Equipment operation and maintenance intelligent management method and system based on artificial intelligence
By using parallel sparse and dense retrieval paths and adaptive weight generation, the problem of low efficiency in traditional operation and maintenance models is solved, and the efficiency and accuracy of intelligent management of equipment operation and maintenance are improved.
Patent Information
- Application Number
- CN202511648711.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-01-27
AI Technical Summary
In the operation and maintenance of existing equipment, the traditional operation and maintenance mode relies on human experience, which is inefficient and difficult to deal with complex faults. When integrating sparse and dense retrieval, the existing intelligent question answering system cannot dynamically adjust the weight according to the query characteristics, resulting in insufficient recall or decreased accuracy.
An artificial intelligence-based approach is adopted to obtain user queries through parallel sparse and dense retrieval paths, adaptive weight generation is performed, and sparse and dense retrieval results are dynamically fused to generate the final answer.
It enables dynamic adjustment of weights based on query characteristics, improving the accuracy and reliability of search results, ensuring a high degree of relevance between knowledge sources and query intent, and enhancing the efficiency and accuracy of intelligent operation and maintenance management.
Smart Images

Figure CN121412397A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of deep learning, and more specifically, to an intelligent management method and system for equipment operation and maintenance based on artificial intelligence. Background Technology
[0002] With the deepening of digital and intelligent transformation across industries, equipment systems in modern industry, energy, information technology, and other fields are becoming increasingly complex. Their operation and maintenance generate massive amounts of multi-source, heterogeneous data, including operation logs, monitoring indicators, and alarm events. Traditional operation and maintenance models heavily rely on manual experience, requiring maintenance personnel to manually filter and analyze massive amounts of information. This is not only inefficient but also prone to response delays and misjudgments when facing sudden or unknown complex faults, making it difficult to ensure the continuous and stable operation of equipment and business systems.
[0003] To address this challenge, the industry has begun exploring the application of artificial intelligence technologies, particularly intelligent question-answering systems that combine retrieval and generation, to assist equipment operation and maintenance. These systems typically receive natural language queries from maintenance personnel and search for answers from vast knowledge bases such as technical manuals, historical fault cases, and maintenance work orders. Hybrid retrieval technology, which combines sparse retrieval (such as the BM25 algorithm) and dense retrieval (such as vector similarity search), is widely adopted as a mainstream solution to improve retrieval performance. However, existing technologies typically employ a static weighted fusion strategy when fusing the results of these two retrieval paths, pre-setting a fixed fusion weight for sparse and dense retrieval scores. This one-size-fits-all approach has a significant technical flaw: it ignores the inherent differences between user queries. For example, when a user query contains precise entity keywords, such as the alarm code "0xFA01" or a specific equipment model, sparse retrieval should dominate due to its accurate term frequency matching capability. Conversely, when a user query is a vague semantic description, such as abnormal equipment vibration, dense retrieval, which can understand the same semantic meaning behind different expressions such as abnormal vibration and excessive unit shaking, should be given more emphasis. Existing fixed-weight methods cannot dynamically adjust according to the characteristics of the query itself, leading to systematic problems of insufficient recall (the retrieved document was not found) or decreased precision (a large number of irrelevant documents were retrieved) when facing specific query types, seriously affecting the accuracy of the final answer.
[0004] Therefore, an optimized intelligent management solution for equipment operation and maintenance is desired. Summary of the Invention
[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide an intelligent management method and system for equipment operation and maintenance based on artificial intelligence.
[0006] According to one aspect of this application, an artificial intelligence-based intelligent management method for equipment operation and maintenance is provided, comprising: Retrieve user queries input by the user; The user query input is processed through parallel sparse and dense search paths to obtain sparse search results, dense search results, and query embedding vectors. Determine adaptive weights based on user queries and query embedding vectors; Based on adaptive weights, the scores of sparse and dense search results are standardized and dynamically fused to obtain the fused result. The fusion results are rearranged and candidates are selected to obtain Top-K candidate documents; The final answer is generated based on the Top-K candidate documents and user queries.
[0007] According to another aspect of this application, an artificial intelligence-based intelligent management system for equipment operation and maintenance is provided, comprising: The user query input acquisition module is used to acquire user queries input by the user. The parallel retrieval module is used to process user query inputs into parallel sparse and dense retrieval paths to obtain sparse retrieval results, dense retrieval results, and query embedding vectors. The weight generation module is used to determine adaptive weights based on user queries and query embedding vectors; The score standardization and fusion module is used to perform score standardization and dynamic fusion of sparse and dense search results based on adaptive weights to obtain the fused result. The candidate selection module is used to rearrange and select candidates from the fusion results to obtain Top-K candidate documents; The final answer generation module is used to generate the final answer based on the Top-K candidate documents and user queries.
[0008] Compared to existing technologies, this application provides an AI-based intelligent management method and system for equipment operation and maintenance. After parallel execution of sparse and dense searches to ensure comprehensive recall, it does not employ fixed fusion rules. Instead, it uses an adaptive weight network to interactively analyze the deep semantics and structured entity features of user queries, generating an optimal, dynamic fusion weight for each query in real time. This weight accurately determines the current query's dependence on keyword matching and semantic understanding, and accordingly performs intelligent dynamic weighting and rearrangement of the two search results, achieving precise candidate document selection based on a specific strategy for each query. This dynamic fusion mechanism fundamentally solves the problem of static weights being unable to adapt to query diversity, ensuring that the knowledge source provided to the downstream generation model is always highly relevant to the current query intent, thereby significantly improving the accuracy and reliability of the final generated answer. Attached Figure Description
[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.
[0010] Figure 1 This is a flowchart of an AI-based intelligent management method for equipment operation and maintenance according to an embodiment of this application; Figure 2 This is a data flow diagram of the AI-based intelligent management method for equipment operation and maintenance according to an embodiment of this application; Figure 3 This is a flowchart illustrating how the AI-based intelligent management method for device operation and maintenance, according to an embodiment of this application, uses parallel sparse and dense search paths to obtain sparse search results, dense search results, and query embedding vectors from user query inputs. Figure 4 This is a flowchart illustrating the process of determining adaptive weights based on user queries and query embedding vectors in the AI-based intelligent management method for device operation and maintenance according to embodiments of this application. Figure 5 This is a block diagram of an AI-based intelligent management system for device operation and maintenance according to an embodiment of this application. Detailed Implementation
[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.
[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.
[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.
[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.
[0015] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.
[0016] To address the technical problem of existing technologies that generally employ fixed weights when fusing sparse and dense search results, failing to adaptively adjust weights based on differences in query characteristics (such as keyword-driven or semantic-driven), this application proposes an AI-based intelligent management method for device operation and maintenance. This method begins by acquiring a user query and simultaneously feeding it into parallel sparse and dense search paths. The former generates sparse search results through keyword matching, while the latter generates dense search results and a key query embedding vector through deep semantic understanding. Furthermore, the system does not directly fuse these two sets of results. Instead, it utilizes the user query and query embedding vector to perform deep analysis through an adaptive weight network, accurately determining the query's dependence on keyword matching and semantic understanding, and generating a dynamic adaptive weight in real time. Subsequently, this adaptive weight is used to dynamically fuse and reorder the standardized sparse and dense search scores, selecting the most relevant Top-K candidate documents. Finally, these high-quality candidate documents, along with the original user query, are fed into a large language model as context to generate accurate and reliable final answers, effectively overcoming the limitations of fixed-weight fusion strategies.
[0017] The technical solution of this application proposes an intelligent management method for equipment operation and maintenance based on artificial intelligence. Figure 1 This is a flowchart of an AI-based intelligent management method for device operation and maintenance according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in an AI-based intelligent management method for device operation and maintenance according to an embodiment of this application. Figure 1 and Figure 2As shown, the AI-based intelligent management method for device operation and maintenance according to an embodiment of this application includes the following steps: S100, obtaining a user query input by the user; S200, inputting the user query into parallel sparse and dense retrieval paths to obtain sparse retrieval results, dense retrieval results, and query embedding vectors; S300, determining adaptive weights based on the user query and query embedding vectors; S400, performing score standardization and dynamic fusion on the sparse and dense retrieval results based on the adaptive weights to obtain fusion results; S500, reordering and selecting candidates from the fusion results to obtain Top-K candidate documents; S600, generating the final answer based on the Top-K candidate documents and the user query.
[0018] Specifically, in step S100, the user query input by the user is obtained. It should be understood that since the user-input query contains the user's direct operational needs, it is the fundamental starting point and basis for triggering the entire intelligent management process and providing the original basis for all subsequent analysis and decision-making steps. Therefore, in the technical solution of this application, the user query input by the user is obtained to initiate the subsequent parallel retrieval path and adaptive weight determination process. This ensures that the entire technical solution has a clear processing object and a predetermined problem-solving orientation, thereby activating the complete intelligent management and question-answering workflow.
[0019] More specifically, in a concrete example of this application, the implementation process of obtaining a user query input by the user is as follows: Maintenance personnel operate their terminal devices, such as computers or mobile terminals, to access a client application that provides a graphical user interface. This interface has dedicated text input fields, such as a query box or dialog box. The maintenance personnel use input devices such as keyboards to enter strings describing equipment malfunctions or maintenance problems, such as how to handle the main transformer protection device alarm code 0xFA01, into the text input field. After completing the input, the maintenance personnel perform a submit action, such as clicking the query button or pressing the Enter key. At this time, the client application captures the string in the text input field, encapsulates it into a standardized data request packet, and sends it to the backend server of the intelligent management system via a network communication protocol. Upon receiving the data packet, the server's application programming interface parses it and extracts the original query string, which constitutes the initial input data for all subsequent processing steps.
[0020] Specifically, in step S200, the user query input is processed in parallel using sparse and dense retrieval paths to obtain sparse retrieval results, dense retrieval results, and a query embedding vector. It should be understood that sparse and dense retrieval each have their inherent technical advantages and limitations when handling different types of queries, and a single retrieval technique cannot simultaneously meet all the requirements for precise keyword matching and deep semantic understanding. Therefore, in the technical solution of this application, the user query input is further processed in parallel using sparse and dense retrieval paths to obtain sparse retrieval results, dense retrieval results, and a query embedding vector. This fully utilizes the advantages of sparse retrieval in precise keyword matching and the advantages of dense retrieval in semantic similarity understanding, achieving complementary capabilities between the two retrieval paradigms. This maximizes the coverage of all potentially relevant candidate sets, providing a high-quality and comprehensive information foundation for subsequent dynamic fusion and accurate ranking, and avoiding the omission of key information due to the bias of a single retrieval method.
[0021] Figure 3 This is a flowchart illustrating the AI-based intelligent management method for device operation and maintenance according to embodiments of this application, which involves parallel sparse and dense search paths for user query input to obtain sparse search results, dense search results, and query embedding vectors. Figure 3 As shown, step S200 includes: S210, performing sparse retrieval and keyword scoring on the user query to obtain sparse retrieval results; S220, performing dense representation generation and semantic retrieval on the user query to obtain dense retrieval results and query embedding vector.
[0022] Accordingly, in step S210, sparse retrieval and keyword scoring are performed on the user query to obtain sparse retrieval results. It should be understood that in equipment operation and maintenance scenarios, user queries often contain clearly targeted entity keywords, such as specific equipment models, alarm codes, or technical terms. The literal matching degree of these keywords is a very strong signal for judging document relevance. Therefore, in the technical solution of this application, sparse retrieval and keyword scoring are performed on the user query to obtain sparse retrieval results, thereby specifically capturing and quantifying this relevance based on precise word frequency matching. This ensures that documents containing the exact same key identifiers as those in the user query are effectively recalled and assigned high scores, thus providing a high-precision candidate set for the entire retrieval system and compensating for the shortcomings of simple semantic retrieval, which may ignore or dilute the importance of specific keywords.
[0023] More specifically, in a concrete example of this application, the implementation process of sparse retrieval and keyword scoring for user queries is as follows: First, the received user query string, such as "How to handle the main transformer protection device alarm code 0xFA01", undergoes text preprocessing. This preprocessing includes word segmentation, breaking the query down into a list of terms, such as "main transformer protection device", "alarm code", "0xFA01", "how", and "handle", and removing non-discriminating stop words such as "how" and "handle", forming a core keyword set. Subsequently, this core keyword set is submitted to a sparse retrieval engine with a pre-built inverted index. The engine uses this inverted index to quickly locate all documents in the document library that contain one or more of the aforementioned core keyword elements, forming a preliminary candidate document set. Finally, for each document in the candidate document set, the BM25 algorithm is applied to calculate its keyword relevance score to the user query, i.e., the sparsity score. This score comprehensively considers the frequency of keyword occurrences in the document, the rarity of the keyword in the entire document database, and the document length, resulting in documents containing rare but important keywords (such as 0xFA01) receiving higher sparsity scores. Finally, all candidate documents are sorted in descending order according to their sparsity scores, and the top-ranked group of documents is selected as the sparse search results output.
[0024] Accordingly, in step S220, dense representation generation and semantic retrieval are performed on the user query to obtain dense retrieval results and query embedding vectors. It should be understood that due to the lexical mismatch between user queries and knowledge base documents—for example, users may use semantically similar but literally different words to describe the same operational problem—sparse retrieval methods relying solely on keyword matching cannot recall all semantically relevant documents. Therefore, in the technical solution of this application, dense representation generation and semantic retrieval are further performed on the user query to obtain dense retrieval results and query embedding vectors. This maps the query and documents to a unified semantic vector space and performs matching based on semantic similarity, thereby overcoming the barrier of literal expression. This effectively improves the recall rate of the retrieval system, ensuring that important documents highly relevant to the query intent but with different wording can also be successfully retrieved, while generating query embedding vectors that characterize the deep semantics of the query for subsequent adaptive weight determination.
[0025] Specifically, in this embodiment of the application, dense representation generation and semantic retrieval of user queries to obtain dense retrieval results and query embedding vectors include: inputting the user query into a language model encoder to obtain a query embedding vector; and performing an approximate nearest neighbor search on the query embedding vector in a vector database to obtain dense retrieval results.
[0026] More specifically, the implementation method for dense representation generation and semantic retrieval of user queries is as follows: The received user query string, such as "I feel that the vibration of wind turbine No. 2 has been a bit abnormal recently," is sent to a processing pipeline dedicated to semantic understanding. First, the user query is input into a language model encoder. Specifically, the query string is fed into a pre-trained language model encoder. This encoder performs deep context encoding on the input text sequence and generates a fixed-dimensional high-dimensional floating-point vector that can represent the complete semantic intent of the query through a pooling operation, such as extracting the output vector of a specific marker. This vector is the query embedding vector. In particular, in a specific example of this application, the language model encoder here is determined to be a pre-trained language model based on the Transformer bidirectional encoder architecture, such as the BERT model. Specifically, first, the user query string is input formatted. This processing includes adding a special classification marker [CLS] at the beginning of the string and adding a separator marker [SEP] at the end. Then, using a dictionary that accompanies the pre-trained model, the formatted text sequence is converted into a fixed-length sequence of numeric IDs consisting of an input ID, a mask ID, and a paragraph ID. Subsequently, the sequence of numeric IDs is used as input and fed into multiple stacked Transformer encoder layers of the BERT model for forward propagation computation. In each layer, a self-attention mechanism captures deep contextual dependencies between all tokens in the input sequence, generating a hidden state sequence rich in contextual information. After computation through all encoder layers, the final hidden layer output vector corresponding to the [CLS] classification label at the beginning of the input sequence is extracted. This vector, designed during training to aggregate global semantic information of the entire input sequence, is used as the query embedding vector representing the user's query. Finally, the query embedding vector is optionally normalized to become a unit vector to optimize the efficiency and stability of subsequent cosine similarity calculations in the vector database. Further, the query embedding vector is subjected to an approximate nearest neighbor search in the vector database. Specifically, the query embedding vector generated in the previous step is used as a query probe to perform a search task in a vector database that pre-stores the embedding vectors corresponding to all documents in the knowledge base. This database utilizes an approximate nearest neighbor search algorithm to efficiently calculate the cosine similarity between the query embedding vector and the massive number of document embedding vectors in the database. This similarity value is used as a density score. Documents are sorted based on this score, and the group of documents with the highest scores and their corresponding density scores are returned to form dense search results.
[0027] Specifically, in step S300, adaptive weights are determined based on the user query and query embedding vector. It should be understood that existing technologies using static fusion weights when fusing sparse and dense search results cannot adapt to the diversity of user queries. For example, when a query contains precise technical entities, the contribution of the two retrieval paths should differ from that of a query with ambiguous semantic descriptions. Fixed weights cannot reflect this difference, resulting in poor retrieval accuracy. Therefore, in the technical solution of this application, adaptive weights are further determined based on the user query and query embedding vector. This allows for in-depth analysis of the structured entity information and overall semantic information of the query, generating an optimized fusion strategy in real time for each independent retrieval task. This enables the subsequent fusion process to dynamically prioritize sparse or dense search results based on the nature of the query itself, fundamentally solving the technical bottleneck of fixed weights' inability to adapt to query diversity, thereby significantly improving the accuracy of search results.
[0028] Figure 4 This is a flowchart illustrating the process of determining adaptive weights based on user queries and query embedding vectors in an AI-based intelligent management method for device operation and maintenance according to embodiments of this application. (See attached flowchart.) Figure 4 As shown, step S300 includes: S310, performing entity recognition and entity feature vectorization on the user query to obtain an entity feature vector; S320, performing feature interaction based on semantic manifold gradient flow on the query embedding vector and the entity feature vector to obtain an entity-query semantic interaction feature matrix; S330, inputting the entity-query semantic interaction feature matrix into an adaptive weight network to obtain adaptive weights.
[0029] Accordingly, in step S310, entity recognition and entity feature vectorization are performed on the user query to obtain an entity feature vector. It should be understood that since the query embedding vector, which only contains global semantic information, cannot explicitly distinguish the structured entity information contained in the query string that is crucial for determining the query type, such as a specific alarm code or device model, these entities are the key basis for determining whether the query should focus more on keyword matching or semantic understanding. Therefore, in the technical solution of this application, entity recognition and entity feature vectorization are performed on the user query to obtain an entity feature vector. This explicitly extracts these highly informative structured features from the original, unstructured user query text and transforms them into a numerical representation that can be processed by the subsequent neural network model. This provides an independent data dimension specifically representing the key entities in the query for subsequent feature interaction and weight determination steps, enabling the final adaptive weight decision to be based simultaneously on the global semantics of the query and its internal key structured information, thus making the decision basis more comprehensive and accurate.
[0030] Specifically, in a concrete example of this application, the process of performing entity recognition and entity feature vectorization on a user query to obtain an entity feature vector includes: First, applying a named entity recognition model pre-trained on a corpus in the equipment operation and maintenance domain to process the user query string. For example, regarding the query string "How to handle the main transformer protection device alarm code 0xFA01", the named entity recognition model analyzes the text and identifies two predefined entity types contained therein: an entity of type "main transformer protection device" (equipment model) and an entity of type "alarm code 0xFA01". After completing entity recognition, the entity feature vectorization operation is then performed. This operation creates an initial vector of all zeros based on the length of a predefined list of entity types (e.g., containing all possible entity types such as alarm code, equipment model, location information, etc.). Then, iterates through all entity types identified in the previous step and sets the value at the dimension index position corresponding to the identified entity type (equipment model and alarm code in this example) in the vector to 1. Through this encoding process, an entity feature vector that accurately reflects the entity types contained in the user query and can be calculated by the subsequent model is finally generated.
[0031] Accordingly, in step S320, feature interaction based on semantic manifold gradient flow is performed on the query embedding vector and entity feature vector to obtain the entity-query semantic interaction feature matrix. It should be understood that in the hybrid retrieval scenario of intelligent operation and maintenance, simple feature concatenation has inherent technical defects. It cannot explicitly model the nonlinear modulation effect of entity information on query semantics. That is, juxtaposing the query embedding vector representing the global semantics of the query with the structured entity feature vector identifying key entities treats them as independent and equal information sources, thus ignoring the master-slave modulation relationship where the query entity, as the core anchor, defines and constrains the scope of application of the query semantics. This results in the loss or weakening of the crucial information of the nonlinear modulation effect of the entity on semantics during the fusion stage. Specifically, in practical application scenarios, entities included in the query, such as equipment models or alarm codes, are not ordinary features parallel to the query semantics, but rather core anchors defining and constraining the scope of application of that semantics. For example, the semantic connotation of the query "abnormal equipment vibration" will fundamentally change depending on whether the associated entity is a large transformer or a centrifugal water pump. A simple concatenation operation merely places two vectors side-by-side, hoping that the downstream neural network can implicitly learn this complex modulation relationship. This is not only inefficient but also results in poor model interpretability. More importantly, this concatenation operation is essentially a preparatory step for linear combination, which has an upper limit in its information representation ability. It cannot directly capture the nonlinear, interactive transformation relationship between the two vector spaces, leading to the partial loss or weakening of crucial information about the nonlinear modulation effect of entities on semantics during the feature fusion stage.
[0032] To address the insufficient information representation capability caused by simple feature concatenation, a preferred embodiment of this technical solution proposes a feature interaction generation method based on semantic manifold gradient flow. This method explicitly models the nonlinear modulation effect of entity information on the global query semantics. In this application, the query embedding vector and entity feature vector are further subjected to feature interaction based on semantic manifold gradient flow to obtain an entity-query semantic interaction feature matrix. This proactively and systematically simulates and realizes the nonlinear guidance and refinement process of entity information on query semantics. This generates a feature representation that deeply integrates the interaction between the two, resulting in a stronger information representation. This feature representation is no longer a simple combination of the two types of information, but rather the result of deep semantic modulation by the entity, providing higher-quality input for the downstream adaptive weight network.
[0033] Specifically, in this embodiment, feature interaction based on semantic manifold gradient flow is performed on the query embedding vector and the entity feature vector to obtain the entity-query semantic interaction feature matrix, including: constructing a semantic potential field based on the entity feature vector; performing semantic modulation along the gradient flow on the query embedding vector based on the semantic potential field to obtain the modulated query embedding vector; and performing feature interaction on the modulated query embedding vector and the entity feature vector to obtain the entity-query semantic interaction feature matrix.
[0034] More specifically, a semantic potential field is constructed based on entity feature vectors. It should be understood that since the entity feature vectors extracted from user queries are discrete, structured data representations, they cannot directly engage in meaningful mathematical interaction with the query embedding vectors representing a continuous semantic space, nor can they be directly used for subsequent gradient calculations. Therefore, in the technical solution of this application, a semantic potential field is further constructed based on the entity feature vectors to transform discrete entity information into a differentiable function that can produce continuous influence in a high-dimensional semantic space. This allows for the mathematical definition of a gravitational field determined by entity information, providing a computable mathematical basis for the subsequent precise modulation and guidance of query semantics along the gradient flow, thereby explicitly modeling the master-slave modulation relationship of entities on semantics.
[0035] In a specific example of this application, the implementation of constructing a semantic potential field based on entity feature vectors is as follows: First, a learnable linear transformation matrix is used to map the input entity feature vector from its original feature space to a semantic space with the same dimension as the query embedding vector, resulting in an entity target vector. This entity target vector semantically represents the ideal semantic center defined by all entities in the query. Subsequently, the weighted distance between any semantic vector and the entity target vector is defined as the potential energy value of the semantic vector in the potential field through a Gaussian kernel function, thereby constructing the semantic potential field function. This step mathematically constructs a gravitational field centered on specific entity information in the query. The semantic potential field is constructed using the following formula: in, For semantic potential field, Represents any semantic vector in the semantic space. For a scalar hyperparameter, To map entity feature vectors to a learnable weight matrix in the semantic space, For entity feature vectors, This represents the norm of a vector. Its significance lies in the fact that when the query entity is a main transformer, a center of interest is defined within a vast semantic space. All related semantic concepts, such as overheating and abnormal noise, should be attracted to the vicinity of this center. Furthermore, for a generalized query about abnormal equipment vibration, if the associated entity is a large transformer, this potential field will pull the semantic connotation of vibration towards abnormal, dangerous structural shaking; while if the associated entity is a centrifugal water pump, it will pull towards semantic directions exceeding the normal operating amplitude range, thus achieving refined contextual constraints on semantics. This achieves the purpose and effect of transforming entity information into a smooth, differentiable scalar field.
[0036] More specifically, based on the semantic potential field, the query embedding vector is semantically modulated along the gradient flow to obtain the modulated query embedding vector. It should be understood that since the previously constructed semantic potential field merely defines a gravitational distribution determined by entity information in the semantic space, the original query embedding vector representing the query generalization intent itself remains unchanged and independent of the entity information, failing to reflect the guiding and constraining role of the entity on the semantics. Therefore, in the technical solution of this application, the query embedding vector is further semantically modulated along the gradient flow based on the semantic potential field to obtain the modulated query embedding vector, thereby actively and non-linearly utilizing the constructed potential field to adjust the original query semantics. In this way, through a mathematically rigorous and interpretable optimization process, the original query semantic vector can be moved to a new position that better reflects its precise meaning in the specific entity context, thus generating a finely tuned query semantic vector that truly integrates entity information and better reflects the user's real and specific intent.
[0037] In a specific example of this application, the implementation of semantic modulation of the query embedding vector along the gradient flow based on the semantic potential field is as follows: First, the gradient of the semantic potential field function at the location of the original query embedding vector is calculated. This gradient is a vector with the same dimension as the query embedding vector, and its direction points in the direction of the fastest growth of the potential field function value. In the contextual sense, this gradient is like a semantic flow vector, indicating how the original, relatively generalized query semantics should be fine-tuned to better align with the entity context, thereby maximizing its fit with the interest center defined by the entity. Subsequently, the original query embedding vector is moved along the direction of this gradient vector by a step size controlled by the learning rate parameter, thus obtaining a new query embedding vector modulated by entity information. This mechanism no longer simply appends entity information to semantic information, but uses the potential field constructed by entity information to guide the direction of change of semantic information. The whole process is like pulling a general query intent precisely to a more specific and professional semantic location through the attraction generated by the entities it contains. Finally, through this update iteration, a new query embedding vector deeply modulated by entity information is obtained. This can be expressed by the following formula: in, This represents the modulated new query embedding vector. This is the original query embedding vector. A scalar learning rate controls the amplitude of the modulation. For variables The gradient operator.
[0038] More specifically, feature interaction is performed on the modulated query embedding vector and entity feature vector to obtain an entity-query semantic interaction feature matrix. It should be understood that after obtaining the semantic vector deeply modulated by entity information, simply concatenating it with the original entity feature vector fails to fully utilize the established nonlinear modulation relationship between the two, only retaining their independent information. A more informative fusion method than simple concatenation is needed to combine the modulated semantics with the original entity information for use by downstream models. Therefore, in the technical solution of this application, feature interaction is further performed on the modulated query embedding vector and entity feature vector to obtain an entity-query semantic interaction feature matrix. This matrix, through vector outer product operations, generates a feature matrix that explicitly represents the interaction strength between each dimension of the modulated semantics and different entity types. This constructs a highly information-dense semantic-entity interaction graph, providing the downstream adaptive weight network with richer, fine-grained information about which semantic dimension is strongly correlated with which entity type, compared to concatenated vectors. This generates a more expressive final feature representation, improving the accuracy of weight prediction.
[0039] In a specific example of this application, the implementation method for feature interaction between the modulated query embedding vector and the entity feature vector is as follows: First, the modulated query embedding vector generated in the previous step and the original entity feature vector are obtained. Then, a vector outer product operation is performed on these two vectors. Since the modulated query embedding vector is a d-dimensional vector and the entity feature vector is a k-dimensional vector, the outer product operation produces a d×k two-dimensional matrix, which is the entity-query semantic interaction feature matrix. This step constructs a semantic-entity interaction graph with extremely high information density. Unlike the concatenation operation, which only retains independent information, each element in the interaction tensor generated by the outer product represents the interaction strength between a certain modulated semantic dimension and a certain entity type. For example, one column of this tensor might clearly show that the alarm code entity mainly activates the semantic dimensions related to urgency and classification in the semantic vector, while another column might show that the device model entity mainly activates the semantic dimensions related to physical attributes and maintenance history. This provides the downstream neural network model with richer, finer-grained information about which semantic dimension is strongly correlated with which entity type, compared to the concatenated vector, thus generating a more expressive final feature representation. The modulated query embedding vector and entity feature vector are used for feature interaction using the following formula: in, The entity-query semantic interaction feature matrix represents the entity-query semantic interaction feature matrix. This is the modulated query embedding vector. For entity feature vectors, This represents the vector outer product operator.
[0040] Through the above preferred embodiments, the concept of gradient flow from differential geometry is introduced to upgrade the original, limited-expressive linear features into a refined feature interaction generation process capable of explicitly modeling the nonlinear modulation relationship between entities and semantics. This mechanism no longer passively relies on downstream models for implicit learning, but actively and principledly completes contextual fine-tuning of the query semantics during the feature engineering stage. By constructing a semantic potential field and modulating the query embedding vector along its gradient direction, a feature tensor rich in fine-grained interaction information is finally generated using outer product operations. This approach provides the subsequent weight prediction network with inputs that have higher information density and are more aligned with the business scenario logic, thereby enabling more accurate judgment of a specific query's dependence on sparse versus dense retrieval. Ultimately, this aims to significantly improve the prediction accuracy of dynamic weights in the hybrid retrieval system, thereby comprehensively optimizing the ranking quality of retrieval results and improving the accuracy and reliability of the entire intelligent question answering system.
[0041] Accordingly, in step S330, the entity-query semantic interaction feature matrix is input into an adaptive weight network to obtain adaptive weights. It should be understood that, since the entity-query semantic interaction feature matrix generated in the preceding steps is a high-dimensional, information-rich feature representation, but it is not the single scalar weight ultimately needed to guide the fusion of retrieval results, the system lacks a mechanism to automatically and accurately map this complex feature representation into the final decision weights. Therefore, in the technical solution of this application, the entity-query semantic interaction feature matrix is further input into an adaptive weight network to obtain adaptive weights, thereby utilizing the powerful nonlinear function fitting capability of neural networks to learn and establish the complex mapping relationship between deep interaction features and optimal fusion weights. In this way, the fine-grained features carefully constructed in the preceding steps can be transformed into a precise, quantified, and physically meaningful dynamic weight value, thus providing a direct, data-driven decision basis for the subsequent dynamic fusion steps.
[0042] Specifically, in a concrete example of this application, the implementation of inputting the entity-query semantic interaction feature matrix into an adaptive weight network to obtain adaptive weights involves first preprocessing the input entity-query semantic interaction feature matrix. For example, for a two-dimensional matrix of dimension d×k, a flattening operation is performed to transform it into a one-dimensional vector of dimension d*k. Subsequently, this one-dimensional vector is used as input and fed into a feedforward neural network, i.e., the adaptive weight network, which is pre-trained through supervised learning. The data processing flow of this network includes the input vector first passing through one or more hidden layers of the network. In each hidden layer, the data undergoes linear transformation and non-linear activation functions (such as ReLU) to extract and combine higher-order patterns from the features layer by layer. Finally, the data is passed to the output layer of the network, which contains only one neuron and uses the Sigmoid activation function. The characteristics of the Sigmoid function ensure that the final output value of the network is strictly constrained within the range of 0 and 1. This output value is then determined as the adaptive weight for the current user query.
[0043] Specifically, in step S400, based on adaptive weights, the sparse and dense search results are score-standardized and dynamically fused to obtain a fusion result. It should be understood that because the sparse search scores (such as BM25 scores) and dense search scores (such as cosine similarity) generated by parallel search paths are fundamentally heterogeneous in numerical scale, distribution range, and statistical significance, they cannot be directly compared or weighted. If fused directly without processing, one type of score will dominate the final result due to its numerical magnitude, thus causing the adaptive weights to lose their intended regulatory role. Therefore, in the technical solution of this application, based on adaptive weights, the sparse and dense search results are further score-standardized and dynamically fused to obtain a fusion result. This first maps the two heterogeneous scores to a unified, dimensionless common scale. Then, using the adaptive weights generated in the previous step and tailored to the current query, the standardized scores are dynamically weighted and summed. This ensures that the fusion process is conducted on a fair and comparable basis, allowing the adaptive weights to effectively balance and regulate the sparse and dense retrieval paradigms, ultimately resulting in a robust final ranking that accurately reflects the query intent.
[0044] More specifically, in a concrete example of this application, the implementation of score standardization and dynamic fusion of sparse and dense search results based on adaptive weights involves first creating a global candidate document set containing all unique documents from both the sparse and dense search result lists. Then, score standardization is performed on the scores of both result lists. For example, the system uses a min-max standardization method. For each document in the sparse search result list, its standardized sparse score is obtained by subtracting the minimum of all original sparse scores in the list from its original score, and then dividing by the difference between the maximum and minimum original sparse scores. Similarly, the same standardization operation is performed on the scores in the dense search result list. Through this process, the sparse and dense scores of all documents are normalized to the range of 0 to 1. After standardization, dynamic fusion is then performed. The final fused score is calculated by traversing each document in the global candidate document set and using a previously determined adaptive weight (e.g., 0.15 for the query "How to handle alarm code 0xFA01 of the main transformer protection device"). For a specific document, its fusion score is calculated by multiplying the document's standardized sparse score by (1 minus the adaptive weight) and then adding the document's standardized dense score by the adaptive weight. If a document is not included in a search result list, its corresponding standardized score is 0.
[0045] Specifically, in step S500, the fusion results are rearranged and candidate selection is performed to obtain Top-K candidate documents. It should be understood that since the fusion result generated in the previous step is a set containing all candidate documents and their corresponding fusion scores, this set itself is unordered and may contain a large number of documents with low relevance. Directly providing it to the downstream large language model would result in low-quality and inefficient answer generation due to information redundancy and context window limitations. Therefore, in the technical solution of this application, the fusion results are further rearranged and candidate selection is performed to obtain Top-K candidate documents. This establishes a clear, final relevance ranking of all candidate documents from high to low based on the carefully calculated fusion scores from the previous steps, and extracts a subset with a controllable quantity and the highest quality. This provides the final answer generation module with a highly refined, high signal-to-noise ratio source of contextual information, ensuring that the model can focus on the most relevant knowledge, thereby significantly improving the accuracy, relevance, and efficiency of the final generated answer.
[0046] More specifically, in a concrete example of this application, the implementation of rearranging and selecting candidates for the fusion results is as follows: The system first receives the fusion result from the previous dynamic fusion step, which is a set containing several documents and their corresponding fusion scores. Then, based on the fusion scores, all document-score pairs in the set are sorted in descending order. Finally, according to a pre-set integer parameter K (e.g., K is set to 3), the top K documents are truncated from the top of the sorted document list. This truncated, ordered list of documents constitutes the final Top-K candidate documents and is output as contextual knowledge for the next stage of the large language model to generate the answer.
[0047] Specifically, in step S600, the final answer is generated based on the Top-K candidate documents and the user query. It should be understood that since the Top-K candidate documents provided in the preceding steps are merely a highly relevant list of knowledge sources, rather than a directly user-facing, actionable answer, the user still needs to read, understand, and synthesize these documents to solve their problem, which fails to achieve the final closed loop of the intelligent question-answering system. Therefore, in the technical solution of this application, the final answer is further generated based on the Top-K candidate documents and the user query. This leverages the deep semantic understanding, logical reasoning, and natural language generation capabilities of a large language model to integrate, refine, and reconstruct the retrieved, fragmented knowledge items around the user query. In this way, the system output is transformed from a list of documents into a direct, complete solution that conforms to natural language expression habits, thereby providing users with an end-to-end intelligent question-answering service without secondary processing, completing the crucial leap from information retrieval to knowledge services.
[0048] More specifically, in a concrete example of this application, the implementation of generating the final answer based on Top-K candidate documents and the user query involves first constructing a structured prompt. This prompt is a carefully designed text template containing explicit system role instructions, the content of the retrieved Top-K candidate documents, and the original user query. Subsequently, this complete structured prompt is used as input to invoke a large language model (e.g., a pre-trained generative model based on the Transformer architecture). Upon receiving the prompt, the large language model first reads and understands the provided contextual information, i.e., the content of the Top-K candidate documents, and then focuses on the user's specific question, searching, extracting, and integrating all relevant information from the context. Finally, based on its internal knowledge and powerful generative capabilities, the large language model organizes this integrated information into a logically clear and fluent text, which is then output as the final answer.
[0049] In summary, the AI-based intelligent management method for device operation and maintenance according to the embodiments of this application is explained. After performing sparse and dense searches in parallel to ensure comprehensive recall, it does not employ fixed fusion rules. Instead, it uses an adaptive weight network to interactively analyze the deep semantics and structured entity features of the user query, generating an optimal, dynamic fusion weight for each query in real time. This weight accurately determines the current query's dependence on keyword matching and semantic understanding, and accordingly performs intelligent dynamic weighting and rearrangement of the two search results, achieving precise candidate document selection based on a specific strategy for each query. This dynamic fusion mechanism fundamentally solves the problem that static weights cannot adapt to query diversity, ensuring that the knowledge source provided to the downstream generation model is always highly relevant to the current query intent, thereby significantly improving the accuracy and reliability of the final generated answer.
[0050] Furthermore, an intelligent management system for equipment operation and maintenance based on artificial intelligence is also provided.
[0051] Figure 5 This is a block diagram of an AI-based intelligent management system for device operation and maintenance according to an embodiment of this application. Figure 5 As shown, the AI-based intelligent management system 100 for device operation and maintenance according to an embodiment of this application includes: a user query input acquisition module 110, used to acquire user queries input by users; a parallel retrieval module 120, used to parallelize sparse retrieval paths and dense retrieval paths of user query input to obtain sparse retrieval results, dense retrieval results and query embedding vectors; a weight generation module 130, used to determine adaptive weights based on user queries and query embedding vectors; a score standardization and fusion module 140, used to perform score standardization and dynamic fusion of sparse retrieval results and dense retrieval results based on adaptive weights to obtain fusion results; a candidate selection module 150, used to rearrange and select candidates from the fusion results to obtain Top-K candidate documents; and a final answer generation module 160, used to generate a final answer based on Top-K candidate documents and user queries.
[0052] As described above, the AI-based intelligent management system 100 for device operation and maintenance according to embodiments of this application can be implemented in various wireless terminals, such as servers with AI-based intelligent management algorithms for device operation and maintenance. In one possible implementation, the AI-based intelligent management system 100 for device operation and maintenance according to embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the AI-based intelligent management system 100 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the AI-based intelligent management system 100 can also be one of many hardware modules of the wireless terminal.
[0053] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.
Claims
1. A method for intelligent management of equipment operation and maintenance based on artificial intelligence, characterized in that, include: Retrieve user queries input by the user; The user query input is processed through parallel sparse and dense search paths to obtain sparse search results, dense search results, and query embedding vectors. Determine adaptive weights based on user queries and query embedding vectors; Based on adaptive weights, the scores of sparse and dense search results are standardized and dynamically fused to obtain the fused result. The fusion results are rearranged and candidates are selected to obtain Top-K candidate documents; The final answer is generated based on the Top-K candidate documents and user queries.
2. The intelligent management method for equipment operation and maintenance based on artificial intelligence according to claim 1, characterized in that, The user query input is processed in parallel using sparse and dense search paths to obtain sparse search results, dense search results, and query embedding vectors, including: Sparse search and keyword scoring are performed on user queries to obtain sparse search results; Dense representation generation and semantic retrieval are performed on user queries to obtain dense retrieval results and query embedding vectors.
3. The intelligent management method for equipment operation and maintenance based on artificial intelligence according to claim 2, characterized in that, Dense representation generation and semantic retrieval of user queries are performed to obtain dense retrieval results and query embedding vectors, including: The user query is input into the language model encoder to obtain the query embedding vector; The query embedding vector is used to perform an approximate nearest neighbor search in the vector database to obtain dense retrieval results.
4. The intelligent management method for equipment operation and maintenance based on artificial intelligence according to claim 1, characterized in that, Based on user queries and query embedding vectors, adaptive weights are determined, including: Entity recognition and entity feature vectorization are performed on user queries to obtain entity feature vectors; The query embedding vector and the entity feature vector are subjected to feature interaction based on semantic manifold gradient flow to obtain the entity-query semantic interaction feature matrix. The entity-query semantic interaction feature matrix is input into the adaptive weight network to obtain adaptive weights.
5. The intelligent management method for equipment operation and maintenance based on artificial intelligence according to claim 4, characterized in that, The query embedding vector and entity feature vector are subjected to feature interaction based on semantic manifold gradient flow to obtain the entity-query semantic interaction feature matrix, including: Construct a semantic potential field based on entity feature vectors; Based on the semantic potential field, the query embedding vector is semantically modulated along the gradient flow to obtain the modulated query embedding vector. The modulated query embedding vector and entity feature vector are subjected to feature interaction to obtain the entity-query semantic interaction feature matrix.
6. The intelligent management method for equipment operation and maintenance based on artificial intelligence according to claim 5, characterized in that, Based on entity feature vectors, a semantic potential field is constructed, including: constructing the semantic potential field using the following formula, where the formula is: in, For semantic potential field, Represents any semantic vector in the semantic space. For a scalar hyperparameter, To map entity feature vectors to a learnable weight matrix in the semantic space, For entity feature vectors, The first norm of a vector is denoted by 1.
7. The intelligent management method for equipment operation and maintenance based on artificial intelligence according to claim 5, characterized in that, To obtain an entity-query semantic interaction feature matrix, the modulated query embedding vector and entity feature vector are subjected to feature interaction using the following formula: in, The entity-query semantic interaction feature matrix represents the entity-query semantic interaction feature matrix. This is the modulated query embedding vector. For entity feature vectors, This represents the vector outer product operator.
8. An intelligent management system for equipment operation and maintenance based on artificial intelligence, characterized in that, include: The user query input acquisition module is used to acquire user queries input by the user. The parallel retrieval module is used to process user query inputs into parallel sparse and dense retrieval paths to obtain sparse retrieval results, dense retrieval results, and query embedding vectors. The weight generation module is used to determine adaptive weights based on user queries and query embedding vectors; The score standardization and fusion module is used to perform score standardization and dynamic fusion of sparse and dense search results based on adaptive weights to obtain the fused result. The candidate selection module is used to rearrange and select candidates from the fusion results to obtain Top-K candidate documents; The final answer generation module is used to generate the final answer based on the Top-K candidate documents and user queries.
Citation Information
Cited By
Text classification method, system and equipment and medium
CN121681836A
A text classification method, system, device, and medium
CN121681836B
