A large model-based knowledge base application method and system

Abstract

The application discloses a kind of knowledge base application methods and systems based on large model, can utilize large model, artificial intelligence or AI to quickly and efficiently obtain content information, to solve the answer content isolated generated in existing operation and maintenance question and answer database system, lack executable decision path, cannot cyclically utilize historical maintenance experience and the technical problem that operation and maintenance log is not fully utilized.Method includes: the initial water affair question text input by user is based on the multi-round question and answer reverse driving and path dynamic growth in the state of decision node;Experience knowledge node is compiled to water affair execution log array, and the incremental update processing of experience knowledge node is carried out on the success rate attribute feature;The received new water affair question text is processed by question fingerprint quick matching;And the new water problem text that cannot be quickly matched is mounted with supplementary knowledge node processing;Based on water affair execution log array and water affair patch log array, generate water affair operation and maintenance report.

Description

Technical Field

[0001] This application relates to the fields of water data processing and intelligent operation and maintenance technology, and in particular to a knowledge base application method and system based on a large model. Background Technology

[0002] In engineering organization and management, a knowledge base is often built based on data materials such as laws and regulations, industry standards, product and equipment manuals, and work logs to facilitate enterprise reference and project correlation. For example, in wastewater treatment projects within the large-scale water conservancy industry, daily operation and maintenance involve numerous legal regulations and a large number of complex equipment. Water pipe networks are the "blood vessels" of the urban water cycle system, encompassing multiple stages such as raw water transportation, tap water supply, wastewater collection and treatment, and rainwater discharge. Throughout the water pipe network, numerous types of equipment and facilities are deployed. These devices typically have their own operation manuals, maintenance guidelines, safety specifications, fault code tables, and other documentation. When maintenance personnel handle equipment malfunctions or perform routine maintenance, they need to frequently consult these scattered documents and rely on their accumulated experience to determine the cause of the problem, develop repair plans, execute operating procedures, and prevent safety risks.

[0003] Furthermore, with the development of information technology, some water utilities have begun to build equipment management systems or knowledge base systems to digitally store equipment documents and provide keyword search functions. However, the limitations and problems of such systems are particularly prominent in the actual operation and maintenance of water conservancy networks (especially in sewage treatment).

[0004] Wastewater treatment plants are critical infrastructure for urban water environment management. Their daily operation and maintenance involve a large number of complex pieces of equipment, such as magnetic levitation blowers, electromagnetic flow meters, hydraulic pressure sensors, various pumps, aeration devices, and sludge dewatering machines. These devices are typically equipped with their own operation manuals, maintenance guidelines, safety regulations, fault code tables, and other documentation. When troubleshooting equipment malfunctions or performing routine maintenance, maintenance personnel need to frequently consult these scattered documents and rely on their accumulated experience to determine the cause of the problem, develop maintenance plans, execute operating procedures, and prevent safety risks.

[0005] In other words, the aforementioned wastewater operation and maintenance management systems still have certain technical limitations in practical applications: First, existing systems generally adopt a flat interaction mode of user-initiated retrieval + system-returned document fragments, which is prone to inefficiency and missing key steps; Second, existing systems lack the ability to extract the causal chains, temporal chains, and conditional branches implicit in equipment documents in a structured manner, and fail to explicitly extract these deep relational structures and organize them into a knowledge network that can be used for reasoning; Third, the question-and-answer records and maintenance logs of existing systems are often isolated text records, lacking the ability to be structured and reused.

[0006] Therefore, based on the aforementioned problems existing in current knowledge bases, this invention provides a knowledge base application method and system based on a large model. Summary of the Invention

[0007] This application provides a knowledge base application method and system based on a large model to solve the following technical problems: the answers generated by existing sewage treatment maintenance question and answer systems are isolated, lack executable decision paths, cannot reuse historical maintenance experience, and do not make full use of operation and maintenance logs.

[0008] The embodiments of this application adopt the following technical solutions: On one hand, this application provides a knowledge base application method based on a large model, including: extracting multimodal content and decomposing hierarchical semantics from water-related files uploaded by the client to construct a water graph database; using the water graph database, performing multi-round question-and-answer reverse-drive and dynamic path growth based on decision node states on the initial water-related question text input by the user to obtain a water-related question decision path tree; visually encoding the water-related question decision path tree for relevant decision node types, and based on an event listening function, recording the execution status of the user's interactive maintenance processing nodes in real time to generate a water-related execution day. The water affairs execution log array is processed by compiling experience knowledge nodes, and incrementally updating the experience knowledge nodes with success rate attribute features using unique log fingerprints to obtain an updated water affairs experience node set. Based on the unique log fingerprints and the water affairs experience node set, new water affairs problem texts are quickly matched with problem fingerprints. New hydrological problem texts that cannot be quickly matched are supplemented with knowledge nodes to obtain a water affairs patch log array. A water affairs operation and maintenance report is generated based on the water affairs execution log array and the water affairs patch log array.

[0009] This application's embodiments generate a decision path tree through a reverse-driven, multi-round question-and-answer process, providing maintenance personnel with a complete and executable solution from risk identification to execution of procedures and emergency response, rather than fragmented text snippets. This significantly reduces the difficulty and risk of on-site decision-making. Furthermore, by compiling maintenance logs into experience knowledge nodes and dynamically updating them to the knowledge graph, each maintenance operation provides the system with learning and training opportunities, enabling continuous accumulation and intelligent reuse of operational experience and solving the problem of traditional experience loss. Conflict detection and versioning technologies allow newly added knowledge to take effect without disrupting the existing authoritative document structure and can be dynamically selected based on contextual conditions, enhancing the system's robustness. Moreover, complete operational records and frequency statistics of decision path usage provide managers with quantitative analysis data, enabling the identification of frequently used decision paths or paths with low success rates.

[0010] In one feasible implementation, water-related files uploaded by the client are subjected to multimodal content extraction and hierarchical semantic decomposition to construct a water graph database. Specifically, this includes: uploading the water-related files to the wastewater treatment system via the client; wherein the water-related files include at least: a 4G pressure sensor instruction manual, a magnetic levitation blower operation manual, an electromagnetic flowmeter instruction manual, a water pump operation manual, an aeration device operation manual, and a sludge dewatering machine operation manual; extracting the internet media type identifier and raw byte stream data for each water-related file; performing content extraction processing on the water-related files according to the file format type based on the parser corresponding to the internet media type identifier and raw byte stream data to obtain an initial paragraph sequence; and using a pre-trained deep learning named entity recognition model to perform named entity recognition and annotation on the text entity sequence of the initial paragraph sequence to obtain a water equipment entity sequence. The water equipment entity sequence includes at least: text content, tag type, and the starting and ending character positions within a paragraph. Using a relation type dictionary, dependency parsing is performed on the semantic text of the water equipment entity sequence. Based on the dependency relations and core verbs in the statements, relation triples are generated from the semantic text to determine the entity semantic relations for each water equipment entity. The relation type dictionary includes: parameters, standard operating procedures, prohibited items, risk identification, preconditions, and emergency response measures for water equipment. Each relation triple in the entity semantic relations is defined as a directed edge, and each entity in the entity semantic relations is defined as a node. Based on the directed edges and the nodes, a water graph database is constructed. The storage attributes of each node in the water graph database include: entity text content, entity type, source document path, and source document page number.

[0011] This application's embodiments achieve standardized processing of different water-related documents through multimodal content extraction and hierarchical semantic decomposition, improving data consistency and usability. It can quickly extract key information from documents, such as parameters, operating procedures, and prohibitions in equipment manuals, improving the efficiency of information retrieval and processing. Furthermore, the constructed water graph database can display the relationships between water equipment in a graph form, aiding in the understanding and analysis of complex water systems. Deep learning models are also used for named entity recognition and dependency parsing, enhancing the system's semantic understanding of water documents. Simultaneously, through relation triple generation, the semantic relationships of each water equipment entity can be clearly defined, providing a foundation for subsequent reasoning and analysis. Based on the constructed graph database, water management decisions, such as equipment maintenance, fault diagnosis, and process optimization, can also be supported.

[0012] In one feasible implementation, after constructing the water resources map database based on the directed edges and the nodes, the method further includes: attaching a set of dimension labels to each entity node in the water resources map database; wherein the initial value of each dimension label is empty, and the label dictionary of the dimension labels includes: equipment type, operation stage, risk level, and applicable working conditions; according to the label dictionary, and through the rule engine of the sewage treatment system, automatically assigning label types to the entity text content to obtain a set of assigned labels for each entity node; performing serialization processing on the assigned label set to obtain extended label attribute features; and extending and storing the extended label attribute features with each node in the water resources map database to obtain a water resources map database with attached dimension labels.

[0013] In one feasible implementation, before obtaining a water issue decision path tree by performing multi-round question-and-answer reverse-drive and path dynamic growth based on decision node states on the initial water issue text input by the user through the water graph database, the method further includes: extracting issue entity features from the initial water issue text using a deep learning named entity recognition model to obtain a core entity set; and identifying a set of issue entities with an abnormal label type; using the issue entity set, performing label type matching processing on the water graph database for the main nodes, and based on the identified main nodes, determining adjacent nodes and associated subgraphs directly connected to the main nodes through arbitrary relationships; wherein, the main node is the node matched with the issue entities in the issue entity set; based on the issue entity set, An initial decision path skeleton is constructed; wherein the initial decision path skeleton includes a root node and an empty decision path array, and each element in the decision path array represents a decision node, and the node type of each decision node includes at least: a precondition check node, a standard operating procedure node, a risk warning node, a prohibited item node, and an emergency response node; according to the progressive order of the attribute items in the node type, the directed edges corresponding to the node type in the associated subgraph are queried, and several target nodes corresponding to the directed edges are sequentially filled into the decision nodes of the initial decision path skeleton; wherein the progressive order of the attribute items is the process processing order of each node type; according to the causal relationship between the target nodes, the corresponding decision nodes are defined with node association relationships under parent-child relationships to obtain the decision path skeleton.

[0014] This application's embodiments utilize a multi-round question-and-answer system to intelligently understand and respond to users' water-related questions, providing immediate decision support. Furthermore, by employing a deep learning model to extract problem entity features, the system can accurately identify and locate key information within water-related issues, improving the accuracy of question-and-answer responses. It can also identify problem entities with abnormal status labels, facilitating rapid location and response to potential problems or faults. Moreover, through a water-related graph database, the system can efficiently utilize the relationships and knowledge between water-related equipment to provide users with precise decision-making paths. Based on the state of decision nodes, the system can dynamically grow decision paths to adapt to different problem scenarios and decision-making needs. The constructed decision path tree provides users with a clear decision-making process, aiding in the understanding and execution of complex decision-making tasks.

[0015] In one feasible implementation, the initial water-related question text input by the user is processed through a multi-round question-and-answer reverse-drive and dynamic path growth based on the decision node status using the water map database to obtain a water-related question decision path tree. Specifically, this includes: traversing all attribute nodes of type precondition check in the decision path skeleton, and generating follow-up question strings based on the pending status of the attribute nodes; wherein the follow-up question string includes at least: string format, required equipment parameters, and equipment operation requirements; pushing the follow-up question string to the user interface and starting a timeout timer; based on the answer status of the follow-up question string, performing feature word judgment processing on the user-input answer string for confirmation and / or negation status to obtain follow-up question feedback result information; if the follow-up question feedback result information is confirmation status information, then updating the status of the precondition check decision node to the passed status; and using the water map database, the precondition check... Starting from a decision node, the operational process features in the queried standard operating procedure decision nodes are sequentially executed; the queried standard operating procedure decision nodes are then identified as child nodes of the current node; if the follow-up feedback information is an error status, the status of the precondition check decision node is updated to a failed status; and through the associated subgraph, directed edge queries with relational types are performed on the precondition check decision nodes to identify decision nodes with remedial measures; the target node among the decision nodes with remedial measures is identified as a new standard operating procedure decision node; the new standard operating procedure decision node is identified as a child node of the current node, and a new follow-up question string is generated based on the new standard operating procedure decision node; this process is repeated until all follow-up feedback information is a confirmed status and no new precondition check decision nodes are generated, thus generating the water issue decision path tree with water equipment operation features and a question-and-answer branch structure.

[0016] This application's embodiments can automatically handle water-related issues, generating decision path trees and reducing reliance on manual decision-making. It can also dynamically adjust decision paths based on user input and decision node status to adapt to different problems and situations. Furthermore, through follow-up questions and user feedback, the system can accurately handle problems, reducing misunderstandings and errors. The user interface displays follow-up questions and operation procedures, improving clarity and comprehension for users. The status of decision nodes (e.g., passed or failed) is updated in real time, allowing both users and the system to track decision progress. Operations are performed according to standard operating procedures, ensuring the standardization and safety of water equipment operation. Simultaneously, the generated decision path tree has a question-and-answer branching structure, facilitating user understanding of different decision paths and possible outcomes. Combining the recording and display of decision paths also helps accumulate water management knowledge and experience, promoting knowledge sharing.

[0017] In one feasible implementation, the water issue decision path tree is visually encoded with respect to the types of decision nodes. Based on an event listener function, the execution status of user interaction maintenance processing nodes is recorded in real time to generate a water execution log array. Specifically, this includes: recursively expanding the water issue decision path tree and assigning a unique user interface component identifier to each decision node; using the user interface component identifier, performing front-end rendering processing on the node type of each decision node, and sorting all rendered user interface components by color depth to obtain the view area interface of the water issue decision path tree; wherein, the user interface component includes: a precondition check. The system includes type nodes, standard operation process type nodes, risk warning type nodes, prohibited items type nodes, and emergency response type nodes. After generating the view area interface, the event listener function is bound to each rendered view area interface. Through the event listener function, the execution status of the interaction data of each decision node in the water issue decision path tree is recorded in real time. Based on the user's checkbox and supplementary content actions, the action status of the checkboxes and confirmation boxes in the water issue decision path tree is captured for the relevant interaction maintenance processing nodes. Based on the interaction data and action status data, the water execution log array based on the water issue decision path tree and the corresponding unique identifier are generated.

[0018] In one feasible implementation, the water affairs execution log array is processed to compile experience knowledge nodes, and the experience knowledge nodes are incrementally updated with success rate attribute features using unique log feature fingerprints to obtain an updated set of water affairs experience nodes. Specifically, this includes: performing a distributed task scan search on the water affairs execution log data group, and marking water affairs execution log data groups with the tag type "problem solved" as data groups to be compiled; extracting water affairs equipment identifiers, initial problem text features, and execution step sequence features from the data groups to be compiled; concatenating the extracted related features into strings to generate the unique log feature fingerprint based on semantic vector retrieval; and compiling experience knowledge nodes from the water affairs execution log array that meets the "problem solved" feature according to the unique log feature fingerprint. The node attributes of the experience knowledge nodes include: unique feature fingerprint, equipment identifier, fault phenomenon, and solution summary. The process involves: establishing a bidirectional relationship between the experience knowledge nodes and adding them, along with the two directed edges, to the water resources graph database. The first directed edge points from the experience knowledge node to the corresponding equipment entity node, and the second edge points from the experience knowledge node to the root node of the water resources problem decision path tree used in this maintenance. The process then performs a similarity check on the unique feature fingerprints of subsequent new experience knowledge nodes. If the new log unique feature fingerprints are similar, the success rate attribute and the number of successful resolutions attribute of the corresponding experience knowledge node are summed. If the new log unique feature fingerprints are not similar, the corresponding new experience knowledge node is added to the water resources graph database. Based on whether the new log unique feature fingerprints are similar, incremental updates are performed on the decision nodes in the water resources graph database to obtain the updated set of water resources experience nodes.

[0019] This application's embodiments, by compiling and processing experience knowledge nodes into water management execution log arrays, can quickly identify and resolve common problems, reduce repetitive processing, and improve overall work efficiency. Furthermore, by combining incremental updates of the success rate attribute of experience knowledge nodes, solutions can be continuously optimized, increasing the success rate of problem resolution. Experience knowledge nodes can also be stored in a structured manner in a database, facilitating management and retrieval, and contributing to knowledge accumulation and inheritance. Simultaneously, by extracting unique feature fingerprints from the logs, rapid filtering and analysis of large amounts of water management execution log data can be achieved, improving the intelligence level of log processing. Moreover, water management execution log data can be transformed into experience knowledge nodes, enhancing the value of the data and providing strong support for subsequent decision support and optimization.

[0020] In one feasible implementation, based on the unique feature fingerprint of the log and the set of water affairs experience nodes, a fast matching process is performed on the received new water affairs problem text for problem fingerprints; and new hydrological problem texts that cannot be quickly matched are supplemented with knowledge nodes to obtain a water affairs patch log array. Specifically, this includes: calculating the new log unique feature fingerprint of the new water affairs problem text; performing a query matching process on the set of water affairs experience nodes in the water affairs graph database based on the new log unique feature fingerprint to obtain a query matching result; if the query matching result shows that a matching feature fingerprint exists, the corresponding experience knowledge node is determined as a high-priority standard operating procedure type decision node, and the standard operating procedure type decision node is... Insert it at the front of the water issue decision path tree corresponding to the new water issue text; if the query matching result is a mismatch feature fingerprint, then compare the new knowledge node attributes of the new water issue text with the original knowledge node attributes of similar water issue texts to determine the supplementary knowledge data object; determine the supplementary knowledge data object as the patch knowledge node in the water graph database, and based on the graph structure position of the similar experience knowledge node in the water graph database, mount the patch knowledge node to the similar experience knowledge node to obtain the water patch log array; wherein, the similar experience knowledge node is compiled from the water execution log array corresponding to the similar water issue text that meets the problem-solved feature.

[0021] This application embodiment enables the system to quickly respond to emerging issues and reduce processing time by rapidly matching the problem fingerprints of new water-related problem texts. Furthermore, for new problems that cannot be quickly matched, the system can efficiently handle them by attaching supplementary knowledge nodes, improving problem-solving efficiency. In addition, by attaching supplementary knowledge nodes, the water-related graph database can continuously adapt to new problem scenarios, enhancing the adaptability and flexibility of the knowledge base. Specifically, inserting matched experiential knowledge nodes as high-priority decision nodes at the forefront of the problem decision path tree helps optimize the decision path and improve decision accuracy. Moreover, by comparing the attributes of existing knowledge nodes in similar water-related problem texts, it can be ensured that the supplementary knowledge data objects match the existing knowledge system, thereby improving the quality of problem-solving. Patch knowledge nodes can also be attached to similar experiential knowledge nodes, promoting knowledge sharing and dissemination, and helping to accumulate and expand knowledge in the water-related field. Simultaneously, the automated problem matching and knowledge node attachment process reduces the need for manual intervention and lowers operation and maintenance costs.

[0022] In one feasible implementation, a water affairs operation and maintenance report is generated based on the water affairs execution log array and the water affairs patch log array. Specifically, this includes: aggregating the water affairs execution log array and the water affairs patch log array retrieved from the backend distributed search through the front-end interface of the sewage treatment system; statistically processing the extracted aggregated data for indicators such as the frequency of equipment failures, the number of times and success rate of water affairs problem decision paths, and the number and duration of maintenance personnel handling the issues, to obtain statistical results; and transmitting the statistical results to the front-end interface to generate the water affairs operation and maintenance report.

[0023] On the other hand, this application also provides a knowledge base application system based on a large model, characterized in that the knowledge base application system based on a large model can be executed by at least one processor, so that at least one processor can execute the method of a knowledge base application system based on a large model described in any of the above embodiments.

[0024] This application provides a method and system for a knowledge base application system based on a large model. Compared with the prior art, the embodiments of this application have the following beneficial technical effects: 1. By generating a decision path tree through multi-round question-and-answer sessions driven by reverse engineering, a complete and actionable solution is provided to maintenance personnel, from risk identification to implementation of procedures and emergency response, rather than fragmented text snippets, which significantly reduces the difficulty and risk of on-site decision-making.

[0025] 2. By compiling maintenance logs into experience knowledge nodes and dynamically updating them to the knowledge graph, each maintenance provides learning and training for the system, enabling continuous accumulation and intelligent reuse of maintenance experience, and solving the problem of traditional experience loss.

[0026] 3. Through conflict detection and versioning mounting technology, newly added knowledge can take effect without breaking the original authoritative document structure, and can be dynamically selected according to context conditions, thus enhancing the robustness of the system.

[0027] 4. Complete operation and maintenance records and frequency statistics of decision-making paths provide managers with quantitative analysis basis, which can identify frequently used decision-making paths or paths with low success rates. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1A flowchart illustrating a knowledge base application method based on a large model, provided in an embodiment of this application; Figure 2 This is a schematic diagram of a knowledge base application system based on a large model, provided as an embodiment of this application. Detailed Implementation

[0029] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0030] It should be noted that the technical solution proposed in this application is not limited to wastewater treatment systems. The core technology of this application lies in: constructing a knowledge graph through three-level semantic decomposition of equipment documentation; generating a decision path tree based on user-initiated questions through multi-round question-and-answer processing; and achieving knowledge cyclical evolution by compiling experiential knowledge nodes from maintenance logs. The technical concept of this application also has universality and can be widely applied to the entire water conservancy network field. Specifically, water conservancy networks encompass multiple subsystems such as raw water transmission, tap water supply, wastewater collection and treatment, and rainwater discharge. Each subsystem deploys a large number of devices (such as water pumping stations, pressure regulating valves, flow monitoring instruments, and water quality analyzers). Maintenance personnel also face problems such as scattered documentation, reliance on experience, and missing decision paths. Therefore, the method claimed in this application is not only applicable to wastewater treatment systems but can be adaptively configured to cover the entire water conservancy network environment for dynamic generation of equipment maintenance decisions.

[0031] This application provides a method for applying a knowledge base based on a large model, such as... Figure 1 As shown, the knowledge base application method based on the large model specifically includes steps S101-S106: S101. Extract multimodal content and hierarchical semantic decomposition from the water-related files uploaded by the client to construct a water-related graph database.

[0032] Specifically, the first step is to upload water-related documents to the wastewater treatment system via the client. These water-related documents include at least the following: 4G pressure sensor instruction manual, magnetic levitation blower operation manual, electromagnetic flowmeter instruction manual, water pump operation manual, aeration device operation manual, and sludge dewatering machine operation manual.

[0033] Furthermore, the internet media type identifier and raw byte stream data of each water-related document are extracted. Based on the parser corresponding to the internet media type identifier and the raw byte stream data, the content of the water-related documents is extracted according to the file format type to obtain the initial paragraph sequence.

[0034] In one embodiment, the wastewater treatment system first receives a file uploaded by the client, supporting portable document format, word processing document format, plain text format, and scanned image format. It then obtains the file's Internet media type identifier and raw byte stream data. Next, it calls the corresponding parser based on the Internet media type identifier. If the file format is portable document format, it uses a portable document parsing tool library to extract the text and images; if the file format is scanned image format, it uses an optical character recognition engine to perform text recognition and convert the image into a text stream. Finally, the extracted raw text stream is segmented according to line breaks and periods to form an initial paragraph sequence, and the page number of each paragraph in the file is recorded.

[0035] Furthermore, a pre-trained deep learning named entity recognition model is needed to perform named entity recognition and annotation on the initial paragraph sequence to obtain the water equipment entity sequence. The water equipment entity sequence includes at least: text content, tag type, and the starting and ending character positions within the paragraph.

[0036] In one embodiment, a pre-trained deep learning named entity recognition model is required. This deep learning named entity recognition model adopts an architecture that combines bidirectional encoder representation with bidirectional long short-term memory network and conditional random field. The training dataset of the model contains 5,000 manually annotated water domain corpora. The annotation label types include: equipment name, parameter name, operation action, numerical threshold, and prohibited action.

[0037] In one embodiment, each initial paragraph sequence is then input into the deep learning named entity recognition model. The model outputs a sequence of water equipment entities for that paragraph. Each water equipment entity sequence includes its text content, tag type, and the starting and ending character positions within the paragraph. Then, the entity sequences output from all paragraphs are merged and deduplicated to obtain the entity set for the entire document. For example, from the instruction manual for a magnetic levitation blower, the system extracts entities whose text content is "magnetic levitation blower" and whose tag type is "equipment name"; it extracts entities whose text content is "bearing temperature" and whose tag type is "parameter name"; and it extracts entities whose text content is "must not operate without cooling water" and whose tag type is "prohibited action".

[0038] Furthermore, using a predefined relation type dictionary, dependency parsing is performed on the semantic text of the water equipment entity sequence. Based on the dependency relations and core verbs in the statements, relation triples are generated from the semantic text to ultimately determine the entity semantic relations for each water equipment entity. The relation type dictionary includes: water equipment having parameters, having standard operating procedures, having prohibited items, having risk identification, having preconditions, and having emergency response measures.

[0039] In one embodiment, the wastewater treatment system predefines a relation type dictionary, including: parameters, standard operating procedures, prohibited items, risk identification, preconditions, and emergency response measures. Then, based on dependency parsing technology, semantic relationships between entities are constructed. Specifically, for the semantic sentence "Before starting the magnetic levitation blower, it must be confirmed that the cooling water flow rate is greater than ten cubic meters per hour," the system parses the dependency relationships in the sentence: the logical subject of the core verb "confirmation" is the implicit operator, and the object clause "cooling water flow rate is greater than ten cubic meters per hour" modifies "magnetic levitation blower." Based on this, the system generates two relation triples: the first is "magnetic levitation blower," with the precondition "cooling water flow rate is greater than ten cubic meters per hour," and the second is "startup operation," also with the precondition "cooling water flow rate is greater than ten cubic meters per hour."

[0040] Furthermore, each relation triple in the entity semantic relationship must be defined as a directed edge, and each entity in the entity semantic relationship must be defined as a node. Based on the directed edges and nodes, a water resources map database is constructed. The storage attributes of each node in the water resources map database include: entity text content, entity type, source document path, and source document page number.

[0041] In one embodiment, all generated relation triples also need to be stored in a water resources graph database. In this database, each triple is represented as a directed edge connecting two nodes, and each node represents an entity. The attributes stored in each node include: entity text content, entity type, source document path, and source document page number.

[0042] As a feasible implementation method, a set of dimensional labels needs to be attached to each entity node in the water resources map database. The initial value of each dimensional label is empty, and the label dictionary includes: equipment type, operation stage, risk level, and applicable operating conditions. Then, based on the label dictionary and through the wastewater treatment system's rule engine, the entity text content is automatically assigned label types, resulting in a set of assigned labels for each entity node. Next, the assigned label sets undergo structured data serialization processing to obtain extended label attribute features. Finally, the extended label attribute features are extended and stored along with each node in the water resources map database to obtain a water resources map database with attached dimensional labels.

[0043] In one embodiment, the wastewater treatment system also needs to attach a set of predefined dimension labels to each entity node, with each label initially empty. The label dictionary includes: equipment type, operation stage, risk level, and applicable operating conditions. A rule engine is then run to automatically assign values ​​to the labels, resulting in a set of assigned labels. For example, if the entity text contains keywords such as blower or flow meter, the equipment type label is set to mechanical equipment or instrumentation equipment; if the entity text contains keywords such as not allowed, strictly prohibited, or forbidden, the risk level label is set to high, and the operation stage label is set to under maintenance. Finally, the assigned label set is serialized according to a structured data format and stored as an extended attribute of the node in the graph database, resulting in a water resources graph database with attached dimension labels.

[0044] S102. Using the water resources map database, the initial water resources question text input by the user is subjected to multi-round question-and-answer reverse driving and dynamic path growth based on the decision node state to obtain the water resources question decision path tree.

[0045] Specifically, a deep learning named entity recognition model is first used to extract the features of problem entities from the initial water-related problem text to obtain a core entity set. Then, a set of problem entities with the tag type "abnormal state" is identified.

[0046] In one embodiment, the initial question text input by the user needs to be fed into the aforementioned deep learning named entity recognition model. The system then extracts a set of core entities from the question and labels the entities representing fault phenomena as "abnormal state." For example, if the user inputs "magnetic levitation blower bearing temperature too high," the system extracts two core entities: the first is the text content "magnetic levitation blower" with the label "equipment name," and the second is the text content "bearing temperature too high" with the label "abnormal state."

[0047] Furthermore, by combining the problem entity set, the label type matching process for the main nodes in the water resources map database is performed. Based on the identified main nodes, adjacent nodes and associated subgraphs directly connected to the main nodes through arbitrary relationships are then determined. Here, the main node is the node that matches the problem entity in the problem entity set.

[0048] In one embodiment, a graph query statement is executed in the water graph database, and then the associated subgraphs within the range of the corresponding device and its adjacent nodes are queried. The logic of the query statement is as follows: match all nodes with the label type "device" and the text content equal to "magnetic levitation blower", and then return the node and all adjacent nodes and relationships directly connected to the node through any relationship.

[0049] Furthermore, an initial decision path skeleton needs to be constructed based on the set of problem entities. The initial decision path skeleton contains a root node and an empty decision path array, and each element in the decision path array represents a decision node. The node type of each decision node includes at least the following: prerequisite check node, standard operating procedure node, risk warning node, prohibited items node, and emergency response node.

[0050] In one embodiment, a data structure for a decision path skeleton is first defined, namely, the initial decision path skeleton; wherein the data structure of the initial decision path skeleton adopts a structured data format. This initial decision path skeleton contains a root node (used to record the user's original question) and an empty array of decision paths. Each element in the decision path array represents a decision node, and each decision node contains the following fields: node identifier, node type, required input parameters, node status, and a list of child nodes. The node types include: precondition check nodes, standard operating procedure nodes, risk warning nodes, prohibited items nodes, and emergency response nodes.

[0051] Furthermore, based on the progressive order of the attribute items in the node type, the directed edges corresponding to the node type in the associated subgraph are queried, and the target nodes corresponding to the directed edges are sequentially filled into the decision nodes of the initial decision path skeleton. The progressive order of the attribute items represents the processing order for each node type. Finally, based on the causal relationships between the target nodes, the corresponding decision nodes are defined with parent-child relationships to obtain the decision path skeleton.

[0052] As a feasible implementation, based on the aforementioned association subgraph, the system continues to search for all directed edges with preconditions as the relationship type, and fills the decision path skeleton with the target nodes of these edges as the first decision nodes of the precondition check type. For example, if the precondition node "cooling water flow rate is greater than 10 cubic meters per hour" is found, the system generates a decision node with the node identifier "Precondition Check 1", the node type "Precondition Check", the required input parameter "cooling water flow rate", the node status "pending", and an empty child node list. Then, the system needs to continue searching for all directed edges with risk identification or prohibited items as the relationship type in the same association subgraph, and then fill the subsequent nodes of the skeleton with the target nodes of these edges, establishing parent-child relationships through the child node list field, where the precondition check node is set as the parent node, and the risk identification node and prohibited item node are set as child nodes; finally, the decision path skeleton is constructed.

[0053] Furthermore, it is necessary to traverse all attribute nodes in the decision path skeleton that are of the precondition check type, and generate follow-up query strings based on the pending status of the attribute nodes. These follow-up query strings must include at least: the string format, the required parameters of the device, and the device operation requirements. The follow-up query strings are then pushed to the user interface, and a timeout timer is started.

[0054] In one embodiment, all nodes in the current decision path skeleton that are of the type of precondition check and whose node status is pending are first traversed. For each such node, the system automatically generates a follow-up question string, which is formatted as "Please confirm" followed by the required input parameters for this node, plus the attribute node indicating whether the requirements are met. For example, for node precondition check one, the follow-up question is generated: "Please confirm whether the cooling water flow rate meets the requirement of being greater than ten cubic meters per hour." The generated follow-up question string is then pushed to the user interface, and a timeout timer is started simultaneously, with a default timeout period set to 120 seconds.

[0055] Furthermore, based on the answer status of the follow-up question string, the user-input answer string is processed to determine the feature words of the confirmation and / or negation status, so as to obtain the follow-up question feedback result information.

[0056] As one feasible implementation, the system receives the user's response string in response to the follow-up question. The wastewater treatment system then performs a simple semantic classification on the response string: if the response string contains affirmative words such as "yes," "normal," "satisfied," or "compliant," it is classified as "confirmed"; if the response string contains negative words such as "no," "abnormal," "not satisfied," or "does not comply," it is classified as "denied"; if it cannot be determined, it is classified as "unclear," and the system will repeat the follow-up question once more; thus, it obtains follow-up feedback information based on different feature word judgments.

[0057] Furthermore, if the follow-up feedback information is a confirmation status, the status of the precondition check decision node is updated to "passed". Then, using the water resources map database, starting from the precondition check decision node, the operational process features in the queried standard operating procedure decision nodes are sequentially executed; and the queried standard operating procedure decision nodes are identified as child nodes of the current node.

[0058] As a possible implementation, if the user's answer is classified as "confirmed", the system updates the status of the decision node to "passed"; then it queries the water graph database for all outgoing edges that start from the node and have a relationship type of "with subsequent steps", such as standard operation flow nodes that should be executed after the preconditions are met, and then adds these queried nodes as decision nodes of the new standard operation flow type as child nodes of the current node.

[0059] Furthermore, if the follow-up feedback is an error status, the status of the precondition check decision node is updated to "failed". Then, through a relational subgraph, directed edge queries with relational types are performed on the precondition check decision nodes to identify decision nodes with remedial measures. The target node among these decision nodes with remedial measures is then identified as the new standard operating procedure (SOP) decision node. This new SOP decision node is then designated as a child node of the current node, and a new follow-up question string is generated based on it.

[0060] As a possible implementation, if the user's answer is classified as "denied", the system updates the status of the decision node to "not passed". Then, it queries the water graph database for alternative paths, that is, it finds a directed edge with the current "precondition check" node as the starting point and the relationship type as "with remedial measures", takes its target node (e.g. "check cooling water pump") as the new standard operating procedure type decision node, adds it as a child node of the current node, and generates a new follow-up question: "Please perform the following troubleshooting steps: check cooling water pump".

[0061] Furthermore, the above steps are repeated until all follow-up feedback information is confirmed status information and no new precondition check decision nodes are generated. Then, a water problem decision path tree with water equipment operation characteristics and question-and-answer branch structure is generated.

[0062] In one embodiment, the wastewater treatment system continues to execute the above steps in a loop until the status of all decision nodes on all branches is "passed" or "failed," and no new precondition check nodes are generated. At this point, a complete, populated, branched water problem decision path tree has been generated in the system memory.

[0063] S103. Visualize and encode the decision-making path tree for water issues by specifying the types of decision nodes. Based on the event listening function, record the execution status of the user's interactive maintenance processing nodes in real time and generate a water affairs execution log array.

[0064] Specifically, the decision path tree for water issues is first recursively expanded, and a unique user interface component identifier is assigned to each decision node.

[0065] Furthermore, by using user interface component identifiers, the node type of each decision node is processed for front-end rendering. All rendered user interface components are then sorted by color depth to obtain the view area interface of the decision path tree for water-related issues. The user interface components include: precondition check type nodes, standard operating procedure type nodes, risk warning type nodes, prohibited items type nodes, and emergency response type nodes. After generating the view area interface, event listener functions are bound to each rendered view area interface.

[0066] In one embodiment, the final generated water issue decision path tree is recursively expanded, and a unique user interface component identifier is assigned to each decision node. This identifier is generated by concatenating the interface prefix with the node's identifier and the current time in milliseconds. Then, based on the type of each decision node, different user interface components are rendered on the front end. 1) For nodes with precondition checks: render as a confirmation dialog box with "Yes" and "No" buttons.

[0067] 2) For standard operation flow type nodes: render as an ordered list item with a checkbox in front of it.

[0068] 3) For risk warning type nodes: render a warning text box with a yellow background.

[0069] 4) For prohibited items type nodes: render as a prohibited icon text box with a red border.

[0070] 5) For emergency response type nodes: render them as a button with a red background and the text "Emergency Operation".

[0071] Finally, all rendered components are arranged vertically in a scrollable view area in color depth-first traversal order and then presented to the user.

[0072] Furthermore, it is necessary to use event listener functions to record and process the execution status of the interaction data of each decision node in the water issue decision path tree in real time. Then, based on the user's check mark and supplementary content actions, the action status of the checkboxes and confirmation boxes in the water issue decision path tree is captured and processed for the relevant interaction maintenance nodes.

[0073] Furthermore, based on the interaction data and action status data, a water affairs execution log array based on the water affairs problem decision path tree and its corresponding unique identifier can be generated.

[0074] As a feasible implementation, an execution log array needs to be maintained on the client side, initially empty. Then, an event listener function is bound to each rendered user interface component. When a user clicks a checkbox in a standard operation flow node, the event listener function is triggered. This function records a structured data object containing the following fields: user interface component identifier, action type (set to "checked"), trigger timestamp, and action value (true or false). When the user clicks the "Yes / No" confirmation button in a precondition check node, or clicks the "Emergency Operation" button in an emergency response node, the corresponding event is also recorded, and the user's additional text input is retrieved through a pop-up floating input box. Next, after the user completes all necessary interactive operations (i.e., all checkboxes in all standard operation flow nodes are checked, and all confirmation boxes in all precondition check nodes are clicked), the previously disabled "Complete and Submit" button on the interface becomes clickable. Then, based on the user's interaction data and action status data, a water affairs execution log array based on a water affairs problem decision path tree and its corresponding unique identifier are generated. Finally, after the user clicks the button, the client sends the entire water utility execution log data set, along with the unique identifier for this session, back to the server.

[0075] S104. Compile the experience knowledge nodes in the water affairs execution log array, and incrementally update the experience knowledge nodes with relevant success rate attribute features through the unique feature fingerprint of the log, so as to obtain the updated water affairs experience node set.

[0076] It should be noted that after receiving the water service execution log array and the unique session identifier sent by the client, the server reads the complete history of this Q&A session from the server-side cache database based on the session identifier. Then, it assembles the maintenance log data and stores the structured data as a document in an index named "Maintenance Log" within the distributed search and analysis engine. This index is pre-configured to create inverted indexes for fields such as "device identifier," "initial question text," and "timestamp" to support fast full-text search.

[0077] Specifically, the task scanning distributed search is first performed on the water affairs execution log data group, and the water affairs execution log data group with the tag type of "problem solved" is marked as the data group to be compiled.

[0078] Furthermore, water equipment identifiers, initial question text features, and execution step sequence features are extracted from the data set to be compiled. The extracted related features are then concatenated into strings to generate a unique log feature fingerprint based on semantic vector retrieval.

[0079] In one embodiment, the wastewater treatment system initiates a scheduled task that executes every ten minutes. Then, using a distributed search and analysis engine, water management execution log data sets tagged with "problem resolved" are marked as data sets to be compiled. For each data set meeting the criteria, the "device identifier," "initial problem text," and "execution step sequence" are extracted. This information is then concatenated into a string, and a semantic vector retrieval is performed on this string using an information digest algorithm to generate a unique log feature fingerprint.

[0080] Furthermore, based on the unique fingerprint of the logs, the water affairs execution log array that meets the characteristics of "problem solved" is processed into experience knowledge nodes. The node attributes of the experience knowledge node include: unique fingerprint, device identifier, fault phenomenon, solution summary, occurrence count, and success count. That is, a new node needs to be created in the water affairs graph database first, with the node label defined as "experience knowledge node," and the attributes of this experience knowledge node include: unique fingerprint, device identifier, fault phenomenon, solution summary, occurrence count, and success count.

[0081] Furthermore, it is necessary to establish a bidirectional relationship between the experience knowledge nodes and add the experience knowledge nodes and the two directed edges to the water graph database. The first bidirectional edge points from the experience knowledge node to the corresponding equipment entity node, with the relationship type "has a solution". The second edge points from the experience knowledge node to the root node of the water problem decision path tree used in this maintenance, with the relationship type "originates from".

[0082] Furthermore, a similarity check is performed on the unique feature fingerprints of the newly generated experience knowledge nodes: if the unique feature fingerprints of the new logs are the same, the success rate attribute and the number of successful resolutions attribute of the corresponding experience knowledge node are added together. If the unique feature fingerprints of the new logs are not the same, the corresponding new experience knowledge node is added to the water resources map database.

[0083] Furthermore, based on whether the unique feature fingerprints of the new logs have the same characteristics, the incremental update processing of the decision nodes in the water resources graph database is finally completed, resulting in the updated set of water resources experience nodes.

[0084] In one embodiment, when a new maintenance log is generated, the system calculates the unique fingerprint of each new maintenance log's new experience knowledge node. Then, it queries the water resources map database to check if any experience knowledge nodes with the same unique fingerprint exist; that is, it checks the unique fingerprint of the new log for similarity. If it exists, the "occurrence count" attribute value of that experience knowledge node is incremented. If the final processing result of the new log is "resolved," the "success count" attribute value of that experience knowledge node is also incremented. Finally, the incremental update processing of the decision nodes in the water resources map database is completed, resulting in an updated set of water resources experience nodes.

[0085] S105. Based on the unique feature fingerprint of the logs and the set of water affairs experience nodes, perform fast matching processing on the received new water affairs problem texts. Then, for new hydrological problem texts that cannot be quickly matched, supplementary knowledge nodes are attached to them, resulting in a water affairs patch log array.

[0086] Specifically, the unique feature fingerprint of the new log text for the new water issue is first calculated. Then, based on the unique feature fingerprint of the new log, a query matching process is performed on the set of water experience nodes in the water graph database to obtain the query matching results.

[0087] If the query result shows that a matching feature fingerprint exists, the corresponding experience knowledge node is determined as a high-priority standard operating procedure type decision node, and the standard operating procedure type decision node is inserted at the front of the water issue decision path tree corresponding to the new water issue text.

[0088] In one embodiment, a unique log fingerprint is used as a query condition to retrieve nodes labeled as "experience knowledge nodes" from the water graph database. If a match is found and the "success rate" of the experience knowledge node is greater than 70%, the process of generating the initial decision path skeleton is skipped, and the process directly enters the reverse-driven multi-round question-and-answer process, which directly utilizes the existing water problem decision path tree. Simultaneously, the "solution summary" text content within the experience knowledge node is treated as a high-priority standard operating procedure type decision node and inserted at the forefront of the water problem decision path tree.

[0089] If the query matching result is a non-matching feature fingerprint, then the new knowledge node attributes of the new water affairs problem text are compared with the original knowledge node attributes of similar water affairs problem texts to determine the supplementary knowledge data object.

[0090] Furthermore, the aforementioned supplementary knowledge data objects need to be identified as patch knowledge nodes in the water resources graph database, based on the graph structure position of similar experience knowledge nodes in the water resources graph database. Then, the patch knowledge nodes are attached to the similar experience knowledge nodes, ultimately resulting in a water resources patch log array. Here, the similar experience knowledge nodes are compiled from the water resources execution log array corresponding to similar water resources problem texts that meet the characteristics of problem resolution.

[0091] In one embodiment, the existing knowledge node attributes corresponding to equipment parameters in similar water-related problem texts are first queried from the water resources map database. If the new threshold in the new knowledge is inconsistent with the threshold stored in the existing knowledge node attributes, then the conflict detection logic is triggered. In other words, the wastewater treatment system does not overwrite or delete existing knowledge nodes; instead, it creates a new patch knowledge node. The patch knowledge node's tag attributes include: the original node identifier, the new value, and the applicable conditions. Simultaneously, two relationships are established in the water graph database: the first relationship points from the patch knowledge node to the original knowledge node, with the relationship type being the overwrite attribute; the second relationship points from the patch knowledge node to a condition node (e.g., seasonal conditions), with the relationship type being the applicable attribute. Then, when generating water management decision paths, the wastewater treatment system automatically selects whether to apply the original knowledge node or the patch knowledge node based on the current environmental conditions. Therefore, it is necessary to attach the patch knowledge node to the similar experience knowledge nodes based on their graph structure positions in the water graph database, resulting in a water management patch log array.

[0092] S106. Generate a water affairs operation and maintenance report based on the water affairs execution log array and the water affairs patch log array.

[0093] Specifically, the front-end interface of the wastewater treatment system can be used to aggregate the water management execution log array and the water management patch log array retrieved from the back-end distributed search.

[0094] Furthermore, the frequency of equipment failures, the number of times and success rate of water-related problem decision-making paths, and the number and duration of maintenance personnel handling the extracted aggregated data are statistically processed to obtain statistical results. Finally, the statistical results are transmitted to the front-end interface and a water-related operation and maintenance report is generated.

[0095] As a feasible implementation method, the frequency of equipment failures is grouped and statistically analyzed according to the equipment identification field; the number of times and success rate of water problem decision paths are grouped and statistically analyzed according to the decision tree identifier field; and the number of maintenance personnel handling cases and the duration are grouped and statistically analyzed according to the user identifier field.

[0096] In addition, embodiments of this application also provide a knowledge base application system based on a large model, such as... Figure 2 As shown, the knowledge base application system 200 based on a large model specifically includes: Database module 210 is used to extract multimodal content and hierarchical semantic decomposition of water-related files uploaded by the client to construct a water affairs graph database; The water affairs decision module 220 is used to obtain a water affairs decision path tree by using the water affairs graph database to perform multi-round question-and-answer reverse driving and path dynamic growth based on the decision node status of the initial water affairs question text input by the user. The listening and recording module 230 visualizes and encodes the decision-making path tree of water affairs nodes with relevant decision-making node types, and records the execution status of user interaction maintenance processing nodes in real time based on the event listening function, generating a water affairs execution log array. The water affairs experience matching module 240 is used to compile and process experience knowledge nodes in the water affairs execution log array, and incrementally update the experience knowledge nodes with success rate attribute features through the unique feature fingerprint of the logs to obtain an updated set of water affairs experience nodes; based on the unique feature fingerprint of the logs and the set of water affairs experience nodes, it performs fast matching processing on the problem fingerprint of the received new water affairs problem text; and it performs supplementary knowledge node mounting processing on new hydrological problem texts that cannot be quickly matched to obtain a water affairs patch log array. The backend data integration module 250 generates a water affairs operation and maintenance report based on the water affairs execution log array and the water affairs patch log array.

[0097] This application's embodiments generate a decision path tree through a reverse-driven, multi-round question-and-answer process, providing maintenance personnel with a complete and executable solution from risk identification to execution of procedures and emergency response, rather than fragmented text snippets. This significantly reduces the difficulty and risk of on-site decision-making. Furthermore, by compiling maintenance logs into experience knowledge nodes and dynamically updating them to the knowledge graph, each maintenance operation provides the system with learning and training opportunities, enabling continuous accumulation and intelligent reuse of operational experience and solving the problem of traditional experience loss. Conflict detection and versioning technologies allow newly added knowledge to take effect without disrupting the existing authoritative document structure and can be dynamically selected based on contextual conditions, enhancing the system's robustness. Moreover, complete operational records and frequency statistics of decision path usage provide managers with quantitative analysis data, enabling the identification of frequently used decision paths or paths with low success rates.

[0098] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0099] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0100] The above description is merely an embodiment of this application and is not intended to limit this application. For those skilled in the art, various modifications and variations can be made to the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of this application should be included within the scope of the claims of this application.

Claims

1. A knowledge base application method based on a large model, characterized in that, The method includes: The water-related files uploaded by the client are subjected to multimodal content extraction and hierarchical semantic decomposition to construct a water affairs graph database; Using the water resources map database, the initial water resources question text input by the user is subjected to multi-round question-and-answer reverse driving and dynamic path growth based on the decision node state to obtain a water resources question decision path tree; The water issue decision path tree is visualized and encoded with relevant decision node types, and based on the event listening function, the execution status of the user's interactive maintenance processing nodes is recorded in real time to generate a water execution log array. The water affairs execution log array is processed by compiling experience knowledge nodes, and the experience knowledge nodes are incrementally updated with relevant success rate attribute features by using the unique feature fingerprint of the log, so as to obtain the updated set of water affairs experience nodes. Based on the unique feature fingerprint of the log and the set of water affairs experience nodes, the received new water affairs problem text is subjected to fast matching of problem fingerprints; and new hydrological problem texts that cannot be quickly matched are subjected to supplementary knowledge node mounting processing to obtain a water affairs patch log array. A water affairs operation and maintenance report is generated based on the water affairs execution log array and the water affairs patch log array.

2. The knowledge base application method based on a large model according to claim 1, characterized in that, The water-related files uploaded by the client are subjected to multimodal content extraction and hierarchical semantic decomposition to construct a water resources graph database, specifically including: The water-related documents are uploaded to the wastewater treatment system via the client; the water-related documents include at least the following: 4G pressure sensor instruction manual, magnetic levitation blower operation manual, electromagnetic flow meter instruction manual, water pump operation and use manual, aeration device operation manual, and sludge dewatering machine operation and use manual. Extract the Internet media type identifier and raw byte stream data of each water-related file; based on the parser corresponding to the Internet media type identifier and the raw byte stream data, perform content extraction processing on the water-related files under the file format type to obtain an initial paragraph sequence; By using a pre-trained deep learning named entity recognition model, the initial paragraph sequence is subjected to named entity recognition and annotation of text entity sequences to obtain a water equipment entity sequence; wherein, the water equipment entity sequence includes at least: text content, tag type, and the starting character position and ending character position in the paragraph; Using a relation type dictionary, dependency parsing is performed on the semantic text of the water equipment entity sequence. Based on the dependency relations and core verbs in the sentences, relation triples are generated for the semantic text to determine the entity semantic relations of each water equipment entity. The relation type dictionary includes: water equipment has parameters, standard operating procedures, prohibited items, risk identification, preconditions, and emergency response measures. Each relation triple in the entity semantic relation is defined as a directed edge, and each entity in the entity semantic relation is defined as a node; and the water resources map database is constructed based on the directed edges and the nodes; wherein, the storage attributes of each node in the water resources map database include: entity text content, entity type, source document path, and source document page number.

3. The knowledge base application method based on a large model according to claim 2, characterized in that, After constructing the water resources graph database based on the directed edges and the nodes, the method further includes: A set of dimension labels is attached to each entity node in the water resources map database; wherein, the initial value of each dimension label is empty, and the label dictionary of the dimension label includes: equipment type, operation stage, risk level, and applicable working conditions; Based on the tag dictionary and through the rule engine of the sewage treatment system, the tag type of the entity text content is automatically assigned to obtain the tag set of each entity node. The assigned label set is serialized to obtain extended label attribute features; these extended attribute features are then extended and stored with each node in the water resources map database to obtain a water resources map database with additional dimension labels.

4. The knowledge base application method based on a large model according to claim 1, characterized in that, Before obtaining a water issue decision path tree by using the water graph database and performing multi-round question-and-answer reverse-drive and dynamic path growth based on decision node states on the initial water issue text input by the user, the method further includes: By using a deep learning named entity recognition model, the initial water-related issue text is processed to extract the features of the issue entities, resulting in a core entity set; and the issue entity set with the tag type of abnormal state is identified. Using the set of problem entities, the water resources map database is processed by matching the tag types of the main nodes, and based on the identified main nodes, adjacent nodes and associated subgraphs that are directly connected to the main nodes through any relationship are determined; wherein, the main node is the node that is matched with the problem entities in the set of problem entities. Based on the set of problem entities, an initial decision path skeleton is constructed; wherein the initial decision path skeleton includes a root node and an empty decision path array, and each element in the decision path array represents a decision node, and the node type of each decision node includes at least: a prerequisite check node, a standard operating procedure node, a risk warning node, a prohibited item node, and an emergency response node. Based on the progressive order of the attribute items in the node type, the directed edges corresponding to the node type in the associated subgraph are queried, and the target nodes corresponding to the directed edges are sequentially filled into the decision nodes of the initial decision path skeleton; wherein, the progressive order of the attribute items is the process processing order of each node type; Based on the causal relationship between the target nodes, the corresponding decision nodes are defined with a parent-child relationship to obtain the decision path skeleton.

5. The knowledge base application method based on a large model according to claim 4, characterized in that, Using the aforementioned water resources map database, the initial water resources question text input by the user is subjected to multi-round question-and-answer reverse driving and dynamic path growth based on the decision node state to obtain a water resources question decision path tree, specifically including: Traverse all attribute nodes of type precondition check in the decision path skeleton, and generate follow-up question strings based on the pending status of the attribute nodes; wherein, the follow-up question strings include at least: string format, required parameters of the device, and device operation requirements; The question string is pushed to the user interface, and a timeout timer is started; Based on the answer status of the follow-up question string, the user-input answer string is processed to determine the feature words of the confirmation status and / or negation status, and the follow-up question feedback result information is obtained. If the follow-up feedback information is a confirmation status, then the status of the precondition check decision node is updated to the passed status; and through the water map database, starting from the precondition check decision node, the operation process features in the queried standard operation process decision nodes are sequentially executed; and the queried standard operation process decision nodes are determined as child nodes of the current node. If the follow-up feedback information is an error status, the status of the precondition check decision node is updated to a failed status; and through the association subgraph, a directed edge query with relational type is performed on the precondition check decision node to determine the decision node with remedial measures. The target node in the decision node with remedial measures is identified as the new standard operating procedure decision node; the new standard operating procedure decision node is identified as the child node of the current node, and a new follow-up question string is generated based on the new standard operating procedure decision node. The process is repeated until all follow-up questions and feedback results are confirmed and no new precondition check decision nodes are generated. Then, a water issue decision path tree with water equipment operation characteristics and a question-and-answer branch structure is generated.

6. The knowledge base application method based on a large model according to claim 1, characterized in that, The water affairs decision path tree is visualized and encoded with relevant decision node types. Based on an event listener function, the execution status of user interaction maintenance processing nodes is recorded in real time, generating a water affairs execution log array, specifically including: The water issue decision path tree is recursively expanded, and a unique user interface component identifier is assigned to each decision node. Using the user interface component identifier, the node type of each decision node is rendered in the front end, and all rendered user interface components are sorted by color depth to obtain the view area interface of the decision path tree for water issues; wherein, the user interface components include: precondition check type nodes, standard operation process type nodes, risk warning type nodes, prohibited items type nodes, and emergency response type nodes. After generating the view area interface, the event listener function is bound to each rendered view area interface; The event listening function is used to record and process the execution status of the interaction data of each decision node in the water issue decision path tree in real time; based on the user's check mark action and supplementary content action, the action status of the check mark and confirmation box in the water issue decision path tree is captured and processed for the interaction maintenance processing node. Based on the interaction data and action status data, an array of water affairs execution logs and corresponding unique identifiers are generated based on the water affairs problem decision path tree.

7. The knowledge base application method based on a large model according to claim 1, characterized in that, The water affairs execution log array is processed by compiling experience knowledge nodes, and the experience knowledge nodes are incrementally updated with relevant success rate attribute features using the unique feature fingerprint of the logs, resulting in an updated set of water affairs experience nodes, specifically including: Perform a distributed search of the water affairs execution log data group and mark the water affairs execution log data group with the tag type "problem solved" as the data group to be compiled; Extract water equipment identifiers, initial question text features, and execution step sequence features from the data group to be compiled; and concatenate the extracted associated features into strings to generate the unique feature fingerprint of the log after semantic vector retrieval. Based on the unique feature fingerprint of the logs, the water affairs execution log array that meets the characteristics of "problem solved" is processed into experience knowledge nodes; wherein, the node attributes of the experience knowledge nodes include: unique feature fingerprint, device identifier, fault phenomenon, solution summary, occurrence number, and success number; The experience knowledge nodes are linked by two directed edges, and the experience knowledge nodes and the two directed edges are added to the water affairs graph database. The first directed edge in the two directed edges points from the experience knowledge node to the corresponding equipment entity node, and the second directed edge points from the experience knowledge node to the root node of the water affairs problem decision path tree used in this maintenance. Perform a similarity check on the unique feature fingerprint of the new log of subsequent new experience knowledge nodes; If the unique feature fingerprints of the new logs are identical, the success rate attribute and the number of successful resolutions attribute of the corresponding experience knowledge node will be added together for calculation. If the unique feature fingerprints of the new logs are not identical, the corresponding new experience knowledge nodes will be added to the water resources map database. Based on whether the unique feature fingerprint of the new log has the same characteristics, the incremental update process of the decision nodes in the water affairs graph database is completed to obtain the updated set of water affairs experience nodes.

8. The knowledge base application method based on a large model according to claim 1, characterized in that, Based on the unique feature fingerprint of the log and the set of water affairs experience nodes, the received new water affairs problem text is subjected to fast matching processing of problem fingerprints; Then, new hydrological issue texts that cannot be quickly matched are supplemented with knowledge nodes, resulting in a water affairs patch log array, which specifically includes: Calculate the unique feature fingerprint of the new log text related to the new water issue; Based on the unique feature fingerprint of the new log, a query matching process is performed on the set of water affairs experience nodes in the water affairs graph database to obtain the query matching result; If the query matching result is that a matching feature fingerprint exists, then the corresponding experience knowledge node is determined as a high-priority standard operating procedure type decision node, and the standard operating procedure type decision node is inserted at the front of the water issue decision path tree corresponding to the new water issue text. If the query matching result is a non-matching feature fingerprint, then the new knowledge node attributes of the new water affairs problem text are compared with the original knowledge node attributes of similar water affairs problem texts to determine the supplementary knowledge data object. The supplementary knowledge data object is identified as a patch knowledge node in the water resources graph database, and based on the graph structure position of the similar experience knowledge node in the water resources graph database, the patch knowledge node is mounted to the similar experience knowledge node to obtain the water resources patch log array; wherein, the similar experience knowledge node is compiled from the water resources execution log array that corresponds to similar water resources problem text and meets the characteristics of problem resolution.

9. The knowledge base application method based on a large model according to claim 1, characterized in that, Based on the water utility execution log array and the water utility patch log array, a water utility operation and maintenance report is generated, specifically including: The front-end interface of the sewage treatment system aggregates the water affairs execution log array and the water affairs patch log array retrieved by the back-end distributed search. The frequency of equipment failures, the number of times and success rate of water-related problem decision-making paths, and the number and duration of maintenance personnel handling issues in the extracted aggregated data are statistically processed to obtain statistical results. The statistical results are transmitted to the front-end interface, and the water affairs operation and maintenance report is generated.

10. A knowledge base application system based on a large model, characterized in that, The large-model-based knowledge base application system has instructions that can be executed by at least one processor, so that at least one processor can execute a method for a large-model-based knowledge base application system according to any one of claims 1-9.

Images