A large model industrial knowledge retrieval enhancement method, system and storage medium

Abstract

The application relates to the technical field of intelligent operation and maintenance of industrial equipment, and in particular to a large model industrial knowledge retrieval enhancement method and system and a storage medium. Text blocks of an industrial knowledge text set are mapped into text feature vectors, and are stored in each node of a fault tree according to semantic matching relationships to construct a topological vector database; real-time sensor data of equipment are collected, and a deviation matrix is calculated to lock an initial target node in the fault tree; after a user query request is vectorized, similarity retrieval is performed in a local vector subset of the target node, and when the retrieval score does not satisfy a confidence threshold, the retrieval range is expanded back to the parent node along the fault tree; and text blocks satisfying the condition are jointly input into a large language model together with the query request to generate a diagnostic reply. The application can solve or at least alleviate the problems of false recall and large model diagnosis illusion caused by semantic overlap of industrial vocabularies in existing global vector retrieval, and provides a large model industrial knowledge retrieval enhancement method, system and storage medium.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for industrial equipment, and in particular to a method, system, and storage medium for enhancing large-scale industrial knowledge retrieval. Background Technology

[0002] As large language models are increasingly applied in question answering and text generation, industrial manufacturing and equipment maintenance scenarios place higher demands on their professional capabilities. A common approach in existing technologies is to employ a retrieval-enhanced generative architecture. This involves segmenting enterprise-owned fault knowledge bases, equipment operation manuals, and other textual materials into text blocks and mapping them to high-dimensional vectors to construct a vector database. When a user submits an equipment fault query, relevant text segments are retrieved using vector similarity calculations and provided as context to the large language model, assisting it in generating professionally grounded responses. This approach, to some extent, compensates for the knowledge gaps of large language models in specific industrial fields.

[0003] However, industrial equipment systems contain a large number of semantically similar technical terms and fault descriptions with different physical affiliations. For example, "pressure valve leakage" has corresponding technical documents in hydraulic systems, cooling systems, and lubrication systems, while "abnormal bearing temperature" can appear in multiple components such as spindles, gearboxes, and generators. Existing global vector retrieval mechanisms are prone to falsely recalling semantically similar text segments from different physical systems when faced with such highly overlapping industrial terms. To reduce this false recall rate, some improvement schemes attempt to constrain the process by optimizing text segmentation granularity, introducing keyword pre-filtering, or setting metadata tags. However, these methods are still limited to textual semantic processing and fail to establish a correlation between textual knowledge and the actual physical operating state of the equipment.

[0004] The root cause of these shortcomings lies in the fact that existing retrieval mechanisms use the semantic vector space as the sole retrieval basis, while equipment fault diagnosis is inherently limited by the physical structure and real-time operating status of the equipment. When the retrieval system lacks the ability to perceive the current operating parameters of the equipment, it cannot determine which textual knowledge is physically relevant to the current fault scenario, and can only perform an indiscriminate nearest neighbor search in the global vector space. As the size of the knowledge base increases and the complexity of the equipment system rises, the probability of semantic ambiguity and false recall also increases. Diagnostic suggestions generated by large language models based on misaligned contexts will lack practical reference value and may even lead to operational risks in safety-sensitive industrial scenarios.

[0005] To address the aforementioned problems, this invention proposes incorporating real-time physical sensing data from equipment into the text retrieval stage of a retrieval-enhanced generative architecture. By utilizing sensor data to calculate equipment operational deviation and mapping it to a pre-constructed fault tree topology, a local retrieval space related to the current operating condition is locked at the physical level before semantic retrieval. Furthermore, a topology backtracking mechanism based on the fault tree hierarchy is designed to address situations where local retrieval coverage is insufficient. This achieves, at the engineering level, the constraint and guidance of physical sensing information on the semantic retrieval scope. Summary of the Invention

[0006] To achieve the above-mentioned objectives, this invention provides a method, system, and storage medium for enhancing large-scale industrial knowledge retrieval, aiming to solve or at least mitigate the problems of false recall and diagnostic illusion in large-scale models caused by semantic overlap of industrial terms in existing global vector retrieval.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for enhancing large-scale industrial knowledge retrieval, comprising: Obtain the fault tree topology model of industrial equipment; The text blocks of the industrial knowledge text set are mapped to text feature vectors, and stored in each fault node of the fault tree according to semantic matching relationship to construct a topology vector database. Real-time sensor data of the industrial equipment is collected and the deviation is calculated. Based on the deviation, the initial target node in the fault tree is located. Receive natural language query requests and convert them into query feature vectors, and calculate a first similarity score within a subset of the text feature vectors of the initial target node and its associated child nodes; When the maximum value of the first similarity score is lower than the preset confidence threshold, backtrack along the fault tree to the parent node topology to expand the retrieval space and calculate the second similarity score; Extract text blocks that meet the confidence threshold and input them together with the query request into a large language model to generate diagnostic response text.

[0008] To further realize the present invention, the following technical solutions may be preferred: Preferably, the calculation of deviation includes: The real-time sensor data is denoised and standardized to obtain a preprocessed data stream; Extract the frequency domain and time domain features of the preprocessed data stream; Calculate the statistical distance between the frequency domain features and the time domain features and the preset normal operating condition benchmark data, and use the statistical distance as an element of the deviation matrix.

[0009] Preferably, the step of mapping text blocks of the industrial knowledge text set to text feature vectors and storing them in each fault node of the fault tree according to semantic matching relationships includes: The industrial knowledge text set is divided into multiple text blocks, and each text block is mapped to a text feature vector; Extract device identifier entities and fault mode entities from the text block; Calculate the semantic matching degree between the device identifier entity words and the fault mode entity words and the labels of each node in the fault tree topology model; The text feature vector is stored in the storage partition to which the target node belongs, whose semantic matching degree meets the preset association threshold.

[0010] Preferably, the step of locking the initial target node in the fault tree based on the deviation includes: Set a corresponding deviation activation threshold for each of the aforementioned fault nodes; When the value of any physical measurement point in the deviation matrix exceeds the deviation activation threshold, the underlying fault node bound to that physical measurement point is activated. The underlying fault node with the highest deviation is identified as the initial target node.

[0011] Preferably, the step of backtracking the fault tree to the parent node topology to expand the retrieval space and calculating the second similarity score includes: Extract the direct parent node identifier of the initial target node in the fault tree topology model; Obtain the subset of expanded text feature vectors associated with the direct parent node identifier; The expanded text feature vector subset is merged with the text feature vector subset corresponding to the initial target node, and the second similarity score of the query feature vector is calculated in the merged vector space.

[0012] Preferably, when calculating the second similarity score, the vectors in the expanded text feature vector subset are assigned a distance weight coefficient lower than that of the vectors in the original text feature vector subset.

[0013] Preferably, after generating the diagnostic response text, the method further includes: The query request, the diagnostic response text, and the finally retrieved text block are associated, recorded, and stored to generate a question-answer pair log with traceability tags.

[0014] A large-scale industrial knowledge retrieval enhancement system includes: The data construction module is configured to acquire the fault tree topology model of industrial equipment, map text blocks in the industrial knowledge text set into text feature vectors, and establish an association between the text feature vectors and fault nodes to construct a topology vector database. The real-time sensing module is configured to collect real-time sensor data from the industrial equipment, calculate the deviation matrix between the real-time sensor data and preset normal operating condition benchmark data, and lock the initial target node based on the deviation matrix. The dynamic retrieval module is configured to receive natural language query requests for the industrial equipment and convert them into query feature vectors, calculate a first similarity score within a subset of the text feature vectors of the initial target node and its associated child nodes, and perform topological backtracking to the parent node to calculate a second similarity score when the maximum value of the first similarity score is lower than a confidence threshold. The large model interaction module is configured to extract text blocks that meet the confidence threshold as contextual reference information, and input the contextual reference information and the query request into the large language model to generate diagnostic response text.

[0015] Preferably, the dynamic retrieval module is further configured to extract the direct parent node identifier of the initial target node when performing topology backtracking, obtain the expanded text feature vector subset associated with the direct parent node identifier and perform retrieval by combining it with the original text feature vector subset, and at the same time assign a distance weight coefficient lower than that of the vectors in the original text feature vector subset to the vectors in the expanded text feature vector subset.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described large-scale industrial knowledge retrieval enhancement method.

[0017] The beneficial effects of this invention are: The large-scale industrial knowledge retrieval enhancement method, system, and storage medium provided by this invention constructs a vector database partitioned storage structure corresponding to the fault tree topology hierarchy, and uses the deviation calculation results of real-time sensor data to lock the initial retrieval space in the fault tree. This limits the semantic retrieval scope to a local vector subset associated with the current operating condition of the equipment. This approach eliminates interference from semantically similar texts between different physical systems, reduces the probability of large language models generating illusory content due to contextual misalignment when generating diagnostic responses, and improves the accuracy and practical usability of industrial fault diagnosis responses.

[0018] Furthermore, the topology backtracking mechanism of this invention, when the initial local retrieval confidence is insufficient, expands the retrieval range in an orderly manner according to the parent-child hierarchical relationship of the fault tree, and distinguishes and sorts textual knowledge from different levels using distance weighting coefficients. This mechanism enables the system to gradually locate relevant knowledge text even in situations such as sensor pointing deviation or cross-system coupling failures, avoiding regression to a global search due to a single retrieval failure, and ensuring the retrieval stability and output reliability of the system under complex industrial conditions. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system architecture of the present invention.

[0020] Figure 2 This is a flowchart of the online phase of the method of the present invention.

[0021] Figure 3 This is a schematic diagram of the topology backtracking principle of the present invention.

[0022] Figure 4 This is a retrieval response time distribution chart. Detailed Implementation

[0023] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] 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 embodiments of the present invention, and not all embodiments. 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. Example 1

[0025] This embodiment provides a large-scale industrial knowledge retrieval enhancement system, such as... Figure 1 As shown, the system is deployed in a hybrid architecture between edge computing nodes and cloud servers in an industrial setting, and consists of a data construction module, a real-time sensing module, a dynamic retrieval module, and a large model interaction module. Figure 1 As shown on the left, the offline data construction end is responsible for processing multi-source heterogeneous industrial knowledge through text segmentation, vectorization, and entity word semantic matching, and then storing it in the topology vector database according to fault tree nodes. For example... Figure 1As shown on the right, the online retrieval generator receives real-time sensor data and user query requests. After deviation calculation, target node locking, local similarity retrieval, and topological backtracking when necessary, the large language model generates diagnostic response text.

[0026] The data construction module operates offline, with its input source being the multi-source heterogeneous knowledge assets accumulated by industrial enterprises. In the CNC machine tool production line involved in this embodiment, the knowledge sources mainly include equipment operation manuals, maintenance history records, failure mode and consequence analysis reports, and compilations of typical failure cases. The data construction module first obtains the fault tree topology model of the industrial equipment. This fault tree topology model is compiled by reliability engineers based on historical failure statistics and equipment structural relationships, and has a tree-like hierarchical structure. Taking a CNC machine tool as an example, the root node of the fault tree represents the entire machine tool, the second-level nodes represent subsystems such as the spindle system, servo feed system, cooling and lubrication system, and electrical control system, and the third-level nodes are further subdivided into specific components, such as spindle bearings, spindle motor encoders, proportional servo valves, and cooling pumps.

[0027] The data construction module then processes the industrial knowledge text set. The system segments the text set into multiple text blocks according to semantic integrity, taking into account paragraph boundaries and contextual coherence to ensure that a single text block is not truncated when describing a complete fault phenomenon or maintenance operation. After segmentation, the system uses a pre-trained industrial domain word embedding model to map each text block into a text feature vector. This word embedding model has been fine-tuned for domain adaptation on technical documents and maintenance work orders in the industrial maintenance field. Specifically, based on a general pre-trained language model, it uses comparative learning and fine-tuning with equipment manual text pairs and maintenance work order text pairs within the industrial domain, enabling its embedding space to effectively distinguish semantically similar but different system-related expressions in industrial scenarios. For example, it maps "hydraulic system pressure valve" and "cooling system pressure valve" to vector positions with sufficient discriminative power.

[0028] Establishing the association between text feature vectors and fault nodes is the core step in the data construction module. The system extracts equipment identifier entities (such as "spindle bearing" and "cooling pump") and fault mode entities (such as "abnormal vibration" and "overheating") from text blocks using named entity recognition. It then calculates the semantic matching degree between these entities and the labels of each node in the fault tree topology model. Text feature vectors whose semantic matching degree meets a preset association threshold are stored in the independent storage partition corresponding to the target node. Each fault tree node has an independent vector storage space, corresponding to an index partition in the vector database. After the entire process is completed, the topology vector database achieves alignment between the knowledge text semantic space and the equipment's physical topology.

[0029] The real-time sensing module continuously operates online. This module collects real-time data from various sensor channels during the operation of industrial equipment via a data acquisition interface, including signals of various physical quantities such as temperature, pressure, displacement, and vibration acceleration. The raw data collected contains noise and dimensional differences. The system performs denoising and standardization on the real-time sensor data to obtain a preprocessed data stream. Based on this, the system extracts the time-domain and frequency-domain features of the preprocessed data stream. Time-domain features include root mean square value, peak factor, and kurtosis, while frequency-domain features include dominant frequency distribution and energy density of each frequency band. The system compares the feature vector of the current acquisition cycle with the pre-existing normal operating condition benchmark data in the system, using multivariate statistical distance for measurement. Multivariate statistical distance is chosen because there is often a covariance relationship between the signals of various sensor channels in industrial equipment. Single-variable distance cannot eliminate interference caused by linear correlation between variables. Multivariate statistical distance, by introducing the inverse of the covariance matrix to perform weighted anti-correlation processing on each dimension, can more accurately measure the deviation of the current operating state of the equipment from the normal operating condition. The calculation results are filled into a deviation matrix. The rows of this matrix correspond to each sensor channel, and the columns correspond to different statistical feature dimensions. The matrix elements reflect the degree of deviation of the corresponding channel from the normal operating condition in the corresponding feature dimension.

[0030] The real-time perception module further determines the initial target node based on the deviation matrix. The system pre-sets a corresponding deviation activation threshold for each leaf node in the fault tree, determined based on the statistical distribution of historical fault data. When the value of a specific physical measurement point in the deviation matrix exceeds the activation threshold of the leaf node bound to that measurement point, the leaf node is marked as active. If multiple leaf nodes are activated simultaneously, the system selects the lowest-level fault node with the highest deviation as the initial target node. The determination of this node signifies that the system has completed the initial physical localization of potential fault areas.

[0031] The dynamic retrieval module starts working after receiving a natural language query request for the industrial equipment input by the maintenance personnel. The system converts the query request into a query feature vector using the same word embedding model as the text feature vector, ensuring that the query vector and the knowledge text vector are in the same semantic space. The dynamic retrieval module performs similarity calculations on the subset of text feature vectors corresponding to the initial target node and its directly associated child nodes to obtain a first similarity score.

[0032] The confidence level of the first similarity score is controlled by a preset confidence threshold. When the maximum value of the first similarity score reaches or exceeds the confidence threshold, it indicates that highly relevant knowledge text has been retrieved in the current local vector space, and the system directly extracts the corresponding text block. When the maximum value of the first similarity score is lower than the confidence threshold, it indicates that the knowledge coverage of the current local vector space is insufficient to meet the user's query needs, and there may be cross-system coupling faults or sensor pointing deviations. At this time, the system initiates a topology backtracking mechanism.

[0033] like Figure 3 As shown, the execution process of topology backtracking is as follows: The system extracts the initial target node ( Figure 3 The identifier of the direct parent node of node A in the fault tree topology model (labeled as node A). Figure 3 The system retrieves an expanded subset of text feature vectors associated with its parent node (labeled as node B) from the topological vector database. The system merges the original vector subset of node A with the expanded vector subset associated with node B to form an expanded retrieval space. The second similarity score of the query feature vectors is recalculated in the merged vector space. When calculating the second similarity score, vectors in the expanded text feature vector subset are assigned a lower distance weight coefficient than those in the original text feature vector subset, ensuring that locally relevant knowledge that is physically closer to the initial fault source maintains priority in the ranking. If the second similarity score still does not meet the confidence threshold, the system continues to backtrack to the next higher-level parent node, repeating the above process until the confidence threshold is met or the root node of the fault tree is reached.

[0034] The large-scale model interaction module is responsible for generating the final diagnostic content. The system extracts top-ranking text blocks that meet the confidence threshold from the search results as contextual reference information, and assembles them with the maintenance personnel's original query request according to a preset prompt word template. The assembled input is fed into the large language model, which performs logical reasoning and language organization within a defined context, generating diagnostic response text and returning it to the user. After generating the diagnostic response text, the system associates and stores the query request, the diagnostic response text, and the text blocks used in the final retrieval, generating a question-answer pair log with traceability tags for subsequent quality audits and knowledge base iterations.

[0035] Example 2

[0036] Based on the system in Embodiment 1, this embodiment provides a method for enhancing large-scale industrial knowledge retrieval. This embodiment uses the scenario of an abnormality in the servo feed system of a CNC machine tool as an example for illustration.

[0037] During the offline phase, the system retrieves the fault tree topology model file of the CNC machine tool from the enterprise knowledge management platform. This file is stored in a tree-like data structure, with the root node labeled "CNC-Machine-001". Its subordinate subsystem nodes include "SpindleSystem", "ServoFeedSystem", "CoolingSystem", "ElectricalControl", etc. Each subsystem node continues to expand downwards to the component-level leaf nodes, such as "ServoFeedSystem" which contains nodes such as "BallScrew", "LinearGuide", "ServoMotor", "GratingScale", etc.

[0038] The system synchronously acquires an industrial knowledge text set, which originates from operation and maintenance manuals provided by equipment manufacturers, internal maintenance work order records, and industry standard documents. The system segments the text set according to semantic paragraphs. For example, a complete description of "a decrease in ball screw preload leading to an increase in backlash in the feed direction" is retained as a separate text block, without being truncated between "decreased preload" and "increased backlash." Each segmented text block is mapped into a text feature vector using an industrial domain word embedding model.

[0039] During the association establishment process, the system performs named entity recognition on each text block, extracting device identifier entities and fault mode entities. Taking a text block containing the description "ball screw preload reduction" as an example, the system identifies the device identifier entity "ball screw" and the fault mode entity "preload reduction". The system then calculates the semantic matching degree between "ball screw" and all node labels in the fault tree. The "BallScrew" node achieves the highest matching degree and exceeds the preset association threshold; therefore, the feature vector of this text block is stored in the independent storage partition of the "BallScrew" node. Similarly, text blocks involving "insufficient guide rail lubrication" and "servo motor overheating" are associated with corresponding nodes such as "LinearGuide" and "ServoMotor", respectively.

[0040] Entering the online operation phase, such as Figure 2As shown, the sensor channels of the CNC machine tool continuously output real-time data. Assume that at a certain moment, the real-time sensing module detects abnormal fluctuations in the data from the vibration acceleration and displacement sensors in the Y-axis direction of the servo feed system. The system performs noise reduction processing on the raw data to eliminate power frequency interference and random impulse noise, and then performs standardization. The system extracts the time-domain features (the root mean square value of the vibration signal is higher than normal) and frequency-domain features (a new energy peak appears at a specific frequency) within the current acquisition window, and performs multivariate statistical distance calculation between the vector composed of these features and the feature vector of the normal operating condition reference data. This distance metric, through the weighting effect of the inverse covariance matrix, eliminates interference from linear correlations between sensor channels. The calculation results show that the element values ​​corresponding to the Y-axis vibration channel and the Y-axis displacement channel in the deviation matrix exceed the preset activation threshold.

[0041] The system checks the faulty leaf nodes bound to the two physical measuring points. The Y-axis vibration sensor is bound to the "BallScrew" node, and the Y-axis displacement sensor is also bound to the "BallScrew" node. If both measuring points point to the same leaf node, that node is directly identified as the initial target node. If the two measuring points point to different leaf nodes, the system selects the one with the highest deviation.

[0042] At this point, the maintenance personnel input the query via the terminal: "Y-axis feed positioning accuracy has decreased, backlash is excessive, possible causes and solutions?" The dynamic retrieval module converts this query text into a query feature vector. The system performs similarity calculations on the text feature vector subsets corresponding to the "BallScrew" node and its child nodes (if any). The "BallScrew" node storage partition contains text feature vectors related to lead screw preload adjustment, lead screw wear judgment, and backlash compensation methods. The query feature vector and the text block feature vector describing "backlash adjustment methods" achieved a high similarity score, with the maximum value of the first similarity score exceeding the confidence threshold. The system directly extracts this text block and its adjacent top-ranking text blocks.

[0043] As an illustration of the topology backtracking scenario (e.g.) Figure 3 As shown), assuming the maintenance personnel input a query of "abnormal noise and temperature rise during Y-axis movement," the query describes a fault that could potentially involve both the ball screw body and the guide rail lubrication components. In the "BallScrew" node ( Figure 3 In the local vector space of node A, since this subset mainly contains text related to ball screws and not knowledge related to guide rail lubrication, the maximum value of the first similarity score failed to reach the confidence threshold. The system initiates topology backtracking to extract the direct parent node "ServoFeedSystem" of the "BallScrew" node. Figure 3The system retrieves the expanded text feature vector subset from all child nodes (including "BallScrew", "LinearGuide", "ServoMotor", and "GratingScale") under the "ServoFeedSystem" node, using the identifier of node B. The system then merges this expanded subset with the original "BallScrew" subset (e.g., ...). Figure 3 The merged retrieval space (highlighted by curly braces) is shown in the merged vector space. Similarity calculation is performed in the merged vector space. At this point, the text feature vector from the "LinearGuide" node regarding "insufficient guide rail lubrication leading to frictional heating and abnormal noise" achieves a high match with the query vector. The system assigns a lower distance weight coefficient to vectors from the expanded subset during calculation, ensuring that relevant content in the original "BallScrew" subset maintains a certain priority in the ranking. After the merged retrieval, the maximum value of the second similarity score meets the confidence threshold, and the system extracts the top-ranked text blocks.

[0044] According to statistics from the verification process, about 92% of the queries that triggered topology backtracking met the confidence threshold after one backtracking (i.e., extending to the direct parent node), while only about 8% of the queries required two or more backtrackings.

[0045] The large model interaction module concatenates the extracted text blocks with the original query request and inputs them into the large language model. Based on contextual reference information and integrating knowledge of both the lead screw and guide rail, the large language model generates a diagnostic response text containing possible cause analysis and step-by-step handling suggestions, which is then returned to the maintenance personnel. After generation, the system associates and stores the query request, the generated diagnostic response text, and the retrieved text blocks as a question-and-answer pair log with traceability tags.

[0046] Example 3

[0047] This embodiment provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the large-scale industrial knowledge retrieval enhancement method in Embodiment 2.

[0048] The technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0049] Example 4

[0050] To verify the actual effectiveness of the above technical solution, a month-long comparative test was conducted on a CNC machine tool production line. The test knowledge base contained approximately 120,000 maintenance records and 300 equipment manuals. The topology vector database covered three subsystems: the spindle system, the servo feed system, and the cooling and lubrication system, and a total of 45 leaf nodes corresponding to sensor channels were mounted.

[0051] The comparative scheme adopts the general RAG architecture of traditional global vector retrieval, which stores all text feature vectors in a single flat vector database and performs nearest neighbor retrieval across the entire database during a query. The present invention's scheme, as an embodiment, is implemented according to the above system and method.

[0052] The test constructed 500 sets of simulated fault queries, of which 350 sets were single-system fault queries and 150 sets were cross-system compound fault or vaguely described queries. Evaluation metrics included Top-3 knowledge recall rate (the proportion of the first three text blocks returned by the retrieval that contain the correct answer) and diagnostic illusion rate (the proportion of diagnostic responses generated by the large model that contain misleading content unrelated to the actual fault).

[0053] In 350 single-fault query scenarios, the comparative solution achieved a Top-3 recall rate of 76.5% and a hallucination rate of 24.2%. Analysis of the erroneous cases revealed that the comparative solution incorrectly recalled text blocks related to "cooling system pressure valve leakage" to queries for "hydraulic servo system pressure anomaly." This was because the semantic vectors of words like "pressure" and "leakage" in the two text types were close, making global retrieval unable to distinguish their physical affiliation. The proposed solution, after locking the initial target node using sensor deviation, performs the retrieval only within the corresponding local space, eliminating interference vectors from the cooling system. This resulted in a Top-3 recall rate of 94.1% and a hallucination rate reduced to 3.8%. In single-fault scenarios, the proposed solution improved the recall rate by 17.6 percentage points and reduced the hallucination rate by 20.4 percentage points compared to the comparative solution.

[0054] In 150 scenarios involving complex faults and fuzzy queries, the Top-3 recall rate of the comparative solution decreased to 52.3%, while the phantom rate increased to 38.6%. The proposed solution, when the initial local retrieval at a node fails to reach the confidence threshold, expands to the parent node range through topology backtracking, maintaining a Top-3 recall rate of 86.7% and a phantom rate of 7.2% despite an additional retrieval latency of approximately 75ms. Even in complex fault scenarios, the recall rate of the proposed solution is still higher than that of the comparative solution in single fault scenarios. In queries triggering topology backtracking, 92% of searches were completed after one backtracking, while 8% required a second backtracking.

[0055] like Figure 4As shown, in the concurrent query stress test, when the system processes 20 query requests simultaneously, the average response time of the retrieval phase (including deviation calculation, local retrieval, and possible topology backtracking) is 142ms. Specifically, deviation matrix calculation takes approximately 18ms, local vector retrieval takes approximately 35ms, topology backtracking (when triggered) adds an average of approximately 75ms, and large language model generation takes approximately 890ms. The total time spent in the retrieval phase is significantly lower than the time spent generating the large language model. The latency increase from topology backtracking accounts for a small percentage of the overall end-to-end response and does not affect the system's real-time requirements in industrial settings.

[0056] During a month-long deployment and operation, the system processed approximately 3,200 real query requests, with about 22% triggering the topology backtracking mechanism. Based on feedback from maintenance personnel, the usability rate of diagnostic response texts (the percentage of responses deemed valuable for troubleshooting by maintenance personnel) was 91.3%, a significant improvement compared to 67.8% when deploying the traditional RAG solution. The question-and-answer log traceability function helped identify and supplement 47 fault handling solutions missing from the original knowledge base.

[0057] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for enhancing industrial knowledge retrieval in a large-scale model, characterized in that, include: Obtain the fault tree topology model of industrial equipment; The text blocks of the industrial knowledge text set are mapped to text feature vectors, and stored in each fault node of the fault tree according to semantic matching relationship to construct a topology vector database. Real-time sensor data of the industrial equipment is collected and the deviation is calculated. Based on the deviation, the initial target node in the fault tree is located. Receive natural language query requests and convert them into query feature vectors, and calculate a first similarity score within a subset of the text feature vectors of the initial target node and its associated child nodes; When the maximum value of the first similarity score is lower than the preset confidence threshold, backtrack along the fault tree to the parent node topology to expand the retrieval space and calculate the second similarity score; Extract text blocks that meet the confidence threshold and input them together with the query request into a large language model to generate diagnostic response text.

2. The method according to claim 1, characterized in that, The calculation of deviation includes: The real-time sensor data is denoised and standardized to obtain a preprocessed data stream; Extract the frequency domain and time domain features of the preprocessed data stream; Calculate the statistical distance between the frequency domain features and the time domain features and the preset normal operating condition benchmark data, and use the statistical distance as an element of the deviation matrix.

3. The method according to claim 1, characterized in that, The process of mapping text blocks of the industrial knowledge text set to text feature vectors and storing them in each fault node of the fault tree according to semantic matching relationships includes: The industrial knowledge text set is divided into multiple text blocks, and each text block is mapped to a text feature vector; Extract device identifier entities and fault mode entities from the text block; Calculate the semantic matching degree between the device identifier entity words and the fault mode entity words and the labels of each node in the fault tree topology model; The text feature vector is stored in the storage partition to which the target node belongs, whose semantic matching degree meets the preset association threshold.

4. The method according to claim 1, characterized in that, The step of locking the initial target node in the fault tree based on the deviation includes: Set a corresponding deviation activation threshold for each of the aforementioned fault nodes; When the value of any physical measurement point in the deviation matrix exceeds the deviation activation threshold, the underlying fault node bound to that physical measurement point is activated. The underlying fault node with the highest deviation is identified as the initial target node.

5. The method according to claim 1, characterized in that, The process of backtracking the fault tree to the parent node topology to expand the retrieval space and calculating the second similarity score includes: Extract the direct parent node identifier of the initial target node in the fault tree topology model; Obtain the subset of expanded text feature vectors associated with the direct parent node identifier; The expanded text feature vector subset is merged with the text feature vector subset corresponding to the initial target node, and the second similarity score of the query feature vector is calculated in the merged vector space.

6. The method according to claim 5, characterized in that, When calculating the second similarity score, the vectors in the expanded text feature vector subset are assigned a distance weight coefficient that is lower than that of the vectors in the original text feature vector subset.

7. The method according to claim 1, characterized in that, After generating the diagnostic response text, the following is also included: The query request, the diagnostic response text, and the finally retrieved text block are associated, recorded, and stored to generate a question-answer pair log with traceability tags.

8. A large-scale industrial knowledge retrieval enhancement system, characterized in that, include: The data construction module is configured to acquire the fault tree topology model of industrial equipment, map text blocks in the industrial knowledge text set into text feature vectors, and establish an association between the text feature vectors and fault nodes to construct a topology vector database. The real-time sensing module is configured to collect real-time sensor data from the industrial equipment, calculate the deviation matrix between the real-time sensor data and preset normal operating condition benchmark data, and lock the initial target node based on the deviation matrix. The dynamic retrieval module is configured to receive natural language query requests for the industrial equipment and convert them into query feature vectors, calculate a first similarity score within a subset of the text feature vectors of the initial target node and its associated child nodes, and perform topological backtracking to the parent node to calculate a second similarity score when the maximum value of the first similarity score is lower than a confidence threshold. The large model interaction module is configured to extract text blocks that meet the confidence threshold as contextual reference information, and input the contextual reference information and the query request into the large language model to generate diagnostic response text.

9. The system according to claim 8, characterized in that, The dynamic retrieval module is further configured to extract the direct parent node identifier of the initial target node when performing topology backtracking, obtain the expanded text feature vector subset associated with the direct parent node identifier and perform retrieval by combining it with the original text feature vector subset, and at the same time assign a distance weight coefficient lower than that of the vectors in the original text feature vector subset to the vectors in the expanded text feature vector subset.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the large-scale industrial knowledge retrieval enhancement method according to any one of claims 1 to 7.

Images