Data retrieval method and apparatus, electronic device, and storage medium
By combining multi-dimensional retrieval and hybrid database retrieval, the problem of poor adaptability in existing operation and maintenance knowledge retrieval technologies has been solved, achieving efficient and comprehensive operation and maintenance knowledge retrieval, forming a complete operation and maintenance handling chain, and improving the accuracy and efficiency of retrieval results.
Patent Information
- Application Number
- CN202611120549.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-27
- Publication Date
- 2026-11-10
AI Technical Summary
Existing knowledge retrieval technologies for operations and maintenance cannot simultaneously address both precise constraint matching and fuzzy semantic matching, resulting in insufficient completeness of retrieval results. These technologies suffer from poor adaptability, limited filtering dimensions, and an inability to achieve multi-source knowledge linkage retrieval.
This study employs multi-dimensional retrieval, hybrid database retrieval, relevance ranking, and relationship expansion retrieval methods. It utilizes text semantic extraction, structured attribute extraction, and data expansion extraction, combined with hybrid database matching and relationship expansion, to integrate knowledge sets for comprehensive analysis and supplementation.
It improves the accuracy and response efficiency of operation and maintenance knowledge retrieval, enhances the completeness and reliability of results, forms a complete knowledge chain for operation and maintenance handling, reduces misunderstandings, and improves the work efficiency of operation and maintenance personnel.
Smart Images

Figure CN123068529A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, and in particular to a data retrieval method, apparatus, electronic device, and storage medium. Background Technology
[0002] In equipment operation and maintenance knowledge retrieval scenarios, user input often exhibits characteristics such as non-standard expressions, incomplete constraints, unstable semantic expression, and missing contextual information, placing high demands on the comprehensive adaptability of the retrieval system. Current existing operation and maintenance knowledge retrieval technologies have a single retrieval mode, failing to simultaneously handle precise constraint matching and fuzzy semantic matching, resulting in insufficient completeness of retrieval results and significant technical defects. Specifically, single precise matching retrieval can only achieve hard matching of explicit constraints such as keywords, equipment numbers, error codes, work order numbers, and time ranges, failing to identify non-standard synonymous expressions and implicit fault phenomena, and struggling to recall cross-document related knowledge, leading to missed effective knowledge. Single semantic retrieval, on the other hand, can only achieve text semantic similarity matching, easily ignoring key explicit constraints in operation and maintenance scenarios, resulting in redundant and inaccurate recall results. In summary, existing retrieval technologies suffer from poor adaptability, a single filtering dimension, and an inability to achieve multi-source knowledge linkage recall, ultimately leading to incomplete retrieval results. Summary of the Invention
[0003] In view of this, the purpose of this application is to provide a data retrieval method, apparatus, electronic device and storage medium to overcome all or part of the shortcomings of the prior art.
[0004] To achieve the above objectives, this application provides a data retrieval method, comprising: in response to receiving retrieval data, performing multi-dimensional retrieval on the retrieval data to obtain an extracted data set corresponding to the retrieval data; using a pre-constructed hybrid retrieval database to perform matching retrieval on the extracted data set to obtain a retrieval knowledge set; based on the extracted data set, ranking each retrieval knowledge in the retrieval knowledge set according to its relevance to obtain a ranked knowledge set; using the hybrid retrieval database to perform relation expansion retrieval on the ranked knowledge set to obtain an expanded knowledge set; and integrating the ranked knowledge set and the expanded knowledge set to obtain a related knowledge set corresponding to the retrieval data.
[0005] Based on the same inventive concept, this application also provides a data retrieval device, comprising: an extraction module configured to, in response to receiving retrieval data, extract the retrieval data using multiple retrieval dimensions to obtain an extracted data set corresponding to the retrieval data; a first retrieval module configured to, using a pre-constructed hybrid retrieval database, perform matching retrieval on the extracted data set to obtain a retrieval knowledge set; a sorting module configured to, based on the extracted data set, sort each retrieval knowledge in the retrieval knowledge set according to its relevance to obtain a sorted knowledge set; a second retrieval module configured to, using the hybrid retrieval database, perform relation expansion retrieval on the sorted knowledge set to obtain an expanded knowledge set; and an integration module configured to, integrate the sorted knowledge set and the expanded knowledge set to obtain a related knowledge set corresponding to the retrieval data.
[0006] Based on the same inventive concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.
[0007] As can be seen from the above, the data retrieval method, apparatus, electronic device, and storage medium provided in this application include: In response to receiving retrieval data, extracting data from the retrieval data across multiple retrieval dimensions to obtain an extracted data set corresponding to the retrieval data, thereby achieving a comprehensive analysis of the retrieval data; Utilizing a pre-constructed hybrid retrieval database, performing matching retrieval on the extracted data set to obtain a retrieval knowledge set, expanding the data participating in the matching retrieval and enriching the retrieval knowledge; Based on the extracted data set, ranking each retrieval knowledge in the retrieval knowledge set according to its relevance to obtain a ranked knowledge set; quantifying the matching closeness according to the actual retrieval needs of the retrieval user, prioritizing the display of knowledge most closely matching the retrieval requirements; Utilizing the hybrid retrieval database, performing relational expansion retrieval on the ranked knowledge set to obtain an expanded knowledge set; and integrating the ranked knowledge set and the expanded knowledge set to obtain a related knowledge set corresponding to the retrieval data, thereby comprehensively supplementing the ranked knowledge set. Attached Figure Description
[0008] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0009] Figure 1This is a schematic flowchart of the data retrieval method according to an embodiment of this application;
[0010] Figure 2 This is a diagram illustrating the overall architecture of dynamic context-enhanced retrieval using a hybrid retrieval database, as described in an embodiment of this application. Figure 3 This is a schematic diagram of the data retrieval device according to an embodiment of this application; Figure 4 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation
[0011] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.
[0012] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0013] As described in the background section, in equipment operation and maintenance knowledge retrieval scenarios, user input often exhibits characteristics such as non-standard expressions, incomplete constraints, unstable semantic expression, and missing contextual information, placing high demands on the comprehensive adaptability of the retrieval system. Current existing operation and maintenance knowledge retrieval technologies have a single retrieval mode, failing to simultaneously address both precise constraint matching and fuzzy semantic matching, resulting in insufficient completeness of retrieval results and significant technical defects. Specifically, single precise matching retrieval can only achieve hard matching of explicit constraints such as keywords, equipment numbers, error codes, work order numbers, and time ranges, failing to identify non-standard synonymous expressions and implicit fault phenomena, and making it difficult to recall cross-document related knowledge, resulting in missed effective knowledge. Single semantic retrieval can only achieve text semantic similarity matching, easily ignoring key explicit constraints in operation and maintenance scenarios, leading to redundant and inaccurate recall results. Returning only a single hit result is insufficient to support continuous access to the "phenomenon-cause-handling-result" chain in operation and maintenance scenarios. In summary, existing retrieval technologies suffer from poor adaptability, limited filtering dimensions, and an inability to achieve multi-source knowledge linkage retrieval, ultimately resulting in incomplete retrieval results.
[0014] With the increasing complexity of information and communication systems and network structures, a large amount of operational data is generated during the maintenance process, characterized by dispersed sources, inconsistent formats, and significant differences in timeliness. This type of data typically includes monitoring metrics, real-time alarms, operation logs, work order records, standard procedures, historical fault cases, configuration lists, locally uploaded documents, local notes, and information returned by third-party interfaces. Due to the heterogeneous data sources, inconsistent semantic expressions, and complex temporal relationships, existing operational knowledge management and retrieval methods still have the following shortcomings: 1. Data is scattered, making unified utilization difficult. Monitoring, logs, alarms, work orders, configurations, documents, and other information are stored on different platforms, lacking a unified data object model. This forces operations and maintenance personnel to frequently switch between multiple systems, resulting in low efficiency in cross-system information association.
[0015] 2. Imbalance in knowledge organization granularity. Existing solutions often only create simple indexes for the entire document, making it difficult to adopt differentiated segmentation and expression methods for different types of knowledge such as procedure paragraphs, log event windows, alarm sequences, table content, and image descriptions. This can easily lead to problems such as overly coarse or overly fragmented recall results.
[0016] 3. Insufficient semantic understanding capabilities. The operation and maintenance corpus contains a large number of aliases, abbreviations, synonyms, internal system codes, and instances where equipment models and business terms are mixed. Traditional keyword retrieval struggles to accurately understand the true intent in natural language queries and also has difficulty handling cross-source knowledge alignment issues.
[0017] 4. Insufficient dynamic knowledge integration capabilities. Alarms, logs, work orders, and third-party interface data have significant time sensitivity, while traditional knowledge bases are mostly built offline and statically, making it impossible to dynamically adjust retrieval strategies based on real-time events, contextual sessions, and operational scenarios.
[0018] 5. Insufficient knowledge conflict and credibility control. Operations and maintenance scenarios involve both public domain knowledge such as standards, specifications, and regulations, and private domain knowledge such as internal procedures, failure cases, operational experience, and on-site notes. Existing solutions lack effective mechanisms for knowledge priority, version consistency, source credibility, and access control, easily leading to inconsistent or even unexecutable results.
[0019] 6. Insufficient traceability of results. Existing question-and-answer knowledge services often directly output conclusions, lacking source binding, version information, evidence fragments, and confidence level control, making it difficult to meet the requirements of operation and maintenance work for standardization, auditability, and verifiability.
[0020] Therefore, there is an urgent need for a method that can achieve unified access, standardized processing, knowledge objectification organization, hybrid index construction, and dynamic enhanced retrieval for multi-source heterogeneous operation and maintenance data, so as to improve the accuracy, response efficiency, standardization level, and knowledge reuse capability of operation and maintenance knowledge retrieval.
[0021] In view of this, embodiments of this application propose a data retrieval method, referring to... Figure 1 This includes the following steps: Step 101: In response to receiving the search data, perform multi-search dimension extraction on the search data to obtain the extracted data set corresponding to the search data.
[0022] In this step, it is checked whether retrieval data has been received. The retrieval data can be input by the user through the visual screen of the retrieval system, or remotely sent to the retrieval system. The retrieval system is a system capable of implementing the retrieval method for the data in this application. For example, the retrieval data may be various operation and maintenance related queries, fault descriptions, or retrieval command texts input by operation and maintenance personnel. Before using the retrieval system to conduct knowledge retrieval, the user must first complete system login and identity authentication. Upon receiving the retrieval data, multi-dimensional extraction is performed on the retrieval data to obtain the corresponding extracted data set. Key information is extracted from the retrieval data from multiple retrieval dimensions to achieve a comprehensive analysis of the retrieval data.
[0023] Step 102: Using a pre-built hybrid retrieval database, perform matching retrieval on the extracted data set to obtain a retrieval knowledge set.
[0024] In this step, the extracted dataset is matched and searched in a pre-built hybrid retrieval database to obtain a retrieval knowledge set. The extracted dataset includes data extracted from multiple retrieval dimensions. By performing matching and retrieval on the extracted dataset, the data participating in the matching and retrieval is expanded, thereby enriching the retrieval knowledge.
[0025] It should be noted that the hybrid retrieval database is used to centrally store various structured operation and maintenance field data, operation and maintenance text knowledge documents, and knowledge association retrieval index data. For example, the stored content includes structured field data such as equipment ledger parameters, work order information, and alarm codes, as well as unstructured operation and maintenance knowledge text such as fault handling procedures, historical work order cases, fault review reports, and alarm parsing documents.
[0026] Step 103: Based on the extracted data set, sort each retrieval knowledge in the retrieval knowledge set according to its relevance to obtain a sorted knowledge set.
[0027] In this step, the multiple search terms in the knowledge set are initially unordered. Based on the extracted data set, each search term in the knowledge set is ranked by relevance, resulting in a ranked knowledge set. For example, the ranked knowledge set might contain multiple independent and complete fault handling documents, each of which can be used independently as complete handling information by maintenance personnel. Relevance ranking quantifies the degree of matching based on the actual search needs of users, prioritizing the display of knowledge that best matches their search requirements.
[0028] Step 104: Using the hybrid retrieval database, perform relation expansion retrieval on the sorted knowledge set to obtain the expanded knowledge set.
[0029] In this step, the sorted knowledge set consists of core knowledge points that directly match the user's query. These are mostly fragmented, independent knowledge items, unable to form a complete knowledge chain for operation and maintenance. Therefore, a relational expansion search is performed on the sorted knowledge set in a hybrid retrieval database to obtain an expanded knowledge set. For example, the expanded knowledge set consists of complete supplementary documents such as accompanying alarm descriptions, preliminary troubleshooting steps, and historical cases of similar faults, linked by the knowledge relationship retrieval index. Retrieving chained data layer by layer from the hybrid retrieval database automatically fills in the information gaps caused by a single search, achieving the goal of comprehensively supplementing the sorted knowledge set without needing to initiate multiple searches to piece together relevant content.
[0030] Building upon the initial recall, this application further leverages event associations, object associations, process associations, and contextual connections between knowledge objects to dynamically enhance and expand highly relevant seed results. As a result, the retrieval system output is no longer limited to isolated fragments of single hits, but can be extended to related investigation steps, historical cases, handling records, associated alarms, work order processes, and procedural basis, thus forming a more complete chain of evidence or handling. Compared to existing retrieval methods that only return discrete text results, this application reduces comprehension biases caused by missing context, improves the efficiency of operations and maintenance personnel in directly utilizing retrieval results, and enhances the reference value of results in fault location, experience reuse, and handling tracing.
[0031] Step 105: Integrate the sorted knowledge set and the extended knowledge set to obtain the associated knowledge set corresponding to the retrieved data.
[0032] In this step, the ranked knowledge set comprises core, precise knowledge aligned with the search user's query intent, while the expanded knowledge set consists of supplementary materials derived from knowledge association links. To facilitate viewing for the search user, the ranked and expanded knowledge sets are integrated to obtain the associated knowledge set corresponding to the search data. After obtaining the associated knowledge set corresponding to the search data, this set is returned to the search user for their review. Through knowledge integration, the precisely matched core knowledge and the surrounding related knowledge obtained from the linked search are organized into a structured overall data set. This retains the core, precise content needed by the search user while supplementing the auxiliary related information required for the entire business chain, achieving comprehensive coverage of search results from a content perspective.
[0033] This application is applicable to scenarios involving unified access, standardized processing, knowledge organization, and enhanced retrieval of multi-source heterogeneous operation and maintenance data, including monitoring platforms, log platforms, alarm platforms, work order systems, configuration management systems, locally uploaded data, local notes, and third-party interfaces. Through unified data object modeling, standardized compilation, knowledge objectification processing, hybrid index construction, dynamic enhanced retrieval, and feedback update mechanisms, it achieves high-quality organization and high-precision retrieval of operation and maintenance knowledge.
[0034] The above scheme, in response to received retrieval data, extracts data from multiple retrieval dimensions to obtain an extracted data set, achieving a comprehensive analysis of the retrieval data. Using a pre-constructed hybrid retrieval database, the extracted data set is matched to obtain a retrieval knowledge set, expanding the data participating in the matching retrieval and enriching the retrieval knowledge. Based on the extracted data set, each retrieval knowledge in the retrieval knowledge set is ranked by relevance, resulting in a ranked knowledge set. The matching closeness is quantified according to the actual retrieval needs of the user, prioritizing the knowledge most relevant to the retrieval request. Using the hybrid retrieval database, the ranked knowledge set is used for relational expansion retrieval to obtain an expanded knowledge set. The ranked knowledge set and the expanded knowledge set are then integrated to obtain a related knowledge set corresponding to the retrieval data, comprehensively supplementing the ranked knowledge set.
[0035] In some embodiments, the multi-dimensional retrieval extraction includes text semantic extraction, structured attribute extraction, and data expansion extraction; the multi-dimensional retrieval extraction of the retrieval data to obtain the extracted data set corresponding to the retrieval data includes: preprocessing the retrieval data to obtain target retrieval data; performing text semantic extraction on the target retrieval data to obtain text semantic data; performing structured attribute extraction on the target retrieval data to obtain attribute data; performing data expansion extraction on the target retrieval data to obtain expanded data; and constructing the extracted data set by combining the text semantic data, the attribute data, and the expanded data.
[0036] In this embodiment, the search data input by the user contains invalid content. To improve the data quality of the search data, preprocessing is performed to obtain target search data. For example, preprocessing includes text cleaning, character normalization, error correction and regularization, and stop word removal. After the above preprocessing operations, the original messy search data is organized into regularized target search data. Text semantic extraction is then performed on the target search data to obtain text semantic data. For example, a pre-trained semantic model is used to perform text semantic extraction on the target search data. Text semantic extraction can understand the inherent meaning of sentences, so that subsequent retrieval is not limited to literal text comparison. It can identify non-standardized expressions in a predetermined scenario, and then combine vector similarity to recall semantically similar data to obtain text semantic data, thus solving the drawback of pure keyword retrieval, which is prone to missing relevant knowledge. For example, the user input search data is "Server number X crashed and I can't log in to the system. I need troubleshooting methods and to view similar fault work orders from this month." The target search data is "Server number X is down and I can't log in to the system. I need troubleshooting solutions and to retrieve similar fault work order records from this month." The semantic meaning of the target search data is "Server failure and downtime, business system unable to log in." The standard written description in the knowledge base is "Server process abnormality caused business system login failure." Traditional keyword search only compares text characters; if two texts use different words, they cannot be matched. Semantic extraction converts two sentences into semantic vectors. Since both express the same fault essence, their vector values are similar, and the corresponding troubleshooting documents can be matched based on similarity.
[0037] Structured attribute extraction is performed on the target retrieval data to obtain attribute data. Structured attribute extraction can combine Named Entity Recognition (NER) and regular expression matching technologies. Regular expressions are used to accurately match formatted fields such as device numbers, error codes, and time ranges. Then, a named entity recognition model intelligently extracts various entity attribute information from the text, organizing it into standardized attribute data to provide precise hard-line filtering boundaries for the retrieval process. Attribute data can include at least one of the following: device name, device number, system name, module name, work order number, alarm number, error code, time range, or source range. Structured attribute extraction is used to accurately capture the hard constraints in the target retrieval data. Relying on precise attribute data, the retrieval scope is narrowed, filtering out a large amount of irrelevant knowledge base content, compensating for the problem of broad scope and easy inclusion of irrelevant information in pure semantic retrieval, and improving the accuracy of retrieval results. For example, the hard constraint attribute data in the target retrieval data is extracted: Device Number = X, Time Range = This Month. Using this set of structured fields, the knowledge base is precisely filtered, directly excluding fault data from other numbered servers and not generated this month, narrowing the retrieval range and improving retrieval accuracy.
[0038] The target retrieval data is expanded and extracted to obtain extended data. For example, a pre-trained text intent classification model is used for data expansion and extraction, supplemented by keyword rule matching as an auxiliary verification method. Extended data can include at least one of the following: fault location intent, handling procedure intent, experience summary intent, related event intent, or historical case intent. By relying on intent recognition, implicit user needs can be proactively captured, semantic intent discrimination is performed on the target retrieval data, deep retrieval needs are identified, and standardized extended data is generated, eliminating the need for users to supplement their queries and improving the completeness of the retrieval data. For example, the intent classification model identifies two types of user needs: obtaining fault handling procedures and querying historical fault work orders, and generates corresponding extended data.
[0039] The textual semantic data, attribute data, and extended data are combined to form the extracted data set, which can be represented as Q: (1) in, It represents textual semantic data, used to characterize the natural language semantic content in user input; Represents attribute data used to characterize explicit search boundaries in user input; This represents extended data used to characterize whether relation expansion, disposal chain expansion, or historical association expansion needs to be performed. Through multi-dimensional retrieval, subsequent retrieval processes no longer treat user input as undifferentiated raw text, but rather decompose it into query objects that can be applied to different index channels. This avoids the decreased matching efficiency and recall bias problems caused by all information being mixed in a single query string in traditional retrieval.
[0040] In some embodiments, the step of using a pre-built hybrid retrieval database to perform matching retrieval on the extracted data set to obtain a retrieval knowledge set includes: for each data point in the extracted data set, based on the retrieval dimension corresponding to the data, calling the underlying retrieval index corresponding to the retrieval dimension in the hybrid retrieval database; retrieving the data in the hybrid retrieval database according to the underlying retrieval index to obtain the initial knowledge corresponding to the data; obtaining the retrieval permissions of the retrieval user corresponding to the retrieval data, and based on the retrieval permissions, performing knowledge removal on the initial knowledge corresponding to each data point to obtain retrieval knowledge; and merging and deduplicating the retrieval knowledge corresponding to all data points to obtain the retrieval knowledge set.
[0041] In this embodiment, for each data point in the extracted dataset, based on the corresponding search dimension, the underlying search index corresponding to that search dimension is invoked in the hybrid search database. The extracted dataset consists of textual semantic data, attribute data, and extended data. Each of these three types of data corresponds to an independent search dimension and forms a one-to-one mapping with the underlying search index: textual semantic data matches the semantic search dimension, corresponding to the invoked vector search index; attribute data matches the attribute filtering dimension, corresponding to the invoked field search index; and extended data matches the knowledge association dimension, corresponding to the invoked relation search index. Following this fixed mapping relationship, the underlying search index matching each search dimension can be invoked in the hybrid search database to perform the search and obtain the initial knowledge corresponding to the data. In the hybrid retrieval database, similarity recall is performed on semantic vectors using vector retrieval indexes to capture fuzzy matching knowledge sets with different textual expressions but consistent semantic meanings. Precise matching is achieved based on explicit constraints of structured attributes using field retrieval indexes, prioritizing candidate knowledge object sets that satisfy constraints such as device object, system object, time range, alarm number, or work order number. Relational retrieval indexes, combined with various enhanced intents carried by extended data, retrieve adjacent knowledge sets that have event associations, process associations, or contextual associations with the target retrieval data. The knowledge from the fuzzy matching knowledge set, candidate knowledge set, and adjacent knowledge set is used as initial knowledge.
[0042] Based on the dedicated underlying search indexes matched by each search dimension, searches are performed in the hybrid search database to obtain the corresponding initial knowledge. The field search index can quickly and accurately locate compliant resources by relying on structured parameters. The vector search index can match faulty data that are different in literal sense but have the same meaning. The relational search index can retrieve upstream and downstream related knowledge in a linked manner. The parallel search of multiple indexes not only avoids the resource consumption and delay caused by traversing the entire database and improves the search response speed, but also effectively reduces the data omission problem caused by a single search mode, and improves the richness and matching quality of the initial knowledge retrieved.
[0043] The system obtains the retrieval permissions of the users corresponding to the retrieved data. Based on these permissions, it removes duplicate knowledge from the initial knowledge corresponding to each piece of data to obtain the retrieval knowledge. The retrieved initial knowledge includes all matching content in the database, which contains confidential data that ordinary operators do not have access to. By removing knowledge based on retrieval permissions, the risk of confidential data leakage can be eliminated at the source. The retrieval knowledge corresponding to all data is merged and deduplicated to obtain the retrieval knowledge set. After permission filtering, the multi-channel retrieval knowledge shows that the same document is repeatedly hit by field retrieval, semantic retrieval, and correlation retrieval. Therefore, it is necessary to uniformly summarize the retrieval knowledge generated from various dimensions and perform deduplication processing to eliminate duplicate and redundant data caused by parallel retrieval from multiple indexes.
[0044] Retrieve knowledge sets Represented as: (2), in, The recall results corresponding to the text semantic extraction dimensions The recall results corresponding to the structured attribute extraction dimensions. The pre-expanded results correspond to the data expansion extraction dimensions. The three types of recall are not mutually exclusive, but rather executed in parallel and then merged. This simultaneously addresses two typical scenarios: one is the precise location scenario where the input contains clear identifying information, and the other is the fuzzy retrieval scenario containing only phenomenon descriptions or colloquial expressions. Parallel recall significantly reduces knowledge omissions caused by a single recall mode.
[0045] In some embodiments, the step of ranking each search knowledge in the search knowledge set based on the extracted data set to obtain a ranked knowledge set includes: performing feature identification on the extracted data set to obtain at least one feature; determining the relevance level of each feature among all predetermined relevance levels; determining the weight corresponding to each relevance level among all predetermined relevance levels based on the relevance level of each feature; performing a weighted calculation of the relevance level of each search knowledge in the search knowledge set and the data corresponding to the search knowledge in the extracted data set based on the weight corresponding to each relevance level to obtain a relevance score corresponding to the search knowledge; and ranking all search knowledge according to the relevance score corresponding to each search knowledge to obtain the ranked knowledge set.
[0046] In this embodiment, the weight corresponding to each relevance level in this application is not fixed in advance, but dynamically adjusted according to the characteristics of the retrieved data. Therefore, feature identification is performed on the extracted data set to extract various query features inherent in the retrieved data. For each feature, the relevance level to which the feature belongs in all predetermined relevance levels is determined. All predetermined relevance levels include field matching degree, semantic similarity degree, relation matching degree, and quality reliability degree. The feature classification is determined by identifying the content attributes in the extracted data set: if the feature is a standardized fixed parameter such as device number, error code, time interval, or work order number, it is classified into the field matching degree category; if the feature is text content such as fault phenomena or system anomalies described by users in natural language, it is classified into the semantic similarity degree category; if the feature is an extended retrieval requirement proposed by users, such as querying historical work orders, viewing handling processes, or retrieving related alarms, it is classified into the relation matching degree category; if the feature is reflected in the update time, verification status, frequency of implementation, official certification marks, or other attribute information of the knowledge document itself, it is classified into the quality reliability degree category. The identified query features are accurately classified into four categories of predefined relevance: field matching, semantic similarity, relationship matching, and quality reliability. This clearly distinguishes the attribute types of different query information and provides a reliable basis for subsequent dynamic allocation of weights for each dimension.
[0047] Based on the relevance of each feature, the weight corresponding to each relevance level among all predetermined relevance levels is determined. Specifically, in response to the lack of available weights for each relevance level corresponding to the user's previous search behavior, the initial weights for each relevance level are adjusted based on the relevance of each feature: the initial weights for each feature's relevance level are increased by a first predetermined adjustment value, and the initial weights for relevance levels that do not belong to each feature are decreased by a second predetermined adjustment value. The increased and decreased initial weights are then normalized to obtain the weights for each relevance level among all predetermined relevance levels. In response to the availability of previous weights for each relevance level corresponding to the user's previous search behavior, the previous weights for each relevance level are adjusted based on the relevance of each feature: the previous weights for each feature's relevance level are increased by a first predetermined adjustment value, and the previous weights for relevance levels that do not belong to each feature are decreased by a second predetermined adjustment value. The increased and decreased previous weights are then normalized to obtain the weights for each relevance level among all predetermined relevance levels.
[0048] When historical user search weight data is unavailable, the system uses a pre-set initial weight as the base. When previous search weights are available, they are used as the adjustment base. The system uniformly adjusts the weights of matching features by adding a first predetermined adjustment value and subtracting a second predetermined adjustment value from the weights of other irrelevant dimensions, based on the relevance of the search features. After normalization, valid weight values are output. This approach can adaptively tilt the corresponding dimension weights in real time according to the needs of a single search, adapting to the ranking optimization requirements of different search scenarios. It can also accumulate user search behavior data to continuously optimize the weight allocation strategy, balancing algorithm stability, scenario adaptability, and personalized search optimization effects.
[0049] When a search request contains explicit device number, work order number, alarm number, or time range, the weight of field matching items is increased; when the search request is mainly a natural language fault description, the weight of semantic matching items is increased; when the search request explicitly includes extended expressions such as "related processing steps," "similar cases," or "related alarms," the weight of relational items is increased; and when there is conflicting knowledge from multiple sources, the weight of quality score items is increased. This allows the ranking mechanism to adaptively optimize for different search scenarios, rather than applying the same static scoring logic to all requests.
[0050] For each piece of retrieval knowledge within the retrieval knowledge set, and considering the dynamic weights assigned to each relevance matching dimension, individual matching scores are calculated between the retrieval knowledge and the extracted data set in terms of field matching, semantic similarity, relation matching, and quality reliability. These scores are then weighted and fused to generate a unified comprehensive relevance score. The relevance score corresponding to the retrieval knowledge is determined using the following formula: Let the candidate knowledge object be Then its initial comprehensive score Represented as: (3) in, This represents the field matching score between the extracted data in the dataset and its corresponding retrieval knowledge. Indicates the semantic matching score. Indicates the score related to the relationship. This represents the quality score of the knowledge object itself. The dynamic weights for each scoring item are given, and the sum of the four weights is 1.
[0051] For example, the field matching score calculates the fit of structured fields through exact string matching, regular expression matching, and numerical range comparison; the semantic similarity score generates text vectors using a pre-trained semantic model optimized for the operations and maintenance domain, and combines this with a cosine similarity algorithm to quantify the semantic fit between two text segments; the relationship matching score relies on a pre-built operations and maintenance knowledge graph, using graph traversal and shortest path solving methods to assess the closeness of knowledge associations based on their distance; the quality and reliability score is a multi-indicator weighted quantification of multiple inherent attributes such as document update time, actual usage frequency, official review markings, and content completeness. These four individual scores are normalized before being substituted into the weighted calculation formula to participate in the overall relevance calculation. A single-dimensional matching score cannot comprehensively measure the fit between search knowledge and user search needs. The importance of each matching dimension varies significantly in different search scenarios, requiring dynamic weights to differentiate the contribution ratio of each dimension. It can amplify the score influence of key matching dimensions according to the actual needs of this search, weaken the interference of unnecessary dimensions, and integrate multiple evaluation indicators such as structured matching, semantic matching, related knowledge matching, and document quality into a quantifiable unified score, so that the knowledge matching degree has a quantifiable numerical basis.
[0052] Based on the relevance score of each search term, all search terms are sorted to obtain a sorted knowledge set. This transforms quantified matching values into a visual content arrangement order, resolving the problem of disorganized knowledge accumulation and the burying of relevant information by irrelevant data after multi-channel retrieval. Search terms with higher relevance and credibility to the user's search intent are prioritized and displayed first, while less relevant content is arranged later. This helps users quickly locate the necessary troubleshooting solutions, work order information, and reference cases, improving overall knowledge retrieval efficiency and the user experience of the search service.
[0053] In some embodiments, the ranked knowledge set includes multiple search knowledge items sorted from high to low relevance; the step of using the hybrid search database to perform relation expansion search on the ranked knowledge set to obtain an expanded knowledge set includes: taking a predetermined number of search knowledge items from the ranked knowledge set in a forward-to-back order as multiple target search knowledge items; for each target search knowledge item, in the hybrid search database, starting from the target search knowledge item, performing a neighborhood expansion search along the relation search index in the hybrid search database to obtain the expanded knowledge item corresponding to the target search knowledge item; and constructing the expanded knowledge set from the expanded knowledge items corresponding to the multiple target search knowledge items.
[0054] In this embodiment, the sorted knowledge set includes multiple search knowledge items ranked from high to low relevance. A predetermined number of search knowledge items are selected from the sorted knowledge set in a sequential order as multiple target search knowledge items to conduct subsequent relation expansion searches. This avoids the problem of low-value expansion content and ineffective consumption of search resources caused by using low-matching invalid knowledge as the source of expansion. The predetermined number is determined based on historical experience. For each target search knowledge item, in the hybrid search database, starting from the target search knowledge item, a neighborhood expansion search is performed along the relation search index in the hybrid search database to obtain the expansion knowledge corresponding to the target search knowledge item. Specifically, starting from the target search knowledge item, the first-order and second-order neighborhood association nodes of the traversal starting point are traversed based on the relation search index pre-built in the hybrid search database. For each neighborhood association node, the relationship strength, contextual continuity, and information supplement contribution of the neighborhood node knowledge to the current search request are extracted between the target search knowledge item and the knowledge corresponding to that neighborhood association node. Based on the relationship strength, contextual continuity, and information supplement contribution, a comprehensive expansion strength value is determined. Neighborhood node knowledge items with a comprehensive expansion strength greater than a preset expansion threshold are selected as expansion knowledge items. The overall propagation strength value is determined using the following formula: Let the target retrieval knowledge be The knowledge of the neighboring nodes that are connected to it is Then the comprehensive extension strength value It can be represented as: (4) in, Represents target retrieval knowledge Knowledge of neighboring nodes The strength of the relationship, This indicates the degree of continuity between the two in the context. Representing knowledge of neighboring nodes The extent to which it supplements the current retrieved data Q. To expand the weights.
[0055] when When the value exceeds the preset expansion threshold, the knowledge of neighboring nodes will be transferred. Incorporate into the expanded knowledge set. During the expansion process, prioritize retaining the following types of knowledge relationships: the relationship between phenomenon records and handling records under the same fault event, the relationship between historical cases under the same equipment object, the relationship between preceding and subsequent steps in the same work order process, the explanatory relationship between alarm records and procedure clauses, and the conclusion relationship between problem descriptions and debriefing summaries.
[0056] Therefore, when faced with typical operational questions such as "how to handle this", "why this error occurred", "how this fault was resolved before", or "what other problems occurred along with this alarm", the retrieval system will not only return isolated fragments, but will be able to return a set of knowledge results with structural continuity.
[0057] By combining a set expansion threshold, qualified neighborhood knowledge is selected as extended knowledge for the target retrieval knowledge, and the extended knowledge corresponding to multiple target retrieval knowledge is formed into an extended knowledge set. By using the relational retrieval index to constrain the retrieval traversal range, and combining it with the expansion strength threshold to filter weakly related invalid data, the missing contextual information and complete processing chain of core knowledge are effectively supplemented, allowing the final output knowledge content to form a systematic solution system. Relational expansion retrieval refers to not directly using the initial hit result as the final output, but rather, based on the user's query intent and the contextual relationship of seed knowledge objects, continuing to expand related knowledge objects along the relational retrieval index, thereby constructing a more complete evidence chain or processing chain.
[0058] It should be noted that the relationship retrieval index is invoked during the matching retrieval phase. Its function is to use the retrieved data as an entry point to match related knowledge directly related to the query statement. This is a positive recall based on user input. This step, on the other hand, is to reverse-engineer the supporting knowledge based on the high-quality knowledge entities that have been hit after the high-quality retrieval results are produced. The two have completely different retrieval entry points, retrieval targets, and data recall scopes. They are not duplicate retrieval behaviors. The initial relationship retrieval is responsible for capturing the main information that directly matches the user's query. This neighborhood expansion retrieval is responsible for supplementing the main information with supporting materials such as processes, cases, alarm explanations, and review records.
[0059] In some embodiments, integrating the ranked knowledge set and the extended knowledge set to obtain the associated knowledge set corresponding to the retrieved data includes: merging the ranked knowledge set and the extended knowledge set to obtain a first knowledge set; performing deduplication, weak relevance filtering, and conflict resolution on the first knowledge set to obtain a second knowledge set; re-ranking the second knowledge set according to relevance to obtain a third knowledge set; for each knowledge in the third knowledge set, obtaining the corresponding associated information and binding the associated information to the knowledge, wherein the associated information includes at least the source anchor of the knowledge; and obtaining the associated knowledge set in response to determining that all knowledge has been bound with associated information.
[0060] In this embodiment, the sorted knowledge set and the extended knowledge set are merged to obtain a first knowledge set. This first knowledge set, formed by merging the two sets, summarizes precisely matched core complete data and related supplementary complete data, gathering all the necessary materials to support fault handling. The first knowledge set undergoes deduplication, weak correlation filtering, and conflict resolution to obtain a second knowledge set. Deduplication technology removes homogeneous maintenance data repeatedly retrieved through multiple search paths and prioritizes retaining the best knowledge with high field completeness, reliable sources, and more recent timeliness. Weak correlation filtering removes invalid content that is only semantically similar but has no actual business relevance and lowers the priority of the remaining weakly related knowledge. For contradictory knowledge, a decision is made based on knowledge quality score, source credibility, and update time. This process eliminates redundant, noisy, and contradictory data, simplifies the knowledge volume, and ensures the authenticity and effectiveness of the information.
[0061] The second knowledge set is reordered according to relevance to obtain a third knowledge set. The arrangement of the mixed knowledge is then standardized and the presentation logic of prioritizing highly matched knowledge is maintained. For each piece of knowledge in the third knowledge set, the corresponding related information is obtained and bound to the knowledge. The related information includes at least the source anchor of the knowledge, and may also include related knowledge, extension paths, and recommended reading order. After all knowledge is bound, the final set of related knowledge is output. This entire integration process eliminates the problems of fragmented, messy, and untraceable information in the original search results. While enriching the breadth of knowledge, it also completes noise reduction and purification, order standardization, and source traceability annotation. Finally, it outputs structured knowledge results with complete evidence sources, related contexts, and extension links to the operation and maintenance personnel, significantly improving the usability, credibility, and reading efficiency of the search results.
[0062] This application retains associated information throughout the knowledge object generation, index construction, and retrieval output processes, ensuring that any retrieval result can be traced back to the original document, record, or event source. It also enables permission verification during the recall, expansion, and output stages to prevent incorrect access to knowledge from different scopes. Furthermore, this invention introduces a retrieval feedback update mechanism. By collecting user feedback information such as clicks, pauses, secondary searches, favorites, confirmations, or abandonments during the retrieval process, it adaptively adjusts field matching weights, semantic matching weights, relationship expansion weights, and ranking strategies. This improves the credibility, verifiability, and security controllability of retrieval results, while also enhancing the invention's continuous adaptability to different operational scenarios.
[0063] In some embodiments, the method for constructing the hybrid retrieval database includes: acquiring multi-source heterogeneous raw data; standardizing the raw data to obtain multiple initial standard knowledge; extracting relationships from the multiple initial standard knowledge to obtain an original set of association relationships; adding an identifier to each initial standard knowledge to obtain standard knowledge; constructing a standard knowledge set from all standard knowledge; performing knowledge replacement on the original set of association relationships based on all standard knowledge to obtain a standard relationship set; extracting and weighting multi-feature dimension knowledge from each standard knowledge in the standard knowledge set to obtain fused knowledge; constructing a field retrieval index based on the standard knowledge set, constructing a vector retrieval index based on all fused knowledge, and constructing a relationship retrieval index based on the standard relationship set, and using the field retrieval index, the vector retrieval index, and the relationship retrieval index as underlying retrieval indexes respectively; and uniformly associating and encapsulating each underlying retrieval index, the standard knowledge set, and the fused knowledge based on the identifier corresponding to each standard knowledge to obtain the hybrid retrieval database.
[0064] In this embodiment, heterogeneous raw data from multiple sources is acquired. Heterogeneous sources refer to raw data originating from multiple independent collection channels, with inconsistencies in storage structure, text format, terminology, or content granularity. Raw data from different sources can enter the retrieval system through file uploads, API retrieval, message subscriptions, or manual submission. Raw data may include document-type data such as operation and maintenance procedures and fault solutions; log-type data of server and application operation processes; alarm-type data from monitoring platforms; work order-type data recording the entire fault handling process; configuration data of static hardware and software parameters; and manual operation and maintenance notes retained by operation and maintenance personnel to record fault experience. The raw data undergoes standardization processing to obtain multiple initial standard knowledge sets, and relationships are extracted from these standard knowledge sets to obtain an initial set of association relationships. Standardization processing includes unified encapsulation, parsing, standardized compilation, consistency correction, and segmentation. Through standardization processing, the raw operation and maintenance data, which is scattered in origin, has different structures, and inconsistent expressions, is transformed into standardized knowledge required for subsequent knowledge representation construction and dynamic enhancement retrieval. The main focus is on addressing the problems that arise from differences in the source, format, naming conventions, time representation, content granularity, and contextual organization of raw data, which make it difficult to directly construct a unified knowledge base and maintain the integrity and traceability of knowledge retrieval results.
[0065] To eliminate differences in format and organization between data from different sources, firstly, each original record in the original data... Perform unified encapsulation to generate the corresponding source object. : (5) in, Indicates from the original record The original content extracted from it, This represents the metadata corresponding to the original content. This represents the location information or reference pointer of the original content in the source system. This indicates a unified encapsulation of mapping functions. Meta-information It should include at least the source identifier, time information, object identifier, and permission information, namely: (6) in, Indicates the data source. Indicates the generation time or collection time. This indicates the associated device, system, module, event, or work order object. Indicates the scope of access control.
[0066] Through the unified encapsulation described above, subsequent processing no longer directly relies on the data structures of each source system, but is parsed and compiled based on the unified source object, thus providing a foundation for multi-source heterogeneous data to enter the same knowledge processing chain.
[0067] After generating a unified source object, content parsing is performed on the source object. For structured data, the correspondence between fields and values is extracted; for semi-structured data, hierarchical relationships, status information, and event association information are extracted; for unstructured text, the semantics of titles, body text, tables, and paragraphs are extracted; for log and alarm data, time, object, event, level, and upstream and downstream relationships are extracted; for mixed text and image data, identifiable text and its association information with corresponding operation and maintenance events are extracted.
[0068] Based on the parsed results, standardization and reorganization are performed. Standardization and reorganization includes time unification, naming consistency, duplicate merging, and conflict correction. For issues related to inconsistent time representations, the original time... By using a unified time standardization function Mapped to standard timestamps : (7) For issues arising from inconsistencies in device names, system names, module names, alarm abbreviations, and historical aliases, the original terminology will be... Mapping to standard terms : (8) in, The terminology normalization function is used to merge multiple different expressions into a unified semantic object, while preserving the original expression and ensuring traceability.
[0069] In order to perform quality control on the parsed records, this application calculates a quality score for each record. : (9) in, Indicates the integrity score. Indicates consistency score, Indicates timeliness score, This indicates the credibility score of the source; And satisfy: (10) when When this happens, the record is marked as a low-quality record and is subject to removal, demotion, or pending review. This is a preset quality threshold.
[0070] For multiple records of the same object from different sources, such as equipment status records, alarm records, or work order records, this application merges them based on object identifier, time proximity, and content similarity. If two records... and satisfy: (11) If the two are determined to be different sources of the same operational fact, a merging process needs to be performed. Represents the comprehensive similarity function. This is the merging threshold.
[0071] Through the above process, operation and maintenance records from multiple sources and with different expressions can be compiled into a consistent standardized result, avoiding a large amount of duplicate knowledge, conflicting knowledge, and low-quality noise knowledge in the subsequent knowledge construction process.
[0072] After standardization and compilation are completed, knowledge segmentation is performed on the standardization results to obtain initial standard knowledge. Unlike segmentation based on a fixed number of words or a fixed length, this application adopts a segmentation strategy oriented towards semantic integrity. That is, the segmentation position is determined by the content structure, semantic continuity, and business boundaries, so that the segmentation result can independently express relatively complete knowledge units without destroying the context.
[0073] Let a standardized content sequence be represented as: (12) in, It represents the basic semantic unit in the content, which can be a paragraph, log fragment, alarm entry, work order step, or table area.
[0074] Divide sequence D into M candidate segments: (13) This application determines the optimal segmentation boundary based on the principle of "high semantic cohesion within segments and small semantic breaks between segments". Its objective function can be expressed as: (14) in, Representing fragments Internal semantic cohesion, Indicates the strength of dependency between adjacent segments. This is the balance coefficient.
[0075] When segmenting data, for document-type content, priority is given to segmenting according to heading level, paragraph topic, and table boundaries; for logs, alarms, and event records, priority is given to segmenting according to time window, event chain, and handling stage; for work orders and fault records, priority is given to segmenting according to problem description, troubleshooting process, handling actions, and results; for mixed text and graphics data, the correspondence between text and graphics is maintained, and mutually explanatory content is not separated.
[0076] If necessary, preserve the contextual overlap between adjacent segments to enhance semantic continuity during subsequent knowledge retrieval. If segments End content and fragments If the content at the beginning of the document is strongly related, then an overlapping window of length l can be retained. This results in an enhanced fragment expression: (15) This avoids the problem description and processing conclusion being separated due to segmentation, thereby improving the accuracy of subsequent knowledge construction.
[0077] After completing the knowledge segmentation, each segment is generated as a knowledge object. : (16) in, Indicates the original content expression, This represents the standardized content expression. Indicates source anchor information, This indicates object association information and permission boundary information. This represents a function that generates knowledge objects.
[0078] To address the relationships between multiple knowledge objects, this application constructs a set of relationships. : (17) in, Representing knowledge objects With knowledge objects The relationship types include, but are not limited to, relationships between devices, events, work orders, pre- and post-processing, or contextual connections.
[0079] Ultimately, the output consists of a set of knowledge objects composed of multiple initial standard knowledge sets. and its original set of relationships : (18) This output serves as input for subsequent knowledge representation construction and retrieval index generation.
[0080] The working principle of standardization can be summarized in three layers: First, raw data scattered across different systems is uniformly converted into processable source objects; second, the messy, repetitive, and conflicting expressions in the source objects are integrated into consistent standardized content; third, the standardized content is segmented into initial standard knowledge that maintains semantic integrity and is suitable for subsequent invocation, and the relationships between the initial standard knowledge are established. Therefore, standardization is not a simple data access process, but a transformation from raw operation and maintenance records to standardized knowledge objects. After this step, subsequent knowledge representation construction no longer faces the problem of fragmented raw data from multiple systems, but is built on a unified knowledge object basis, thereby improving the efficiency of unified knowledge base construction, the accuracy of knowledge retrieval, and the traceability of results.
[0081] This application achieves unified access, standardized compilation, and knowledge objectification preprocessing of multi-source heterogeneous operation and maintenance data, including documents, logs, alarms, work orders, configuration records, and manual notes. This transforms raw records, originally scattered across different systems, formats, and expressions, into knowledge objects with unified structure, regular semantics, and traceable origins. This effectively solves the problems of scattered data sources, inconsistent naming, inconsistent record granularity, and difficulty in connecting contexts in existing technologies. It also reduces the probability of duplicate, conflicting, and low-quality noisy knowledge entering the knowledge base, thereby improving the standardization and usability of the unified knowledge base construction.
[0082] For each initial standard knowledge, an identifier is added to the initial standard knowledge to obtain standard knowledge; all standard knowledge constitutes a standard knowledge set; and based on all standard knowledge, the original set of relationships is replaced with knowledge to obtain a standard relationship set. By assigning a unique identifier to each initial standard knowledge to form standard knowledge and constructing a standard knowledge set, and simultaneously performing entity replacement on the original set of relationships based on standard knowledge, the related entities of the relationships can be transformed from unstandardized raw data units into standard knowledge units with unified identifiers and consistent specifications. On the one hand, this achieves unified encoding of knowledge entities, eliminating matching ambiguities caused by heterogeneous object representations in the original relationships, facilitating index construction, knowledge retrieval, and relationship traversal; on the other hand, it establishes a binding correspondence between standard knowledge and relationships, ensuring accurate linkage between knowledge entities and relationships during subsequent hybrid index construction and dynamic enhanced retrieval, improving the accuracy of relationship expansion queries, and relying on standardized identifiers to achieve unified management of knowledge changes, traceability, and access control.
[0083] After the above processing, the original operation and maintenance data has been transformed into standardized knowledge objects. However, there are still problems between different knowledge objects: text can be searched but semantics cannot be computed; semantics can be expressed but fields are difficult to precisely constrain; local knowledge can be located but contextual relationships are difficult to extend. If only a single inverted index is used, it is easy to miss synonyms, implicit fault descriptions, and cross-document related knowledge; if only a single vector retrieval index is used, it is easy to weaken the ability to hit precise fields such as device number, alarm code, work order number, error code, and time constraint. Therefore, this application constructs a hybrid index structure for operation and maintenance knowledge base to simultaneously take into account precise matching ability, semantic recall ability, and relationship extension ability.
[0084] For each piece of standard knowledge in the standard knowledge set, multi-dimensional feature extraction and weighted fusion are performed to obtain fused knowledge. For example, for this standard knowledge... The standardized text content, attribute information, and contextual information of the standard knowledge are extracted and a corresponding multidimensional knowledge representation is generated to obtain the fused knowledge. Specifically, the standardized text content reflects the semantic subject of the standard knowledge; the attribute information reflects the structured features of the standard knowledge, such as the corresponding device, system, module, event, time, source, and permission boundaries; and the contextual information reflects the sequential connections, relationships with the same event, the same device, or the same processing procedure between the standard knowledge and other knowledge objects.
[0085] Standard knowledge The text semantic representation is denoted as Attributes are represented as The associated context is denoted as The fusion of these three elements yields the fused knowledge of standard knowledge. : , (19) in, To merge weights, and satisfy .
[0086] In practical applications, for text-dominant knowledge objects such as procedures, reports, and experience summaries, the weight of textual semantic representation can be increased; for knowledge objects with strong structural fields such as alarm records, work order records, and configuration records, the weight of attribute representation can be increased; and for knowledge objects with strong correlations between handling chains, event evolution chains, and upstream and downstream faults, the weight of contextual representation can be increased. The resulting unified representation vector... It is no longer a single text expression, but carries three types of information at the same time: semantic content, business attributes, and relational context, making it more suitable for knowledge retrieval in subsequent operation and maintenance scenarios.
[0087] It should be noted that when generating a unified representation, this application does not replace the original text content and structured fields. Instead, it generates a computable representation while retaining the original traceable content. That is, the knowledge object still retains the original content anchor and field anchor to ensure that subsequent hit results can be traced back to the original source.
[0088] After completing the construction of the unified knowledge representation, a field retrieval index is built based on the standard knowledge set. This field retrieval index is divided into a structured attribute sub-index and a text inverted index. The structured attribute sub-index enables precise field filtering, while the text inverted index constructs keyword mappings based on knowledge text segmentation, together achieving precise conditional retrieval. For example, each standard knowledge entry with a unique identifier is traversed, and normalized structured fields such as device number, work order number, error code, timestamp, and permission tag are extracted. Hash equal-value indexes, time partition range indexes, and tag indexes are constructed according to the code, time, and tag, respectively, forming the structured attribute sub-index. At the same time, the standardized text of the standard knowledge is extracted, segmented by the operation and maintenance dictionary, and stop words are removed. An inverted linked list is built with the entry as the key and the standard knowledge ID list as the value, and associated with synonym mappings to obtain the text inverted index. Finally, the two types of sub-indexes are encapsulated into a unified field retrieval index.
[0089] Structured attribute sub-indexes and text inverted sub-indexes are generated for precisely locatable information within knowledge objects. For example, the structured attribute sub-indexes primarily target content with clear identifying significance in operational scenarios, including but not limited to device names, device numbers, system names, module names, work order numbers, alarm numbers, error codes, timestamps, source identifiers, and standardized terminology. The text inverted sub-indexes, on the other hand, perform term segmentation and index registration for standardized text content within knowledge objects, enabling any term to quickly locate the set of knowledge objects containing that term.
[0090] In this process, not only are standard terms indexed, but a mapping index is also created for the original terminology expressions before normalization. This is because operations and maintenance personnel often mix abbreviations, outdated terms, colloquial expressions, and formal terms during searches. Retaining only the standard terminology index would weaken the compatibility of the search entry point; retaining only the original terminology index would lead to synonym splitting and result dispersion. Therefore, this application improves the fault tolerance and consistency of the search entry point by using both standard terminology indexes and original terminology mapping indexes.
[0091] Furthermore, for the time field, object field, and source field, a field partitioning approach is used to establish index boundaries, enabling the subsequent retrieval process to complete the necessary range constraints before semantic recall. Let the field constraint sub-items in the query be... Knowledge objects The corresponding field retrieval index information is Then the set of candidate knowledge objects that satisfy the field constraints Represented as: (20) in, This represents a field matching function used to determine knowledge objects. Does it meet the constraints of device, work order, alarm, error code, time range, or source range in the query object?
[0092] Using the above method, when a search request is explicitly limited to a specific device, work order, or time range, a set of candidate knowledge objects can be obtained first by searching the field index. Then, it enters the semantic matching stage, thereby reducing retrieval overhead and improving hit accuracy.
[0093] After completing the structured attribute sub-index and the text inverted sub-index, a vector retrieval index is constructed based on all fused knowledge, providing a unified representation of vectors. Establish a vector retrieval index. For example, iterate through each piece of fused knowledge that has completed multi-dimensional feature fusion, extracting standardized text semantic vectors, attribute feature vectors, and relational representation vectors from the fused knowledge. These are then weighted and concatenated according to preset weight coefficients to generate a unified high-dimensional embedding vector for the entire fused knowledge. The unique identifier of each piece of fused knowledge is bound to its corresponding embedding vector and stored, constructing a vector library index structure. A vector similarity calculation interface is simultaneously set up. Corresponding knowledge permission verification rules and source information are bound to the vector retrieval index. During retrieval, cosine similarity calculation is used to match semantically similar fused knowledge. An incremental update strategy is configured to update the vector retrieval index synchronously when new or modified fused knowledge is added, ultimately completing the construction of the vector retrieval index.
[0094] This vector retrieval index is used to support the retrieval of knowledge objects that are semantically similar but not entirely consistent in expression. It is especially suitable for operation and maintenance scenarios such as non-standard descriptions of fault phenomena, incomplete expression of problems, and semantic inconsistencies between log errors and experience summaries.
[0095] In this application, each knowledge object generates a corresponding index entry. : ,(twenty one), in, Represents the identifier of a knowledge object. A unified representation vector for this knowledge object. Indicates the original source anchor point. This indicates the field retrieval index information and access control information.
[0096] The creation of the vector retrieval index is not done in isolation, but is simultaneously bound to the original source anchor, field anchor, and permission boundaries of the knowledge object. In this way, in the subsequent recall stage, even if a knowledge object is hit through semantic similarity, the source location, object tag, and access permission information of that knowledge object can be obtained simultaneously, rather than obtaining an isolated vector result that cannot be traced back or audited.
[0097] Unlike general text retrieval, the vector retrieval index in this application emphasizes semantic recall at the "operational knowledge object level," rather than approximate matching of simple natural language sentences. In other words, the vector retrieval index corresponds to a normalized knowledge object, rather than unprocessed raw text fragments, thus its recall results have stronger business completeness and contextual availability.
[0098] In addition to field retrieval indexes and vector retrieval indexes, this application further establishes a relationship retrieval index for the set of relationships between knowledge objects. The relationship retrieval index is used to record event associations, object associations, process associations, and sequential connections between knowledge objects. Based on a standard relationship set, the relationship retrieval index is constructed by, for example, traversing each association in the standard relationship set consisting of two unique standard knowledge identifiers, distinguishing relationship types (such as association with the same device, work order-alarm causal association, fault handling association, etc.), and constructing a forward adjacency index using the standard knowledge ID as the starting point as the primary key, mapping and storing all associated target knowledge IDs and relationship types; and constructing a reverse backtracking index pointing from the target knowledge to the source knowledge, while binding each index entry with an permission verification identifier and a source anchor point. During retrieval, all upstream and downstream associated standard knowledge can be quickly traversed based on any standard knowledge. A synchronous update mechanism for adding, deleting, and modifying relationships is set up, ultimately completing the construction of a relationship retrieval index that supports chain-like extended queries of retrieval results.
[0099] Specifically, when two knowledge objects belong to the same device object, the same maintenance event, the same work order process, the same fault link, or adjacent handling stages, a relationship is established between them. The resulting relationship retrieval index can support contextual expansion of subsequent retrieval results. For example, when a knowledge object is matched, the system can not only return the knowledge object itself, but also further expand along its relationship retrieval index to adjacent handling steps, related alarm records, corresponding work order records, historical fault records of the same device, and related procedure content, thereby avoiding retrieval results remaining at a single fragment level.
[0100] The introduction of relational retrieval indexes transforms the knowledge base of this application from a simple collection of discrete knowledge into a knowledge organization structure with contextual paths. This is particularly important for operational scenarios, as operational knowledge is often not isolated but exists in a chain-like manner of "phenomenon—troubleshooting—location—handling—review." Relational retrieval indexes provide a path foundation for subsequent dynamically enhanced retrieval.
[0101] After the field retrieval index, vector retrieval index, and relation retrieval index are all generated, each of these three types of indexes is used as the underlying retrieval index. This application constructs a unified knowledge representation and a hybrid index structure composed of field retrieval index, inverted index, vector retrieval index, and relation retrieval index. This enables the retrieval process to support both precise location under explicit constraints such as device number, alarm number, work order number, error code, and time range, and semantic recall under conditions of fault phenomenon description, empirical expression, colloquial expression, and synonymous expression. Compared to single-keyword retrieval methods, this invention can reduce the problem of missed detections caused by differences in terminology, non-standard expressions, or incomplete descriptions; compared to single semantic retrieval methods, this invention can reduce the problem of false recalls caused by weakened explicit object constraints, thereby simultaneously improving the accuracy and coverage of operation and maintenance knowledge retrieval.
[0102] Based on the identifier corresponding to each standard knowledge, each underlying retrieval index, standard knowledge set, and fused knowledge are uniformly associated and mixed and encapsulated to obtain a hybrid retrieval database. For example, by traversing all unique identifiers of standard knowledge and using these unique identifiers as a global unified primary key, corresponding retrieval index entries for field retrieval, vector retrieval, and relation retrieval are bound to each other, establishing a one-to-one mapping relationship between the three types of indexes and standard knowledge and fused knowledge. The original text content and structured attribute information of standard knowledge are uniformly associated and mounted under the same primary key with the multi-dimensional semantic fusion representation of fused knowledge, achieving the same source binding of knowledge ontology, fusion features, and multi-dimensional index data. The retrieval indexes for field retrieval, vector retrieval, and relation retrieval are integrated and encapsulated to obtain a hybrid index set. The access permission boundaries, original data source anchors, and relation extension call entry points of each knowledge object are synchronously entered into the hybrid index set, giving it integrated capabilities for index retrieval, permission verification, data traceability, and relation extension. The linkage call rules and incremental update mechanism of various indexes are uniformly configured to ensure the data consistency and synchronous update of multiple indexes, standard knowledge, and fused knowledge, ultimately completing the construction of a hybrid retrieval database that integrates precise field retrieval, keyword retrieval, semantic vector retrieval, and relation extension retrieval.
[0103] The three types of indexes mentioned above are uniformly encapsulated to form a hybrid index set. This hybrid index set not only records the searchable representation of the knowledge object itself, but also records the access permission boundaries, source anchors, and relation extension entry points of the knowledge object.
[0104] In practical implementation, hybrid index encapsulation must at least meet the following requirements: First, any index hit result can be traced back to the original knowledge object; Secondly, any index hit result can retrieve the corresponding source information and object association information; Third, every index hit result must undergo permission boundary verification to prevent knowledge from being incorrectly recalled from different security scopes; Fourth, any index hit result can retain the basic information required for subsequent sorting and enhanced retrieval.
[0105] First, the output knowledge objects are transformed into a unified computable knowledge representation. Then, indexes are built from three dimensions: precise field matching, semantic similarity recall, and contextual relationship expansion. Finally, the three types of indexes are encapsulated into a unified hybrid index structure. Specifically, the field retrieval index addresses precise constraints such as device number, alarm code, error code, and time range; the vector retrieval index addresses semantic recall under synonymous expressions, implicit fault descriptions, and empirical expressions; and the relationship retrieval index addresses contextual completion of search results and expansion of the handling chain. The combination of these three elements enables the unified knowledge base to possess both the precise positioning capabilities of traditional retrieval systems and the semantic understanding and chain-like knowledge expansion capabilities required for operational scenarios.
[0106] Therefore, the above process is not simply about building an index, but rather about constructing a hybrid retrieval infrastructure at the knowledge object level. After this step, subsequent steps, when performing dynamic enhanced retrieval, no longer face a loose, singular, and non-scalable knowledge storage structure, but are instead built on a unified knowledge representation and multi-channel index, thus providing stable support for high-quality retrieval, evidence expansion, and result enhancement.
[0107] In some embodiments, after integrating the sorted knowledge set and the extended knowledge set to obtain the associated knowledge set corresponding to the retrieved data, the method includes: obtaining retrieval feedback information corresponding to the associated knowledge set; determining a correction value corresponding to the weight for each relevance level among all predetermined relevance levels based on the retrieval feedback information; determining an updated weight based on the weight and the correction value; and normalizing the updated weights corresponding to all relevance levels to obtain a target weight corresponding to each relevance level.
[0108] In this embodiment, to continuously optimize the dynamically enhanced retrieval process during use, this application further introduces a retrieval feedback update mechanism. First, retrieval feedback information corresponding to the associated knowledge set is obtained. For example, the retrieval feedback information includes at least one of user click behavior, dwell behavior, secondary retrieval behavior, manual collection behavior, manual confirmation behavior, or result abandonment behavior. For each weight corresponding to a predetermined relevance level, a correction value is determined based on the retrieval feedback information. Specifically, the information corresponding to the relevance level in the retrieval feedback information is mapped to a corresponding positive incentive score or negative penalty score. Positive incentive scores are preset for positive interaction behaviors such as user clicks to retrieve knowledge, prolonged reading, manual collection, and manual confirmation of accepted results; negative penalty scores are preset for negative interaction behaviors such as user quickly swiping away, batch abandonment of results, and repeated secondary retrievals. Based on the positive incentive score or negative penalty score corresponding to each piece of information, the total score for that relevance level is calculated, and the total score is normalized within a range. The normalized value is then set as the correction value corresponding to the weight of that relevance level.
[0109] Based on the quantification of user interaction behaviors corresponding to each relevant dimension, correction values are generated. Positive behaviors are scored with bonuses, and negative behaviors are penalized with deductions. The total score for each dimension is calculated and then normalized to unify the numerical range, avoiding large fluctuations in scores. This generates standardized weight correction values. The actual user experience of operations and maintenance personnel is transformed into a calculable adjustment basis, accurately reflecting the actual value of each matching dimension. Normalization ensures smooth and controllable weight iteration, avoiding drastic parameter shifts caused by single feedback. This allows the entire set of search weights to autonomously adapt and optimize according to actual search scenarios, continuously improving the accuracy of knowledge ranking and the user experience of operations and maintenance searches.
[0110] Based on the weights and correction values, the updated weights are determined. The retrieval system updates the aforementioned dynamic weights in small steps according to feedback information. Let the weight of a certain relevance at the t-th retrieval be . The correction value obtained based on the feedback is The updated weights can then be expressed as: ,(twenty two), in, To update the step size, the value ranges from 0 to 1.
[0111] The updated weights corresponding to all relevance levels are normalized to obtain the target weight for each relevance level. The sum of all relevance level weights is then converged to a fixed standard value (e.g., 1), ensuring that the weights for field matching, semantic similarity, relation matching, and quality reliability are always within a reasonable numerical allocation system. Through this retrieval feedback update mechanism, the retrieval system can gradually learn the preferences of different types of queries for field matching, semantic matching, and relation expansion in the current operation and maintenance scenario. For example, in the device maintenance number retrieval scenario, the system will gradually increase the weight of precise field matching; in the fuzzy fault phenomenon retrieval scenario, the system will gradually increase the weights of semantics and relation expansion. This allows the dynamically enhanced retrieval to gradually evolve from "rule-driven" to "rule and feedback collaboratively driven," improving the retrieval adaptability under long-term use.
[0112] In another embodiment, the user input is first parsed to generate query objects that can be adapted to field retrieval indexes, vector retrieval indexes, and relation retrieval indexes, respectively. Initial candidate knowledge results are obtained through a multi-channel parallel retrieval method. Next, the candidate results are dynamically weighted and sorted based on four indicators: field matching, semantic matching, relational relevance, and knowledge quality. High-scoring seed knowledge is then selected as the starting point for traversal, and knowledge enhancement and expansion are performed based on knowledge relationship paths to supplement the fault handling context, complete the handling chain, and related historical information. Afterwards, all expanded knowledge data is integrated and fused, duplicate content is removed, invalid weakly related information is filtered, and results are reorganized. Simultaneously, user search feedback data is collected to continuously iterate and optimize various search weight parameters. Compared to traditional retrieval models that only perform a single recall and a single fixed ranking, this solution is a dynamically enhanced retrieval solution adapted to operational knowledge business scenarios. Its core purpose is not simply to expand the total amount of knowledge recalled, but rather to rely on a collaborative architecture of multi-channel recall, dynamic weighted ranking, knowledge relationship expansion, and adaptive feedback updates. This ensures that the final retrieval results accurately hit the explicit search elements proposed by the user, while also parsing the user's implicit semantic needs, completing the full business context, and providing traceable source documentation. Based on this processing flow, the integrated knowledge base built in this application can handle both routine daily retrieval needs and complex business scenarios in operations and maintenance, such as fault location, reuse of historical operations and maintenance experience, process retrospection, and multi-dimensional correlation analysis. This effectively improves the accuracy of knowledge base retrieval and reduces the overall time spent by operations and maintenance personnel in troubleshooting and handling faults.
[0113] In another embodiment, such as Figure 2 As shown, Figure 2This application presents an overall architecture diagram for dynamic context-enhanced retrieval using a hybrid retrieval database, applicable to a marketing information technology operations and maintenance scenario. The data sources to be processed include operational monitoring information from a monitoring platform, operational logs from a log platform, alarm records from an alarm platform, handling work orders from a work order system, device and system configuration records from a configuration management system, locally uploaded procedural documents and fault analysis materials, local notes generated by operations and maintenance personnel, and supplementary operations and maintenance information returned by third-party interfaces. These data sources are independent of each other, with different record formats, naming conventions, update frequencies, and content granularities, making them difficult to directly use for unified knowledge base construction and retrieval. Therefore, this application performs unified knowledge base construction and dynamic enhanced retrieval.
[0114] First, multi-source heterogeneous operation and maintenance data are uniformly accessed, standardized and compiled, and preprocessed into knowledge objects.
[0115] The raw data output by the monitoring platform includes the monitored object identifier, monitored item name, sampling time, status value, and trend information; the raw data output by the log platform includes timestamps, host identifiers, application identifiers, log levels, log text, and anomaly information; the raw data output by the alarm platform includes alarm number, alarm object, alarm level, trigger time, recovery time, and alarm description; the raw data output by the work order system includes work order number, fault description, handling process, handler, and closure conclusion; the raw data output by the configuration management system includes device name, system name, module name, version information, and association relationships; locally uploaded materials include procedural documents, operation and maintenance manuals, fault review materials, and operation instructions; local notes include experience summaries and temporary records formed by operation and maintenance personnel based on historical problems; third-party interface data includes supplementary status information and external collaboration records related to operation and maintenance events. The above data enters the processing flow of this invention through file upload, interface retrieval, message subscription, and manual submission, respectively.
[0116] For each original record entering the process, it is first uniformly encapsulated to generate a source object. The source object includes at least the original content, metadata, and original location information. This uniform encapsulation allows data from different platforms to enter the same subsequent processing chain, no longer relying on the original data organization structure of each source system. After uniform encapsulation, content parsing is performed on the source object. For structured data, the correspondence between fields and values is extracted; for semi-structured data, hierarchical relationships and status information are extracted; for unstructured text, the semantics of titles, body text, tables, and paragraphs are extracted; for log and alarm data, time, object, event, level, and upstream / downstream association information are extracted; for mixed text and image data, identifiable text and the text-image correspondence are extracted and associated with relevant operational events.
[0117] After parsing, the results are standardized and compiled. This involves standardizing the time representation across different platforms; normalizing the mapping of device names, system names, module names, alarm abbreviations, and historical aliases; merging records with duplicate content from different sources; resolving conflicting records based on source reliability, time relevance, and field completeness; and downgrading or removing records with many missing fields, unreliable sources, or significant timeliness deficiencies. A quality score is calculated for each record. When a record's quality score falls below a preset threshold, it is marked as a low-quality record and is removed, downgraded, or placed under review. After standardization, the same operational facts, previously inconsistent across different sources, are integrated into a unified representation. Following the standardization results, knowledge segmentation is performed. In this embodiment, procedure documents and operation and maintenance manuals are segmented according to title level, paragraph theme, and table ownership boundaries; logs and alarm records are segmented according to time window, event chain, and handling stage; work order records are segmented according to problem description, troubleshooting process, handling actions, and handling conclusions; for mixed text and image materials, the correspondence between text and images is maintained, and mutually explanatory content is not split. For adjacent segments with strong semantic connections, contextual overlap is preserved at segment boundaries to avoid separating problem descriptions and handling conclusions.
[0118] After segmentation, each segment is used to generate a knowledge object. Each knowledge object includes at least the original content representation, the standardized content representation, source anchor information, and object association and permission boundary information. Simultaneously, based on relationships between knowledge objects such as shared device, shared event, shared work order, sequential processing, and contextual connection, a set of relationships is established between the objects. Finally, the set of knowledge objects and their relationship set are output.
[0119] A knowledge representation is constructed and a hybrid index is generated for the knowledge object set. For each knowledge object, its standardized text content, attribute information, and contextual information are extracted. The standardized text content reflects the semantic subject of the knowledge object; the attribute information reflects the device, system, module, time, source, and permission boundaries corresponding to the knowledge object; and the contextual information reflects the sequential, event, and object relationships between the knowledge object and other knowledge objects. Then, the above three types of information are fused to generate a unified knowledge representation. Thus, the unified knowledge representation can not only express the text semantics but also preserve the object attributes and relational context, thereby providing a unified representation foundation for subsequent accurate retrieval, semantic recall, and relation expansion.
[0120] After the unified knowledge representation is constructed, the first step is to establish a structured attribute sub-index and a text inverted sub-index. The structured attribute sub-index primarily targets content with clear identifying meaning, such as device name, device number, system name, module name, work order number, alarm number, error code, timestamp, source identifier, and standardized terminology. The text inverted sub-index, on the other hand, performs term segmentation and index registration on the standardized text content within the knowledge objects, enabling any term to be quickly located in the set of knowledge objects containing that term. Simultaneously, a mapping index between the original terminology expressions before normalization and the standard terminology is also established to accommodate situations where maintenance personnel mix abbreviations, old names, colloquial expressions, and formal terms during retrieval.
[0121] To enable the field retrieval index to narrow down the candidate scope before semantic recall, a field matching judgment function is set up to determine whether the knowledge object meets the constraints of device, work order, alarm, error code, time range, or source range in the query object. In this way, when the search request is explicitly limited to a specific device, work order, or time range, the candidate knowledge object set can be obtained first through the field retrieval index before entering the semantic matching stage, thereby reducing retrieval overhead and improving hit accuracy.
[0122] After establishing structured attribute sub-indexes and text inverted sub-indexes, a vector retrieval index is built for the unified knowledge representation. This vector retrieval index supports the retrieval of knowledge objects with similar semantics but not entirely consistent expressions, and is particularly suitable for operation and maintenance scenarios where fault phenomenon descriptions are not standardized, problem expressions are incomplete, and there are semantic inconsistencies between log errors and experience summaries. Simultaneously, relationship retrieval indexes are built for relationships between knowledge objects based on the same device, the same operation and maintenance event, the same work order process, the same fault link, and adjacent handling stages. Finally, the field retrieval index, vector retrieval index, and relationship retrieval index are uniformly encapsulated to form a hybrid index set.
[0123] Dynamically enhanced retrieval is performed based on a hybrid index set. The retrieval requests input by operations and maintenance personnel can be queries containing explicit object constraints or natural language fault descriptions. For example, a retrieval request can include a device name, an alarm number, or a work order number, or it can be a natural language expression such as "How to handle frequent alarms from a certain application in the early morning?" or "What are the historical handling steps after a module reports an error?" Upon receiving a retrieval request, the system first parses the request, identifying the semantic query content, field constraint content, and enhanced intent content, and then constructs a unified query object. The field constraint content is used to limit the device, work order, alarm, error code, time range, or source range, while the enhanced intent content indicates whether it is necessary to perform related event extensions, historical case extensions, or handling path extensions.
[0124] After constructing the query object, multi-channel initial recall is performed by calling the field retrieval index, vector retrieval index, and relation retrieval index respectively. Specifically, the field retrieval index prioritizes candidate knowledge objects that meet the object scope and time range requirements based on the field constraints in the query object; the vector retrieval index recalls candidate knowledge objects with similar semantics but inconsistent expressions based on the semantic content of the query object; and the relation retrieval index pre-positions related knowledge objects that may need to be expanded based on the enhanced intent content. Then, the three types of recall results are combined to form an initial candidate set.
[0125] For each candidate knowledge object in the initial candidate set, its comprehensive relevance score is calculated. The comprehensive relevance score considers field matching degree, semantic matching degree, relational relevance degree, and the quality of the knowledge object itself. Through this method, different types of query requests can obtain ranking results adapted to their characteristics. For example, when the search request contains explicit device numbers, work order numbers, alarm numbers, or time ranges, the weight of field matching items is increased; when the search request mainly presents a natural language fault description, the weight of semantic matching items is increased; when the search request explicitly includes extended expressions such as "related processing steps," "similar cases," or "related alarms," the weight of relational relevance items is increased.
[0126] After initial sorting, seed knowledge objects with high scores are selected from the candidate set as the starting point for enhancement and expansion. The system determines whether to include adjacent knowledge objects in the enhancement result set based on the strength of the relationship between the seed knowledge object and its adjacent knowledge objects, the degree of contextual continuity, and the degree of supplementary contribution to the current query. When the candidate expansion strength exceeds a preset expansion threshold, the corresponding adjacent knowledge object is included in the enhancement result set. During the expansion process, priority is given to retaining phenomenon records and handling records under the same fault event, historical cases under the same equipment object, the preceding and following steps in the same work order process, the explanatory relationship between alarm records and procedure clauses, and the conclusion relationship between problem descriptions and debriefing summaries.
[0127] After the enhancement and expansion are completed, the initial candidate results and the enhanced and expanded results are uniformly merged, and deduplication, noise reduction, and conflict resolution are performed. For multiple knowledge objects from different sources but pointing to the same operational fact, results with higher field completeness, more reliable sources, closer time to the target event, and better matching the current query are prioritized for retention; weakly related results that are only semantically similar but lack object or disposal associations are demoted in ranking; for conflicting knowledge results, the resolution is based on knowledge object quality, source credibility, and update time. Finally, the output is a set of search results containing knowledge object content, source anchors, related object information, relationship expansion paths, and recommended reading order.
[0128] To ensure continuous optimization of dynamically enhanced retrieval during use, this embodiment further introduces a retrieval feedback update mechanism. The system collects user click behavior, dwell time, secondary search behavior, favorites behavior, manual confirmation behavior, and result abandonment behavior during the retrieval process, and adaptively updates field matching weights, semantic matching weights, relation expansion weights, and ranking strategies based on the feedback. Therefore, as usage continues, this invention can gradually adapt to retrieval preferences under different operational scenarios, improving retrieval adaptability during long-term operation.
[0129] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.
[0130] It should be noted that some embodiments of this application have been described above. In some cases, the actions or steps described in the above embodiments can be performed in a different order than that shown in the above embodiments and the desired result can still be achieved. In addition, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0131] Based on the same inventive concept, corresponding to any of the above-described embodiments, this application also provides a data retrieval device.
[0132] refer to Figure 3 The data retrieval device includes: Extraction module 10 is configured to, in response to receiving retrieval data, perform multi-dimensional extraction on the retrieval data to obtain an extracted data set corresponding to the retrieval data.
[0133] The first retrieval module 20 is configured to use a pre-built hybrid retrieval database to perform matching retrieval on the extracted data set to obtain a retrieval knowledge set.
[0134] The sorting module 30 is configured to sort each retrieval knowledge in the retrieval knowledge set according to its relevance based on the extracted data set, thereby obtaining a sorted knowledge set.
[0135] The second retrieval module 40 is configured to use the hybrid retrieval database to perform relation expansion retrieval on the sorted knowledge set to obtain an expanded knowledge set.
[0136] The integration module 50 is configured to integrate the sorted knowledge set and the extended knowledge set to obtain the associated knowledge set corresponding to the retrieved data.
[0137] Through the aforementioned device, in response to received retrieval data, multi-dimensional retrieval is performed on the retrieval data to obtain an extracted data set corresponding to the retrieval data, achieving the purpose of comprehensive analysis of the retrieval data. Using a pre-constructed hybrid retrieval database, the extracted data set is matched and retrieved to obtain a retrieval knowledge set, expanding the data participating in the matching retrieval and enriching the retrieval knowledge. Based on the extracted data set, each retrieval knowledge in the retrieval knowledge set is ranked by relevance to obtain a ranked knowledge set. The matching tightness is quantified according to the actual retrieval needs of the user, prioritizing the display of knowledge most closely matching the retrieval requirements. Using the hybrid retrieval database, a relational expansion retrieval is performed on the ranked knowledge set to obtain an expanded knowledge set. The ranked knowledge set and the expanded knowledge set are then integrated to obtain a related knowledge set corresponding to the retrieval data, achieving the purpose of comprehensively supplementing the ranked knowledge set.
[0138] In some embodiments, the multi-dimensional retrieval extraction includes text semantic extraction, structured attribute extraction, and data expansion extraction; the extraction module 10 is further configured to preprocess the retrieval data to obtain target retrieval data; perform text semantic extraction on the target retrieval data to obtain text semantic data; perform structured attribute extraction on the target retrieval data to obtain attribute data; perform data expansion extraction on the target retrieval data to obtain expanded data; and combine the text semantic data, the attribute data, and the expanded data to form the extracted data set.
[0139] In some embodiments, the first retrieval module 20 is further configured to, for each piece of data in the extracted data set, call the underlying retrieval index corresponding to the retrieval dimension in the hybrid retrieval database based on the retrieval dimension corresponding to the data; retrieve the data in the hybrid retrieval database according to the underlying retrieval index to obtain the initial knowledge corresponding to the data; obtain the retrieval permissions of the retrieval user corresponding to the retrieval data; based on the retrieval permissions, remove knowledge from the initial knowledge corresponding to each piece of data to obtain retrieval knowledge; and merge and deduplicate the retrieval knowledge corresponding to all data to obtain the retrieval knowledge set.
[0140] In some embodiments, the sorting module 30 is further configured to perform feature recognition on the extracted data set to obtain at least one feature; for each feature, determine the relevance level of the feature among all predetermined relevance levels; based on the relevance level of each feature, determine the weight corresponding to each of the all predetermined relevance levels; for each retrieval knowledge in the retrieval knowledge set, based on the weight corresponding to each relevance level, perform a weighted calculation of the relevance level of the retrieval knowledge and the data corresponding to the retrieval knowledge in the extracted data set to obtain a relevance score corresponding to the retrieval knowledge; and sort all retrieval knowledge according to the relevance score corresponding to each retrieval knowledge to obtain the sorted knowledge set.
[0141] In some embodiments, the sorted knowledge set includes multiple search knowledge items sorted from high to low relevance; the second search module 40 is further configured to take a predetermined number of search knowledge items from the sorted knowledge set in a forward-to-back order as multiple target search knowledge items; for each target search knowledge item, in the hybrid search database, starting from the target search knowledge item, a neighborhood expansion search is performed along the relation search index in the hybrid search database to obtain the extended knowledge item corresponding to the target search knowledge item; the extended knowledge items corresponding to the multiple target search knowledge items constitute the extended knowledge set.
[0142] In some embodiments, the integration module 50 is further configured to merge the sorted knowledge set and the extended knowledge set to obtain a first knowledge set; perform deduplication, weak correlation filtering, and conflict resolution on the first knowledge set to obtain a second knowledge set; reorder the second knowledge set according to relevance to obtain a third knowledge set; for each knowledge in the third knowledge set, obtain the corresponding associated information and bind the associated information to the knowledge, wherein the associated information includes at least the source anchor of the knowledge; and in response to determining that all knowledge has been bound with associated information, obtain the associated knowledge set.
[0143] In some embodiments, the system further includes a construction module configured to: acquire multi-source heterogeneous raw data; perform standardization processing on the raw data to obtain multiple initial standard knowledge sets; extract relationships from the multiple initial standard knowledge sets to obtain an initial set of association relationships; add an identifier to each initial standard knowledge set to obtain standard knowledge; construct a standard knowledge set from all standard knowledge sets; perform knowledge replacement on the original set of association relationships based on all standard knowledge sets to obtain a standard relationship set; extract and weight fused multi-feature dimension knowledge from each standard knowledge set to obtain fused knowledge; construct a field retrieval index based on the standard knowledge set; construct a vector retrieval index based on all fused knowledge sets; and construct a relationship retrieval index based on the standard relationship set, using the field retrieval index, the vector retrieval index, and the relationship retrieval index as underlying retrieval indexes; and uniformly associate and encapsulate each underlying retrieval index, the standard knowledge set, and the fused knowledge based on the identifier corresponding to each standard knowledge set to obtain the hybrid retrieval database.
[0144] In some embodiments, after integrating the sorted knowledge set and the extended knowledge set to obtain the associated knowledge set corresponding to the retrieved data, an update module is further included. The update module is configured to: obtain retrieval feedback information corresponding to the associated knowledge set; determine a correction value corresponding to the weight for each relevance level among all predetermined relevance levels based on the retrieval feedback information; determine the updated weight based on the weight and the correction value; and normalize the updated weights corresponding to all relevance levels to obtain the target weight corresponding to each relevance level.
[0145] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.
[0146] The apparatus described above is used to implement the corresponding data retrieval method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.
[0147] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the data retrieval method as described in any of the above embodiments.
[0148] Figure 4This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.
[0149] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.
[0150] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.
[0151] The input / output interface 1030 is used to connect input / output modules to realize information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.
[0152] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).
[0153] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.
[0154] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.
[0155] The electronic devices described above are used to implement the corresponding data retrieval methods in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.
[0156] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the data retrieval method as described in any of the above embodiments.
[0157] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.
[0158] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the data retrieval method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.
[0159] Based on the same concept, corresponding to the methods of any of the above embodiments, this application also provides a computer program product, including computer program instructions. When the computer program instructions are run on a computer, the computer causes the computer to execute the data retrieval method as described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.
[0160] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.
[0161] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this disclosed technical solution.
[0162] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.
[0163] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.
[0164] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in detail for the sake of brevity.
[0165] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.
[0166] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.
[0167] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.
Claims
1. A data retrieval method, characterized in that, include: In response to receiving retrieval data, the retrieval data is extracted from multiple retrieval dimensions to obtain the extracted data set corresponding to the retrieval data. Using a pre-built hybrid retrieval database, the extracted data set is matched and retrieved to obtain a retrieval knowledge set; Based on the extracted data set, each retrieval knowledge in the retrieval knowledge set is sorted by relevance to obtain a sorted knowledge set; Using the hybrid retrieval database, a relation expansion retrieval is performed on the sorted knowledge set to obtain an expanded knowledge set; The sorted knowledge set and the extended knowledge set are integrated to obtain the associated knowledge set corresponding to the retrieved data.
2. The method according to claim 1, characterized in that, The multi-dimensional retrieval extraction includes text semantic extraction, structured attribute extraction, and data expansion extraction; The step of extracting data from the search data using multiple search dimensions to obtain the extracted data set corresponding to the search data includes: The search data is preprocessed to obtain the target search data; Text semantic extraction is performed on the target retrieval data to obtain text semantic data; Structured attribute extraction is performed on the target retrieval data to obtain attribute data; The target retrieval data is expanded and extracted to obtain expanded data; The extracted data set is composed of the text semantic data, the attribute data, and the extended data.
3. The method according to claim 1, characterized in that, The step involves using a pre-built hybrid retrieval database to perform matching retrieval on the extracted data set to obtain a retrieval knowledge set, including: For each piece of data in the extracted data set, based on the search dimension corresponding to the data, the underlying search index corresponding to the search dimension is called in the hybrid search database; According to the underlying retrieval index, the data is retrieved in the hybrid retrieval database to obtain the initial knowledge corresponding to the data; Obtain the retrieval permissions of the retrieval user corresponding to the retrieval data, and based on the retrieval permissions, perform knowledge removal on the initial knowledge corresponding to each data to obtain the retrieval knowledge; The retrieval knowledge corresponding to all data is merged and deduplicated to obtain the retrieval knowledge set.
4. The method according to claim 1, characterized in that, Based on the extracted data set, each piece of retrieval knowledge in the retrieval knowledge set is ranked according to its relevance to obtain a ranked knowledge set, including: The extracted data set is subjected to feature recognition to obtain at least one feature; For each feature, determine the degree of relevance of that feature among all predetermined degrees of relevance; Based on the relevance of each feature, determine the weight corresponding to each relevance among all predetermined relevance levels; For each piece of retrieval knowledge in the retrieval knowledge set, based on the weight corresponding to each relevance level, a weighted calculation of the relevance level of the retrieval knowledge and the data corresponding to the retrieval knowledge in the extracted data set is performed to obtain the relevance score corresponding to the retrieval knowledge; All retrieved knowledge is sorted according to the relevance score corresponding to each retrieved knowledge, resulting in the sorted knowledge set.
5. The method according to claim 1, characterized in that, The sorted knowledge set includes multiple retrieval knowledge items sorted from high to low according to their relevance; The step of using the hybrid retrieval database to perform relation expansion retrieval on the ranked knowledge set to obtain an expanded knowledge set includes: A predetermined number of retrieval knowledge items are selected from the sorted knowledge set in a forward-to-back order as multiple target retrieval knowledge items; For each target retrieval knowledge, in the hybrid retrieval database, starting from the target retrieval knowledge, a neighborhood expansion retrieval is performed along the relation retrieval index in the hybrid retrieval database to obtain the extended knowledge corresponding to the target retrieval knowledge; The extended knowledge corresponding to the multiple target retrieval knowledge is used to form the extended knowledge set.
6. The method according to claim 1, characterized in that, The step of integrating the sorted knowledge set and the extended knowledge set to obtain the associated knowledge set corresponding to the retrieved data includes: The ordered knowledge set and the extended knowledge set are merged to obtain the first knowledge set; The first knowledge set is subjected to deduplication, weak correlation filtering and conflict resolution to obtain the second knowledge set; The second knowledge set is reordered according to relevance to obtain the third knowledge set; For each piece of knowledge in the third knowledge set, obtain the corresponding associated information for the knowledge, and bind the associated information to the knowledge, wherein the associated information includes at least the source anchor of the knowledge; In response to the determination that all knowledge has been bound to the corresponding associated information, the associated knowledge set is obtained.
7. The method according to claim 1, characterized in that, The method for constructing the hybrid retrieval database includes: Obtain multi-source heterogeneous raw data, perform standardization processing on the raw data to obtain multiple initial standard knowledge, and extract relationships from the multiple initial standard knowledge to obtain an original set of association relationships; For each initial standard knowledge, an identifier is added to the initial standard knowledge to obtain the standard knowledge; All standard knowledge is used to form a standard knowledge set, and based on all standard knowledge, the original set of relationships is replaced with knowledge to obtain a standard relationship set; For each piece of standard knowledge in the standard knowledge set, multi-feature dimension knowledge is extracted and weighted and fused to obtain fused knowledge; A field retrieval index is constructed based on the standard knowledge set, a vector retrieval index is constructed based on all fused knowledge, and a relationship retrieval index is constructed based on the standard relationship set. The field retrieval index, the vector retrieval index, and the relationship retrieval index are used as the underlying retrieval indexes, respectively. Based on the identifier corresponding to each standard knowledge, each underlying retrieval index, the standard knowledge set, and the fused knowledge are uniformly associated and mixed and encapsulated to obtain the hybrid retrieval database.
8. The method according to claim 1, characterized in that, After integrating the sorted knowledge set and the extended knowledge set to obtain the associated knowledge set corresponding to the retrieved data, the method includes: Obtain the retrieval feedback information corresponding to the associated knowledge set; For each relevance level among all predetermined relevance levels, a correction value corresponding to the weight is determined based on the search feedback information; Based on the weights and the correction values, the updated weights are determined; Normalize the updated weights corresponding to all relevance levels to obtain the target weight for each relevance level.
9. A data retrieval device, characterized in that, include: The extraction module is configured to, in response to receiving search data, perform multi-dimensional extraction on the search data to obtain an extracted data set corresponding to the search data. The first retrieval module is configured to use a pre-built hybrid retrieval database to perform matching retrieval on the extracted data set to obtain a retrieval knowledge set; The sorting module is configured to sort each retrieval knowledge in the retrieval knowledge set according to its relevance based on the extracted data set, thereby obtaining a sorted knowledge set. The second retrieval module is configured to use the hybrid retrieval database to perform relation expansion retrieval on the sorted knowledge set to obtain an expanded knowledge set. The integration module is configured to integrate the sorted knowledge set and the extended knowledge set to obtain the associated knowledge set corresponding to the retrieved data.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.
