Device knowledge base intelligent retrieval method based on large model perception and dynamic index optimization

Abstract

The invention relates to an equipment knowledge base intelligent retrieval method based on large model perception and dynamic index optimization, and aims at solving the problems that an existing equipment knowledge base is low in retrieval precision, poor in semantic understanding and low in response speed. The method comprises the following steps: performing deep semantic understanding and structural analysis on an equipment knowledge base by utilizing a large language model, and constructing a knowledge graph containing entities, attributes and relationships; constructing and dynamically updating a multi-level mixed index structure based on the graph in combination with query load prediction; and receiving a natural language query of a user, performing intention recognition and semantic extension through the large model, and performing efficient retrieval and accurate matching through the index. According to the method, deep semantic understanding and efficient retrieval of complicated and heterogeneous equipment knowledge are realized, and the knowledge acquisition efficiency and accuracy in scenes of equipment operation and maintenance, fault diagnosis and the like are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and information retrieval technology, and in particular to an intelligent retrieval method for device knowledge bases that combines large language model knowledge perception and dynamic index optimization technology. Background Technology

[0002] With the deepening of industrial digital transformation, complex and equipment-intensive fields such as power and manufacturing have accumulated massive amounts of equipment knowledge resources, including diverse and heterogeneous data such as technical specifications, operation and maintenance manuals, fault records, and test reports. These knowledge resources collectively constitute the core equipment knowledge base of an enterprise, and their effective utilization is of vital importance for equipment operation and maintenance, fault diagnosis, and expert decision-making.

[0003] Currently, equipment knowledge base retrieval mainly relies on traditional keyword matching techniques. While these methods are simple to implement, they have significant limitations in practical applications: First, literal matching-based retrieval mechanisms cannot understand the deep semantic intent of user queries; for example, they struggle to return valid results for complex queries such as "possible causes of abnormal vibration." Second, traditional methods lack the ability to perceive professional knowledge structures and cannot uncover the entity relationships and logical connections hidden within the text, resulting in fragmented and incomplete retrieval results. Furthermore, as the scale of the knowledge base continues to expand, a single index structure cannot simultaneously meet the multi-dimensional needs of semantic retrieval, relational queries, and exact matching, leading to low retrieval efficiency and significant response delays.

[0004] In recent years, although some studies have attempted to introduce technologies such as knowledge graphs or vector retrieval, these methods also have new problems: knowledge graph construction relies on a large amount of manual annotation, which is costly and difficult to scale; while vector retrieval has a certain semantic understanding capability, it lacks effective utilization of structured knowledge. At the same time, the index structure of existing retrieval systems is mostly statically constructed, which cannot adaptively optimize according to changes in user query patterns, and performs poorly when faced with dynamically evolving knowledge bases and diverse user needs.

[0005] Especially in specialized fields such as power equipment, there are unique challenges, including understanding professional terminology and adapting to industry standards. Traditional methods cannot accurately grasp the connotation and extension of professional concepts such as "CT saturation characteristics" and "circuit breaker failure to operate," let alone establish a deep connection between equipment failures and handling measures. This deficiency in knowledge retrieval directly affects the efficiency and accuracy of equipment operation and maintenance, and may even lead to safety accidents due to incomplete information acquisition. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides an intelligent retrieval method for device knowledge base based on large model perception and dynamic index optimization, which solves the problems of insufficient semantic understanding, low retrieval efficiency and poor system adaptability in the existing retrieval methods.

[0007] The technical solution of the present invention is as follows:

[0008] A device knowledge base intelligent retrieval method based on large model perception and dynamic index optimization includes the following steps:

[0009] By using a domain-adaptive large language model, we can perform deep semantic analysis on the device knowledge base, extract entities, relationships and attributes, and construct a structured device knowledge graph.

[0010] Based on the constructed device knowledge graph, historical query load is analyzed, a multi-level hybrid index including semantic vector index, graph structure index and keyword index is constructed, and a dynamic update mechanism for the index is established.

[0011] It receives users' natural language queries, uses a large language model for deep intent recognition and semantic expansion, and transforms vague user needs into structured, machine-understandable query vectors and constraints.

[0012] Using the query vector and constraints, collaborative retrieval and matching are performed in the constructed multi-level hybrid index, and the retrieval results are fused and intelligently sorted to return the most relevant knowledge entries.

[0013] Optionally, the step of constructing a structured equipment knowledge graph involves: incrementally pre-training and fine-tuning a general-purpose large language model using professional text corpus in the power equipment field to obtain a domain-adaptive model, Model_domain; guiding the model to extract standardized knowledge triples T = (h, r, t) from the equipment knowledge base text using a preset prompt template, where h represents the head entity, r represents the relation, and t represents the tail entity or attribute value; and performing entity alignment and conflict resolution on all extracted knowledge triples to construct a unified power equipment knowledge graph G = (V, E), where V is the entity set and E is the relation set.

[0014] Optionally, the step of constructing a multi-level hybrid index involves using an embedding model to map entity and relation description texts in the device knowledge graph into high-dimensional vectors and constructing an approximate nearest neighbor index based on HNSW; performing graph summarization on the device knowledge graph and constructing a hierarchical index structure; and simultaneously constructing an inverted index to cover term matching and literal retrieval requirements.

[0015] Optionally, the steps for establishing a dynamic index update mechanism include: optimizing the index structure based on query frequency and performance metrics; periodically analyzing query logs; and adjusting the vector distribution of entities in the semantic index or the summary granularity of the graph structure index when the query frequency freq(e) for a specific entity or relation exceeds a threshold θ or the query pattern changes. The objective function of the optimization process is to minimize the average query latency L.

[0016]

[0017] Among them, t query N represents the time taken for a single query, and N represents the total number of queries.

[0018] Optionally, in the query vector and constraint generation step, the user query input domain is adapted to the model, and its intent category is identified by a softmax classifier; an extended query item set is generated based on the identified intent, and the extended items are filtered by calculating concept similarity; the extended query is converted into a vector representation, and structured constraints for the graph query are generated.

[0019] Optionally, in the collaborative retrieval and matching step, semantic retrieval, graph retrieval, and keyword retrieval are performed in parallel in a multi-level hybrid index based on query vectors and structured constraints; multi-dimensional feature extraction is performed on the retrieval results to construct feature vectors.

[0020] Alternatively, the step of fusing and intelligently sorting the search results involves using a pre-trained sorting model to reorder the candidate results and returning the Top-K results in descending order of score.

[0021] Optionally, the ranking model is trained based on the LambdaMART algorithm, with training data derived from historical click logs and expert annotations. The objective function is to maximize the NDCG@10 metric. It also includes a result diversity guarantee mechanism, which uses a maximum marginal relevance algorithm to deduplicate similar results.

[0022]

[0023] Where S is the set of selected results, λ is the diversity weight parameter, and sim(e,e) j ) represents entities e and e j The similarity is calculated using cosine similarity:

[0024]

[0025] The beneficial effects of this invention are as follows: This invention achieves deep semantic understanding and structured construction of equipment knowledge through a domain-adapted large language model, transforming unstructured text into a knowledge graph rich in semantic relationships, laying the foundation for accurate retrieval. By constructing a multi-level hybrid index adapted to query load and introducing a dynamic optimization mechanism, millisecond-level response is achieved under massive data. This method organically combines the deep semantic understanding capabilities of a large model with efficient index retrieval technology, achieving a leap from "keyword matching" to "semantic and relational understanding," significantly improving the accuracy, completeness, and efficiency of equipment knowledge retrieval. In particular, this method effectively solves the problems of terminology understanding and semantic gaps in professional fields through query intent recognition, semantic expansion, and multi-path collaborative retrieval mechanisms; through dynamic index updates and learning ranking strategies, it ensures the system's adaptability under different query modes and environmental changes. This invention can be widely applied to power equipment operation and maintenance, fault diagnosis, and expert decision-making systems, providing strong knowledge management support for the digital transformation of equipment-intensive industries. Attached Figure Description

[0026] Figure 1 The flowchart shows the intelligent retrieval method for device knowledge base based on large model perception and dynamic index optimization provided by this invention. Detailed Implementation

[0027] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0028] Example

[0029] like Figure 1 As shown, the intelligent retrieval method for device knowledge base based on large model perception and dynamic index optimization provided by this invention includes the following steps:

[0030] S1: Knowledge Structure Perception and Graph Construction: Utilize a domain-adapted large language model to perform deep semantic analysis on the device knowledge base, automatically extract entities, relationships, and attributes, and construct a structured device knowledge graph.

[0031] Incremental pre-training and instruction fine-tuning of a general large language model were performed using specialized text corpus in the field of power equipment to obtain a domain-adapted model, Model_domain. The specialized text corpus includes equipment technical specifications, operation and maintenance procedures, fault analysis reports, and equipment nameplate parameters.

[0032] Unstructured text from the equipment knowledge base is input into the domain adaptation model. The model is guided by a preset prompt template to identify and extract standardized knowledge triples T = (h, r, t), where h represents the head entity, r represents the relation, and t represents the tail entity or attribute value. The entities include specific equipment models of transformers, circuit breakers, and instrument transformers, and the relations include "has a fault", "maintenance method", and "technical parameters".

[0033] Entity alignment and conflict resolution are performed on all extracted knowledge triples to construct a unified power equipment knowledge graph G = (V, E), where V is the set of entities and E is the set of relations. Specifically, entity similarity calculation based on edit distance and conflict detection method based on knowledge graph embedding are used to ensure the consistency of entities and the reliability of relations in the graph.

[0034] S2: Multi-level hybrid index construction and optimization: Based on the knowledge graph constructed in step S1, analyze the historical query load, construct a multi-level hybrid index structure that includes semantic vector index, graph structure index and keyword index, and establish a dynamic update mechanism for the index.

[0035] An embedding model is used to map entity and relation description texts in the knowledge graph into high-dimensional vectors, and an approximate nearest neighbor index based on HNSW is constructed. Graph summarization is performed on the knowledge graph G to construct a hierarchical index structure to support efficient multi-hop relation queries. At the same time, an inverted index is constructed to cover the needs of accurate term matching and fast literal retrieval.

[0036] The dynamic update strategy for the index structure is as follows: Periodically analyze the query logs. When the query frequency (freq(e)) for a specific entity or relation exceeds a threshold θ, or when the query pattern changes, automatically adjust the vector distribution of that entity in the semantic index or the summary granularity of the graph structure index. The objective function of the optimization process is to minimize the average query latency L.

[0037]

[0038] Among them, t query N represents the time taken for a single query, and N represents the total number of queries.

[0039] S3: Query Semantic Understanding and Expansion: Receives user natural language queries, uses a large language model for deep intent recognition and semantic expansion, and transforms ambiguous user needs into structured, machine-understandable query vectors and constraints.

[0040] Query intent recognition and semantic expansion specifically involve inputting the user query Q into the domain adaptation model M. domain The intent category I is identified using a softmax classifier.

[0041] I = argmax(Wi ·h [CLS] +b i )

[0042] Among them, h [CLS] For the hidden layer state marked [CLS], W i and b i These are the parameters for the classification layer.

[0043] Based on the identified intent, expanded query terms Q′={q1,q2,…,q3} are generated, and expanded terms are filtered by calculating concept similarity:

[0044] sim(p,Q)=cos(E(q),E(Q))>θ

[0045] Where E(·) is the text embedding function, and θ is the similarity threshold;

[0046] Query vectorization: Converting the expanded query into a vector representation.

[0047]

[0048] Based on the intent and entity recognition results, structured constraints C for the graph query are generated, including entity type constraints T. e and relational path constraints P r Structured constraints are automatically filled using SPARQL templates, and entity recognition is achieved through sequence labeling.

[0049] y t =softmax(W y h t +b y )

[0050] Where h t Let y be the hidden state of the t-th word. t For entity tags.

[0051] S4: Collaborative retrieval and result ranking: Using the query vector and constraints generated in step S3, collaborative retrieval and matching are performed in the multi-level hybrid index constructed in step S2, and the retrieval results are fused and intelligently ranked to return the most relevant knowledge items.

[0052] Based on query vector v Q With structured constraint C, the following retrieval operations are performed in parallel within the multi-level hybrid index:

[0053] Perform an approximate nearest neighbor search in the semantic vector index to obtain a set R of similar entities. s =e|cos(v e ,v Q>δ; Perform subgraph matching in the graph structure index to obtain the relevant entity set R. g =SubgraphMatch(C) performs a Boolean search on the keyword index to obtain the set of matching documents R. k .

[0054] Multi-dimensional feature extraction is performed on the search results to construct a feature vector:

[0055] f(e) = [cos(v)] e ,v Q ),PageRank(e),Recency(e),Authority(e)]

[0056] Re-rank the candidate results using a pre-trained ranking model:

[0057] score(e) = σ(W) T f(e)+b)

[0058] Where σ is the sigmoid function, which returns the top-K results in descending order of score.

[0059] The semantic retrieval similarity threshold δ = 0.6, and the HNSW algorithm is used for approximate nearest neighbor search. The ranking model is trained based on the LambdaMART algorithm, with training data from historical click logs and expert annotations. The objective function is to maximize the NDCG@10 index. The method also includes a result diversity guarantee mechanism, which uses the maximum marginal relevance algorithm to deduplicate similar results.

[0060]

[0061] Where S is the set of selected results, λ is the diversity weight parameter, and sim(e,e) j ) represents entities e and e j The similarity is calculated using cosine similarity:

[0062]

[0063] This invention also provides a storage medium storing a computer program. When executed by a processor, the computer program implements some or all of the steps in the various embodiments of the intelligent retrieval method for device knowledge base based on large model perception and dynamic index optimization provided by this invention. The storage medium may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0064] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent retrieval of power equipment knowledge base based on large model perception and dynamic index optimization, characterized in that, Includes the following steps: By using a domain-adaptive large language model, we can perform deep semantic analysis on the device knowledge base, extract entities, relationships and attributes, and construct a structured device knowledge graph. Based on the constructed device knowledge graph, historical query load is analyzed, a multi-level hybrid index including semantic vector index, graph structure index and keyword index is constructed, and a dynamic update mechanism for the index is established. It receives users' natural language queries, uses a large language model for deep intent recognition and semantic expansion, and transforms vague user needs into structured, machine-understandable query vectors and constraints. Using the query vector and constraints, collaborative retrieval and matching are performed in the constructed multi-level hybrid index, and the retrieval results are fused and intelligently sorted to return the most relevant knowledge entries.

2. The intelligent retrieval method for power equipment knowledge base based on large model perception and dynamic index optimization according to claim 1, characterized in that, The steps for constructing a structured equipment knowledge graph are as follows: First, incremental pre-training and fine-tuning of a general-purpose large language model using specialized text corpus from the power equipment domain yields a domain-adapted model, Model_domain. Then, the model is guided by a pre-defined prompt template to extract standardized knowledge triples T = (h, r, t) from the equipment knowledge base text, where h represents the head entity, r represents the relation, and t represents the tail entity or attribute value. Finally, entity alignment and conflict resolution are performed on all extracted knowledge triples to construct a unified power equipment knowledge graph G = (V, E), where V is the entity set and E is the relation set.

3. The intelligent retrieval method for power equipment knowledge base based on large model perception and dynamic index optimization according to claim 1, characterized in that, The steps for constructing a multi-level hybrid index are as follows: using an embedding model to map the entity and relation description text in the device knowledge graph into high-dimensional vectors, and constructing an approximate nearest neighbor index based on HNSW; performing graph summarization on the device knowledge graph and constructing a hierarchical index structure; and simultaneously constructing an inverted index to cover term matching and literal retrieval requirements.

4. The intelligent retrieval method for power equipment knowledge base based on large model perception and dynamic index optimization according to claim 3, characterized in that, The steps of establishing the dynamic index update mechanism are as follows: Optimize the index structure based on query frequency and performance metrics; periodically analyze query logs, and when the query frequency freq(e) for a specific entity or relationship exceeds the threshold θ or the query pattern changes, adjust the vector distribution of that entity in the semantic index or the summary granularity of the graph structure index. The objective function of the optimization process is to minimize the average query latency L. Among them, t query N represents the time taken for a single query, and N represents the total number of queries.

5. The intelligent retrieval method for power equipment knowledge base based on large model perception and dynamic index optimization according to claim 1, characterized in that, In the query vector and constraint generation step, the user query input domain is adapted to the model, and the user's intent category is identified by a softmax classifier. An expanded query item set is generated based on the identified intent, and the expanded items are filtered by calculating concept similarity. The expanded query is converted into a vector representation, and structured constraints for the graph query are generated.

6. The intelligent retrieval method for power equipment knowledge base based on large model perception and dynamic index optimization according to claim 1, characterized in that, In the collaborative retrieval and matching step, semantic retrieval, graph retrieval, and keyword retrieval are performed in parallel in a multi-level hybrid index based on query vectors and structured constraints; multi-dimensional feature extraction is performed on the retrieval results to construct feature vectors.

7. The intelligent retrieval method for power equipment knowledge base based on large model perception and dynamic index optimization according to claim 6, characterized in that, The steps for fusing and intelligently ranking the search results are as follows: a pre-trained ranking model is used to re-rank the candidate results, and the Top-K results are returned in descending order of score.

8. The intelligent retrieval method for power equipment knowledge base based on large model perception and dynamic index optimization according to claim 7, characterized in that, The ranking model is trained based on the LambdaMART algorithm, with training data derived from historical click logs and expert annotations. The objective function is to maximize the NDCG@10 metric. It also includes a result diversity guarantee mechanism, which uses a maximum marginal relevance algorithm to deduplicate similar results. Where S is the set of selected results, λ is the diversity weight parameter, and sim(e,e) j ) represents entities e and e j The similarity is calculated using cosine similarity:

Images