts and their corresponding socially ethical target answer labels. It updates model parameters by calculating the "total stage loss" of the structured inference chain and the "total relative advantage" based on human ratings and KL divergence, thereby guiding the model to generate normative answers.

[0009] The drawbacks of one of the existing large-scale model question-answering methods based on thought chains include: although this method optimizes the organization and compliance of the reasoning process through structured reasoning chains and budget constraints, it still directly reasons based on the original question, failing to address initial comprehension biases caused by the mismatch between the question's wording and the model's cognitive framework. Furthermore, if the original input contains ambiguity, noise, or leading information, the structured reasoning path generated by this method may actually reinforce incorrect reasoning directions, failing to fundamentally improve the model's accuracy and robustness in understanding complex semantics. Summary of the Invention

[0010] The embodiments of the present invention provide an intelligent operation and maintenance fault diagnosis method based on problem reconstruction and consensus reasoning, so as to achieve effective automated fault location and diagnosis.

[0011] To achieve the above objectives, the present invention adopts the following technical solution.

[0012] An intelligent operation and maintenance fault diagnosis method based on problem reconstruction and consensus reasoning includes:

[0013] By leveraging the generative capabilities and self-evaluation feedback of large language models, a fault description rewriting model based on direct preference optimization of DPO is constructed.

[0014] Based on the fault description rewriting model, the original fault description is reconstructed in a variety of ways. Through the rejection sampling mechanism based on perplexity degree PPL, the non-standard fault description is transformed into a set of fault query statements.

[0015] Based on the real-time collected fault descriptions and the set of fault query statements, a multi-perspective hybrid consensus diagnostic decision is made, and an intelligent operation and maintenance fault diagnosis result containing root cause location and repair suggestions is output, corresponding to the real-time collected fault descriptions.

[0016] Preferably, the method of constructing a fault description rewriting model based on direct preference optimization (DPO) by utilizing the generative capabilities and self-evaluation feedback of a large language model includes:

[0017] A base language model is constructed, and multi-path sampling is performed on the input original fault description using the base language model to generate multiple candidate rewritten fault descriptions. A diagnostic correctness indicator function and a PPL function are introduced as dual evaluation indicators. Rewritten texts that can guide the base language model to draw correct diagnostic conclusions and conform to the operation and maintenance terminology specifications are marked as positive samples, while rewritten texts that cause the base language model to make incorrect diagnoses are marked as negative samples. A training set containing pairwise preference data is constructed using positive and negative samples. Based on the training set, the base language model is fine-tuned and optimized using the DPO algorithm to obtain a fault description rewriting model with the ability to map colloquial descriptions to standardized operation and maintenance terminology.

[0018] Preferably, the method of constructing a fault description rewriting model based on direct preference optimization (DPO) by utilizing the generative capabilities and self-evaluation feedback of a large language model includes:

[0019] Construct a foundational large language model for a given original fault description. Using a base-based large language model The second sampling generates a set of candidate rewrite fault descriptions. Obtain the fault reasoning and diagnosis results corresponding to each candidate rewritten fault description. The system incorporates filtering logic including between-group and within-group preferences. In the between-group preference setting, rewritten text that guides the model to generate the correct root cause of the failure is preferred over rewritten text that leads to errors. In the within-group preference setting, when both rewritten texts can generate the correct answer, the rewritten text with lower PPL and a more standard operational specification is preferred over the rewritten text with high perplexity. This is a function to indicate the correctness of the diagnostic, where 1 indicates correct and 0 indicates incorrect. If it is a perplexity function, then the preference is on... The construction of [the system] must satisfy the following discrimination formula:

[0020]

[0021] in, This represents a constructed training preference pair. This represents a rewrite of the selected preferred fault description. As a positive sample, The description of the rejected non-optimal fault is rewritten. (for negative samples) This indicates the fault diagnosis reasoning results generated by the base large language model based on the corresponding rewritten description. This is a function to indicate the correctness of the diagnostic, where 1 indicates correct and 0 indicates incorrect. A language perplexity function to measure the fluency of rewritten text generation. The formula, which represents a logical AND relationship, indicates that positive samples must have better diagnostic accuracy than negative samples, or have a lower language perplexity value when both are diagnosed correctly.

[0022] Utilize all training preferences to construct the preference dataset Using preference datasets The DPO algorithm is used to optimize the base large language model. The objective function of the optimization process is... Represented as:

[0023]

[0024] in, This represents the DPO objective loss function that needs to be minimized. This represents the base language model that needs optimization. The reference model serves as the constraint benchmark. Indicates from preference dataset The sampled data contains the original problem. Optimal rewriting and non-optimal rewriting The training sample triples. For mathematical expectation operators, It is the Sigmoid activation function. To control the hyperparameters that penalize deviations from the reference model, This represents the probability value of the base large language model generating a specific rewritten description under given conditions;

[0025] By minimizing the objective function Continuously optimize the base large language model This enables the base large language model By learning implicit operational and maintenance cognitive alignment strategies, we can transform fuzzy fault descriptions into fault query statements that better conform to the internal knowledge graph of the inference model, thus obtaining a fault description rewriting model.

[0026] Preferably, the original fault description is reconstructed in a diversified manner based on the fault description rewriting model. Through a PPL-based rejection sampling mechanism, non-standard fault descriptions are transformed into a set of fault query statements, including:

[0027] Non-standard, colloquial original fault descriptions submitted by users or maintenance personnel are input into the fault description rewriting model. The fault description rewriting model uses a random sampling decoding strategy to generate each original fault description by setting temperature parameters and Top-K sampling. ,generate A set of candidate rewrite fault descriptions that are semantically consistent but differ in expression and terminology combination. ;

[0028] A PPL-based rejection sampling mechanism is introduced as a filter to rewrite the fault description for each candidate in the candidate set. Assuming Depend on A sequence of tokens The composition, and the formula for calculating its PPL are as follows:

[0029]

[0030] in, This represents the calculated candidate rewrite fault description. Language confusion score, This represents the total number of tokens contained in the token sequence. This indicates the first token in the token sequence. Each word element, Indicates the generation of the first All historical context lexical sequences preceding each lexical unit, This indicates the generation of the current word element given the preceding context. The conditional probability, For summation operations, For exponentiation;

[0031] Based on each candidate rewrite fault description The corresponding PPL value corresponds to the set of candidate rewrite fault descriptions. Each candidate rewrite fault description Sort in ascending order, keeping the first few rows. The candidate rewrite fault descriptions with the lowest perplexity are selected as the final set of fault query statements. .

[0032] Preferably, the step of performing multi-perspective hybrid consensus diagnostic decision-making based on the real-time collected fault descriptions and the set of fault query statements, and outputting intelligent operation and maintenance fault diagnosis results corresponding to the real-time collected fault descriptions, including root cause localization and repair suggestions, includes:

[0033] The main diagnostic path is generated through direct reasoning based on real-time collected fault descriptions. The real-time collected fault descriptions are compared with the set of fault query statements. Based on the comparison results and standardized fault query statements, deep reasoning is performed to generate auxiliary diagnostic paths. Consensus decision-making is carried out using a majority voting mechanism based on the main diagnostic path and auxiliary diagnostic paths to eliminate the randomness error of a single perspective. The diagnostic conclusion with the highest frequency among all paths is counted, and the intelligent operation and maintenance fault diagnosis result corresponding to the real-time collected fault description is output. The intelligent operation and maintenance fault diagnosis result includes: a determined description of the root cause of the fault and specific fault repair suggestion steps.

[0034] As can be seen from the technical solutions provided by the embodiments of the present invention above, the method of the present invention does not rely on external manual annotation and achieves cognitive alignment by automatically mapping non-standard original fault descriptions to a standardized operation and maintenance terminology space. In the online diagnosis phase, the method of the present invention uses a PPL rejection sampling mechanism to screen high-quality candidate rewrites that conform to operation and maintenance specifications, and combines the auxiliary diagnostic path based on the standardized reconstructed description with the direct diagnostic path based on the original fault description. It utilizes a hybrid majority voting strategy to eliminate comprehension biases from a single perspective, achieving highly robust automated fault location and diagnosis.

[0035] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A flowchart illustrating an intelligent operation and maintenance fault diagnosis method based on problem reconstruction and consensus reasoning, provided in an embodiment of the present invention.

[0038] Figure 2 This is a flowchart illustrating the construction process of a fault description rewriting model based on correctness and DPO proposed in an embodiment of the present invention.

[0039] Figure 3 This is a flowchart of a diversified reconstruction and PPL rejection sampling method proposed in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram illustrating a decision-making process from multi-path generation to final answer voting, as proposed in an embodiment of the present invention. Detailed Implementation

[0041] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0042] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0043] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0044] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0045] This invention proposes an intelligent operation and maintenance (O&M) fault diagnosis method based on problem reconstruction and consensus reasoning. Instead of relying directly on the original fuzzy description for diagnosis, this invention introduces an O&M-specific rewriting mechanism trained using Direct Preference Optimization (DPO). This O&M-specific rewriting mechanism can transform non-standard fault descriptions into high-quality query statements that conform to O&M domain standards and have clear semantics, and eliminate noise through perplexity (PPL)-based quality screening. Finally, this invention combines a multi-perspective consensus reasoning strategy, integrating reasoning paths from the original description perspective and the standardized reconstruction perspective, effectively bridging the cognitive gap between user-speaking language and professional O&M knowledge, and significantly improving the accuracy and robustness of large language models in complex fault diagnosis tasks.

[0046] The intelligent operation and maintenance fault diagnosis method based on problem reconstruction and consensus reasoning proposed in this invention first constructs a fault description rewriting model specifically for the operation and maintenance domain based on the DPO algorithm. By constructing preference pairs of data with dual constraints of "diagnostic accuracy" and "standardization of operation and maintenance terminology", the trained model can automatically transform the vague and colloquial fault descriptions submitted by users into standardized fault query statements that are more conducive to retrieval and reasoning. Then, the trained rewriting model is used to perform diversified restatements of the original fault descriptions, and the rejection sampling mechanism of PPL is introduced to filter out low-quality rewrites that do not conform to the operation and maintenance context, thus constructing a high-quality fault description candidate set with clear semantics and diverse expressions. Finally, a multi-perspective hybrid consensus reasoning strategy is adopted to combine the reasoning path of "original description perspective" with the reasoning path of "reconstructed professional description perspective". Multiple fault diagnosis paths are generated through a large language model and majority voting is performed to obtain the final fault root cause location and repair suggestions.

[0047] This invention is primarily applied to the field of intelligent operations and maintenance (AIOps) technology for cloud computing data centers and enterprise-level IT infrastructure. Specifically, it is applicable to automated fault root cause localization and diagnosis based on natural language descriptions. These problems typically exist in operation and maintenance work order processing, intelligent customer service consultation, or human-machine collaborative operation and maintenance scenarios. It focuses on solving the problem that traditional models cannot accurately retrieve and reason due to the semantic gap between users' non-professional, colloquial fault descriptions and professional operation and maintenance terminology. For example, when a user inputs a non-professional, colloquial fault description such as "the system keeps spinning and is stuck," this invention first reconstructs it into a professional query statement conforming to operation and maintenance standards, such as "system application response timeout (ResponseTimeout)," through a fault description rewriting model. Subsequently, through PPL-based quality screening and multi-perspective hybrid consensus diagnostic decision-making, it outputs accurate fault root cause localization results, such as "due to long transaction table locks in the backend database, the application layer connection pool cannot be released, thus causing the frontend HTTP request to be suspended," along with corresponding repair suggestions.

[0048] The processing flow of the intelligent operation and maintenance fault diagnosis method based on problem reconstruction and consensus reasoning proposed in this embodiment of the invention is as follows: Figure 1 As shown, the processing steps include the following:

[0049] Step S10: Utilize the generation capabilities and self-evaluation feedback of the large language model to construct a fault description rewriting model based on DPO.

[0050] This step is in the offline training phase, and its input data consists of historical maintenance work order data or a set of simulated original fault descriptions. Instead of directly employing the supervised fine-tuning method relying on large-scale manually labeled data as used in existing technologies, this step innovatively proposes a self-reinforcing preference pair construction strategy. Specifically, a base large language model is first constructed, using a pre-trained generative model based on the Transformer architecture. This model has been pre-trained on a massive general-purpose corpus and possesses general natural language understanding, logical reasoning, and text generation capabilities, but alignment optimization has not yet been performed for the specific intelligent maintenance domain. For the specific model selection, open-source large language models with a certain parameter scale, such as the LLaMA series, Qwen series, or Baichuan series, can be used as the initial policy network.

[0051] The base large language model is used to perform multi-path sampling on the input original fault description to generate multiple candidate rewritten fault descriptions. Then, a diagnostic correctness indicator function and a PPL function are introduced as dual evaluation metrics. Rewritten texts that can guide the model to arrive at correct diagnostic conclusions and conform to operational terminology standards (i.e., low PPL values) are marked as positive samples (Chosen), while high PPL texts that lead to incorrect diagnoses or obscure expressions are marked as negative samples (Rejected). A training set containing pairwise preference data is constructed using the positive and negative samples. Finally, the DPO algorithm is used to fine-tune and optimize the base large language model on the training set, outputting a fault description rewriting model capable of mapping colloquial descriptions to standardized operational terminology, which serves as the basic input tool for step S20.

[0052] Step S20: Based on the fault description rewriting model, the original input fault description is reconstructed in a diversified manner, and non-standard fault descriptions are transformed into fault query statements by filtering through a PPL-based rejection sampling mechanism.

[0053] This step is the preprocessing stage of the online diagnosis phase. Non-standard, colloquial original fault descriptions submitted by users or operations and maintenance personnel are input into the fault description rewriting model. The model uses a random sampling decoding strategy instead of traditional greedy decoding to generate a set of candidate fault descriptions that are semantically consistent but differ in expression, thus covering the potential standard terminology space. Subsequently, a PPL (Problem Logic Point) is innovatively introduced as a quality filter. The PPL value of each candidate fault description is calculated under the distribution of the operations and maintenance corpus, and a rejection sampling mechanism is executed to automatically remove low-quality noise in the candidate fault descriptions that is semantically ambiguous or does not conform to operations and maintenance syntax. Finally, a "high-quality standardized fault query statement set" is output, sorted by PPL value and retaining the Top-K optimal solutions. This fault query statement set will be used as enhancement information input into step S30.

[0054] Step S30: Based on the selected set of fault query statements and the original fault description, perform multi-perspective hybrid consensus diagnostic decision-making and output intelligent operation and maintenance fault diagnosis results containing root cause location and repair suggestions.

[0055] This step is the final reasoning and decision-making stage. Its input data includes not only the set of fault query statements output from step S20 and the original fault description, forming a dual input stream of "standardized reconstruction perspective" and "user's original perspective." Existing technologies typically rely on single-path reasoning, which is prone to illusions or biases. This invention innovatively employs a multi-perspective hybrid consensus reasoning strategy: on the one hand, comparative hints guide the large model to perform deep reasoning combined with standardized fault query statements, generating auxiliary diagnostic paths; on the other hand, direct reasoning is performed based on the original fault description to generate the main diagnostic path. The system collects diagnostic conclusions generated from all paths and uses a majority voting mechanism for consensus decision-making, eliminating the randomness error of a single perspective. The final intelligent operation and maintenance fault diagnosis result specifically includes: a determined description of the root cause of the fault (e.g., "long transaction table lock exists in the backend database") and specific fault repair suggestions.

[0056] Phase 1 is the offline training phase, while Phases 2 and 3 are the online diagnostic phases.

[0057] The above step S10 specifically includes: a flowchart of the construction of a fault description rewriting model based on correctness and DPO proposed in this embodiment of the invention, as shown in the figure below. Figure 2 As shown, to address the issue of misconceptions in terminology within the large-scale operation and maintenance model, this model does not rely on manually labeled rewritten data. Instead, it leverages the inherent generative capabilities and evaluation feedback of the large language model itself to construct a self-reinforcing optimization loop. Specifically, this invention first models the "fault description rewriting" task as rewriting the original fault description... Standardized O&M terminology space mapped to optimal model understanding The process. To obtain high-quality training signals, this invention designs a dual preference pair construction strategy based on answer correctness and language fluency. For a given original fault description Using large language models The second sampling generates a set of candidate rewrite fault descriptions. And obtain the fault reasoning and diagnosis results corresponding to each candidate rewrite. .

[0058] Figure 2 The fault description rewriting model shown illustrates the process of generating a preference dataset from the original fault description through sampling, diagnostic verification, and index calculation. This process involves constructing the preference dataset for training. At that time, this invention defined positive samples (Chosen, ) and negative samples (Rejected, The selection logic for this algorithm includes two dimensions: between-group preference and within-group preference. First, in terms of between-group preference, rewrites that guide the model to generate the correct root cause of the failure are preferred over rewrites that lead to errors. Second, in terms of within-group preference, when both rewrites can generate the correct answer, rewrites with lower perplexity (PPL) and more consistent with standard operational specifications are preferred over rewrites with high perplexity. Definition This is a diagnostic correctness indicator function (1 for correct, 0 for incorrect). If it is a perplexity function, then the preference is on... The construction of [the system] must satisfy the following discrimination formula:

[0059]

[0060] in, This represents a constructed training preference pair. This represents the rewriting of the selected preferred fault description (positive sample). The non-optimal fault description was rewritten (negative sample). This indicates the fault diagnosis inference results generated by the large model based on the corresponding rewritten description. This is a diagnostic correctness indicator function (1 for correct, 0 for incorrect). A language perplexity function to measure the fluency of text generation. The formula represents a logical AND relationship, indicating that positive samples must have better diagnostic accuracy than negative samples, or have a lower language perplexity value when both are diagnosed correctly.

[0061] The preference dataset constructed based on the above This invention employs the DPO algorithm to fine-tune the base model. The DPO algorithm eliminates the need to train a separate reward model; instead, it directly optimizes the policy network. To ensure that it maintains consistency with the reference model While avoiding excessive deviation, maximize the selection of optimal samples. Compared to non-preferred samples The likelihood difference. The rewritten model of this invention optimizes the objective function. Represented as:

[0062]

[0063] in, This represents the DPO objective loss function that needs to be minimized. This represents the policy network to be optimized (i.e., the currently trained rewritten model). The representative reference model serves as the constraint benchmark. Indicates from preference dataset The sampled data contains the original problem. Optimal rewriting and non-optimal rewriting The training sample triples. For mathematical expectation operators, It is the Sigmoid activation function. To control the hyperparameters that penalize deviations from the reference model, This represents the probability value that the model generates a specific rewritten description under given conditions.

[0064] By minimizing this loss function, the model is rewritten. It can learn an implicit operational cognition alignment strategy, which automatically transforms the fuzzy description of the input into a high-quality query statement that is more in line with the internal knowledge graph of the reasoning model while keeping the semantics unchanged, thereby providing a more robust input foundation for the subsequent reasoning process.

[0065] Step S20 specifically includes: after training the fault description rewriting model, the present invention utilizes the fault description rewriting model to perform diversified reconstruction and quality screening of the input fault description during the online diagnosis phase. A flowchart of a diversified reconstruction and PPL rejection sampling method proposed in this embodiment of the invention is shown below. Figure 3 As shown, this process describes the complete steps from diversified generation to perplexity-based filtering. To overcome the risk of losing critical fault information that may result from a single rewrite, this invention first employs a sampling-based decoding strategy instead of deterministic greedy decoding. By setting a high temperature parameter and Top-K sampling, the rewrite model trained in the previous stage is utilized. For each original fault description generate A set of candidate rewrite fault descriptions that are semantically consistent but differ significantly in expression and terminology. This diversified generation aims to simulate different perspectives of operations and maintenance experts at different levels on the same problem, thereby covering the optimal understanding space that the model may possess. Although models trained by DPO tend to generate high-quality rewrites, the random sampling process may still introduce some ambiguous or low-quality noise that does not conform to the syntax of operations and maintenance. To address this, this invention introduces a rejection sampling mechanism based on PPL as a filter. Perplexity is an indicator that measures the uncertainty of the probability distribution of the predicted sequence by a language model. A lower PPL value usually means that the generated sentences are more consistent with the distribution habits of professional operations and maintenance corpora and are more logically coherent. For each rewrite problem in the candidate set... Assuming it is made by A sequence of tokens The composition, and its PPL calculation formula are as follows:

[0066]

[0067] in, This represents the calculated candidate rewrite fault description. Language confusion score, This represents the total number of tokens contained in the descriptive text sequence. Represents the first in the sequence Each word element, Indicates the generation of the first All historical context lexical sequences preceding each lexical unit, This indicates the generation of the current word element given the preceding context. The conditional probability, For summation operations, This is for exponential operations.

[0068] Based on the calculated PPL values, the system selects the candidate set. Sort in ascending order and retain the first few rows. The rewrite problem with the lowest perplexity was selected as the final set of fault descriptions. This screening process can be formally represented as:

[0069] in, This represents the final set of high-quality standardized fault descriptions selected. Represents any rewritten fault description from the candidate set. Based on the original fault description The generated set of all candidate rewrite fault descriptions This indicates the ranking of the candidate rewrites after being sorted in ascending order of language perplexity score. This is a preset threshold for the number of retained samples.

[0070] Through the above steps, this invention successfully eliminates potential low-quality rewrites and retains only the clearest and most standardized K problem variants (along with the original problem) into the subsequent consensus inference module, thereby effectively controlling diagnostic noise while ensuring input diversity.

[0071] The above step S30 specifically includes: a decision-making process from multi-path generation to final answer voting proposed in this embodiment of the invention, as follows: Figure 4As shown. After obtaining a high-quality set of reconstructed fault descriptions that have undergone quality screening, this invention enters the final inference and decision-making stage. To balance the intuitiveness of the user's original perspective with the accuracy of the standardized perspective, this invention adopts a multi-perspective hybrid consensus inference strategy. This strategy no longer relies on a single input path, but simultaneously retains the direct diagnostic path based on the original description perspective and the auxiliary diagnostic path based on the reconstructed description perspective. In specific implementation, the system uses the K optimal rewriting problems selected in the previous stage together with the original input problem as inference input. For the rewritten description, the system uses contrastive prompting to guide the large model to comprehensively understand the "user's original repair report description" and the "standardized fault definition," thereby stimulating a deeper level of fault knowledge association; at the same time, the system also independently performs multi-path sampling on the original description to prevent the drift of core fault phenomena due to rewriting bias.

[0072] Finally, the system collects the root causes or repair suggestions generated by all inference paths and uses a majority voting mechanism for consensus decision-making. This mechanism is based on the assumption of consistency and uses the diagnostic conclusion with the highest frequency across all paths as the final output. Through this multi-perspective, multi-path hybrid consensus, this invention effectively offsets the randomness errors and model illusions in single inference, achieving more robust and accurate automated fault diagnosis results than a single perspective.

[0073] In summary, the intelligent operation and maintenance fault diagnosis method based on problem reconstruction and consensus reasoning designed in this invention has been extensively tested on simulated fault datasets and historical operation and maintenance work order datasets. Compared with several baseline methods such as zero-shot, chain-of-thought, and self-consistency, it has achieved significant improvements in evaluation metrics such as fault location accuracy and diagnostic result recall. Furthermore, compared with existing technologies, this invention effectively bridges the gap between users' non-professional natural language descriptions and the model's internal standard operation and maintenance cognition through a unique DPO rewriting and multi-perspective consensus mechanism. Without introducing large-scale external operation and maintenance knowledge graphs or performing heavy manual annotation, it enhances the robustness and stability of the model in complex fault diagnosis tasks and verifies its general effectiveness on models with different parameter scales.

[0074] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0075] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0076] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0077] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent operation and maintenance fault diagnosis method based on problem reconstruction and consensus reasoning, characterized in that, include: By leveraging the generative capabilities and self-evaluation feedback of large language models, a fault description rewriting model based on direct preference optimization of DPO is constructed. Based on the fault description rewriting model, the original fault description is reconstructed in a variety of ways. Through the rejection sampling mechanism based on perplexity degree PPL, the non-standard fault description is transformed into a set of fault query statements. Based on the real-time collected fault descriptions and the set of fault query statements, a multi-perspective hybrid consensus diagnostic decision is made, and an intelligent operation and maintenance fault diagnosis result containing root cause location and repair suggestions is output, corresponding to the real-time collected fault descriptions.

2. The method according to claim 1, characterized in that, The aforementioned method of constructing a fault description rewriting model based on direct preference optimization (DPO) by leveraging the generative capabilities and self-evaluation feedback of large language models includes: A base language model is constructed, and multi-path sampling is performed on the input original fault description using the base language model to generate multiple candidate rewritten fault descriptions. A diagnostic correctness indicator function and a PPL function are introduced as dual evaluation indicators. Rewritten texts that can guide the base language model to draw correct diagnostic conclusions and conform to the operation and maintenance terminology specifications are marked as positive samples, while rewritten texts that cause the base language model to make incorrect diagnoses are marked as negative samples. A training set containing pairwise preference data is constructed using positive and negative samples. Based on the training set, the base language model is fine-tuned and optimized using the DPO algorithm to obtain a fault description rewriting model with the ability to map colloquial descriptions to standardized operation and maintenance terminology.

3. The method according to claim 2, characterized in that, The aforementioned method of constructing a fault description rewriting model based on direct preference optimization (DPO) by leveraging the generative capabilities and self-evaluation feedback of large language models includes: Construct a foundational large language model for a given original fault description. Using a base-based large language model The second sampling generates a set of candidate rewrite fault descriptions. Obtain the fault reasoning and diagnosis results corresponding to each candidate rewritten fault description. The system incorporates filtering logic including between-group and within-group preferences. In the between-group preference setting, rewritten text that guides the model to generate the correct root cause of the failure is preferred over rewritten text that leads to errors. In the within-group preference setting, when both rewritten texts can generate the correct answer, the rewritten text with lower PPL and a more standard operational specification is preferred over the rewritten text with high perplexity. This is a function to indicate the correctness of the diagnostic, where 1 indicates correct and 0 indicates incorrect. If it is a perplexity function, then the preference is on... The construction of [the system] must satisfy the following discrimination formula: in, This represents a constructed training preference pair. This represents a rewrite of the selected preferred fault description. As a positive sample, The description of the rejected non-optimal fault is rewritten. (for negative samples) This indicates the fault diagnosis reasoning results generated by the base large language model based on the corresponding rewritten description. This is a function to indicate the correctness of the diagnostic, where 1 indicates correct and 0 indicates incorrect. A language perplexity function to measure the fluency of rewritten text generation. The formula, which represents a logical AND relationship, indicates that positive samples must have better diagnostic accuracy than negative samples, or have a lower language perplexity value when both are diagnosed correctly. Utilize all training preferences to construct the preference dataset Using preference datasets The DPO algorithm is used to optimize the base large language model. The objective function of the optimization process is... Represented as: in, This represents the DPO objective loss function that needs to be minimized. This represents the base language model that needs optimization. The representative reference model serves as the constraint benchmark. Indicates from preference dataset The sampled data contains the original problem. Optimal rewriting and non-optimal rewriting The training sample triples. For mathematical expectation operators, It is the Sigmoid activation function. To control the hyperparameters that penalize deviations from the reference model, This represents the probability value of the base large language model generating a specific rewritten description under given conditions; By minimizing the objective function Continuously optimize the base large language model This enables the base large language model By learning implicit operational and maintenance cognitive alignment strategies, we can transform fuzzy fault descriptions into fault query statements that better conform to the internal knowledge graph of the inference model, thus obtaining a fault description rewriting model.

4. The method according to claim 2 or 3, characterized in that, The aforementioned fault description rewriting model performs diversified reconstruction of the original fault description. Through a PPL-based rejection sampling mechanism, non-standard fault descriptions are transformed into a set of fault query statements, including: Non-standard, colloquial original fault descriptions submitted by users or maintenance personnel are input into the fault description rewriting model. The fault description rewriting model uses a random sampling decoding strategy to generate each original fault description by setting temperature parameters and Top-K sampling. ,generate A set of candidate rewrite fault descriptions that are semantically consistent but differ in expression and terminology combination. ; A PPL-based rejection sampling mechanism is introduced as a filter to rewrite the fault description for each candidate in the candidate set. Assuming Depend on A sequence of lexical tokens The composition, and the formula for calculating its PPL are as follows: in, This represents the calculated candidate rewrite fault description. Language confusion score, This represents the total number of tokens contained in the token sequence. This indicates the first token in the token sequence. Each word element, Indicates the generation of the first All historical context lexical sequences preceding each lexical unit, This indicates the generation of the current word element given the preceding context. The conditional probability, For summation operations, For exponentiation; Based on each candidate rewrite fault description The corresponding PPL value corresponds to the set of candidate rewrite fault descriptions. Each candidate rewrite fault description Sort in ascending order, keeping the first few rows. The candidate rewrite fault descriptions with the lowest perplexity are selected as the final set of fault query statements. .

5. The method according to claim 4, characterized in that, The process of performing multi-perspective hybrid consensus diagnostic decision-making based on real-time collected fault descriptions and the set of fault query statements, and outputting intelligent operation and maintenance fault diagnosis results corresponding to the real-time collected fault descriptions, including root cause localization and remediation suggestions, includes: The main diagnostic path is generated through direct reasoning based on real-time collected fault descriptions. The real-time collected fault descriptions are compared with the set of fault query statements. Based on the comparison results and standardized fault query statements, deep reasoning is performed to generate auxiliary diagnostic paths. Consensus decision-making is carried out using a majority voting mechanism based on the main diagnostic path and auxiliary diagnostic paths to eliminate the randomness error of a single perspective. The diagnostic conclusion with the highest frequency among all paths is counted, and the intelligent operation and maintenance fault diagnosis result corresponding to the real-time collected fault description is output. The intelligent operation and maintenance fault diagnosis result includes: a determined description of the root cause of the fault and specific fault repair suggestion steps.

Images