A Method and System for Document Processing in Power Transmission and Transformation Scenarios Based on Neuromorphic Memory Mechanism

Abstract

This invention relates to the field of intelligent document processing technology, specifically to a method and system for document processing in power transmission and transformation scenarios based on a brain-like memory mechanism. The method includes the following steps: generating a query plan based on query intent; recalling relevant documents from a hierarchically modeled power transmission and transformation document knowledge base; and extracting and reasoning information from the recalled documents using a dual-path parallel processing mechanism and a hierarchical abstraction processing mechanism; filling in a structured form according to a predefined template to generate preliminary structured output; simultaneously transforming the extracted and reasoned information into interpretable reasoning chain records; generating memory cases based on the reasoning chain records and confidence assessment results; and storing and optimizing the memory cases using a brain-like memory mechanism. This method possesses the ability to handle complex multi-step reasoning tasks, significantly improves reasoning accuracy, automates document information extraction, and greatly reduces manual processing time.

Description

Technical Field

[0001] This invention relates to the field of intelligent document processing technology, and more specifically, to a document processing method and system for power transmission and transformation scenarios based on a brain-like memory mechanism. Background Technology

[0002] Extracting structured information from documents has always been an important and complex issue in document processing within the power transmission and transformation industry. In power transmission and transformation scenarios, when compiling a power outage maintenance plan document based on a document titled "Field Survey Record of a Certain Inspection Line," most of the information needs to be extracted from the field survey record, while some information, such as the responsible person's information, needs to be extracted from databases / knowledge bases or other files.

[0003] To address the above challenges, the power transmission and transformation industry currently employs the following technical solutions for document information processing:

[0004] The three approaches are: 1) using traditional keyword retrieval methods, which rely on manually pre-set keyword databases for document retrieval based on keyword matching; 2) using basic large language model methods, which directly use general large models for document understanding and information extraction; and 3) using simple RAG (Retrieval Augmentation) technology, which segments documents for vector retrieval and then combines it with large models to generate answers.

[0005] The existing technology has the following drawbacks:

[0006] 1. Low accuracy of information extraction: Traditional methods cannot understand complex power transmission and transformation professional terms and business logic relationships, and cannot handle complex hierarchical relationships and cross-document reference relationships in power transmission and transformation documents, resulting in a high error rate in information extraction;

[0007] 2. Lack of memory capacity: Existing methods lack a memory indexing mechanism similar to the hippocampus in the human brain. Each query requires re-retrieval and analysis, making it impossible to accumulate and reuse historical query experience like the human brain, resulting in low efficiency.

[0008] 3. Lack of neuroplasticity: Traditional systems cannot simulate the brain's synaptic plasticity mechanism, lack the ability to continuously learn and adaptively optimize, and cannot continuously improve performance based on user experience.

[0009] Therefore, a method and system for document processing in power transmission and transformation scenarios based on a brain-like memory mechanism are provided. Summary of the Invention

[0010] The purpose of this invention is to provide a method and system for document processing in power transmission and transformation scenarios based on a brain-like memory mechanism, so as to solve the problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention aims to provide a document processing method for power transmission and transformation scenarios based on a brain-like memory mechanism, comprising the following steps:

[0012] S1. Receive the user's query request for the target document, generate the query intent based on the query request using a pre-trained large language model in the power transmission and transformation field, and generate a query plan based on the query intent.

[0013] S2. Based on the query plan, relevant documents are retrieved from the hierarchically modeled power transmission and transformation document knowledge base, and information extraction and reasoning are performed on the retrieved relevant documents using a dual-path parallel processing mechanism and a hierarchical abstract processing mechanism.

[0014] S3. The information obtained from extraction and reasoning is structured and filled according to the predefined power transmission and transformation industry document template to generate preliminary structured output. At the same time, a brain-like evidence accumulation decision-making mechanism is adopted to transform the information obtained from extraction and reasoning into an interpretable reasoning chain record.

[0015] S4. A multi-source evidence accumulation mechanism is used to assess the confidence of the initial structured output, thereby optimizing the inference chain record. Memory cases are generated based on the inference chain record and the confidence assessment results. A brain-like memory mechanism is used to store and optimize the memory cases.

[0016] As a further improvement to this technical solution, in step S1, after receiving the user's query request for the target document, the system first retrieves historical query cases in the brain-like memory mechanism and calculates the confidence level; if the confidence level of the retrieved similar cases is not lower than the preset confidence threshold, the system calls the memory result corresponding to the similar cases as the output and terminates the subsequent steps; otherwise, step S2 is executed.

[0017] As a further improvement to this technical solution, in S1, after the query plan is executed, a reflection and checking mechanism is introduced for optimization. Specifically, through logical consistency verification and cross-validation, inconsistencies or obvious logical errors are identified and corrected.

[0018] As a further improvement to this technical solution, in step S2, the specific steps for recalling target documents based on a hierarchically modeled power transmission and transformation document knowledge base are as follows:

[0019] S21. Construct a multi-level knowledge base structure for power transmission and transformation documents based on organizational structure, functional classification, and time range.

[0020] S22. First, perform coarse-grained retrieval at a high level to determine the relevant document categories, and then perform fine-grained document retrieval within the determined categories.

[0021] S23. Identify the reference relationships, dependencies, and hierarchical relationships between documents, and construct a document relationship graph;

[0022] S24. Based on the document association graph, perform association expansion retrieval on the documents retrieved through fine-grained retrieval, and add related documents that have a dependency relationship with the recalled documents to the recall scope.

[0023] As a further improvement to this technical solution, in S2, the dual-path parallel processing mechanism includes a ventral path processing sub-step, a dorsal path processing sub-step, and a path fusion sub-step; the hierarchical abstract processing mechanism includes primary feature extraction, intermediate pattern recognition, and advanced semantic understanding.

[0024] As a further improvement to this technical solution, in step S3, the specific steps for structurally filling the extracted and inferred information according to a predefined power transmission and transformation industry document template to generate preliminary structured output are as follows:

[0025] Select a structured output template based on the query type and target document format. Organize and populate the extracted and inferred information on the template according to the predefined structured format. After filling, verify the completeness of the output results.

[0026] As a further improvement to this technical solution, in step S3, the brain-like evidence accumulation decision-making mechanism is used to transform the extracted and reasoned information into an interpretable reasoning chain record. The specific steps are as follows:

[0027] S31. Extract and collect evidence to verify the preliminary structured output, and label the information source for each piece of evidence;

[0028] S32. Based on the reliability, timeliness, and relevance of the evidence source to the current query, assign dynamic weights to each piece of evidence collected; at the same time, identify and label evidence conflicts from different information sources.

[0029] S33. Starting with the initial structured output, construct a logical skeleton composed of multiple reasoning nodes to generate a reasoning chain, and attach weighted evidence and evidence conflicts to the corresponding nodes of the reasoning chain.

[0030] S34. Identify the risk level and problem nodes of the inference chain decision, quantify the uncertainty of the problem nodes in the inference chain, and mark and record the quantified problem nodes in the inference chain.

[0031] S35. Integrate all the information from steps S31 to S34 to generate a structured, interpretable reasoning chain record.

[0032] As a further improvement to this technical solution, in step S4, a multi-source evidence accumulation mechanism is used to assess the confidence level of the preliminary structured output, specifically as follows:

[0033] Based on the evidence weights attached to the nodes, the cumulative evidence strength is calculated. Depending on whether the cumulative evidence strength reaches the confidence threshold preset for decision-making, it is determined whether to implement an early stop strategy or a progressive deep analysis strategy. The adopted decision strategy, the cumulative evidence strength, and the reasoning for the decision are recorded in the inference chain.

[0034] As a further improvement to this technical solution, in step S4, the brain-like memory mechanism simulates the hippocampal indexing mechanism, neuroplasticity mechanism, and memory retrieval and matching to store and optimize memory cases. The specific steps are as follows:

[0035] S41. Simulate the hippocampal indexing mechanism to generate a memory case for each successful historical query case, store the memory cases in a brain-like memory mechanism, and establish a multi-dimensional index for the memory cases;

[0036] S42. For new query requests, use multi-dimensional indexes and multi-modal similarity matching to perform memory retrieval, obtain highly similar memory cases, and then reuse the results of those cases.

[0037] S43. Simulate the neural plasticity mechanism and dynamically adjust the connection weights of memory cases based on user feedback and query confidence.

[0038] S44. Extract the query logic for complex query paths, generate a general query template, and update it to the brain-like memory mechanism.

[0039] On the other hand, the present invention provides a power transmission and transformation scenario document processing system based on a brain-like memory mechanism, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the power transmission and transformation scenario document processing method based on any one of the above-described brain-like memory mechanisms.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] 1. This document processing method and system for power transmission and transformation scenarios based on brain-like memory mechanism achieves long-term storage and efficient retrieval of document information by introducing a brain-like memory mechanism. It overcomes the drawback of traditional methods that require starting from the beginning for each query, and the retrieval efficiency and accuracy are significantly better than traditional vector retrieval methods.

[0042] Meanwhile, by combining brain-like dual-pathway information processing and evidence accumulation decision-making mechanisms, the system has the ability to handle complex multi-step reasoning tasks, its reasoning accuracy has been greatly improved, and it has demonstrated metacognitive capabilities, achieving explainable intelligent reasoning and autonomous error correction.

[0043] 2. This document processing method and system for power transmission and transformation scenarios based on brain-like memory mechanism automates document information extraction, significantly reduces manual processing time, and continuously improves work efficiency through continuous self-optimization. Attached Figure Description

[0044] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Example 1: Please refer to Figure 1 As shown, this embodiment provides a document processing method for power transmission and transformation scenarios based on a brain-like memory mechanism. The core of this method lies in simulating the cognitive mechanism of the human brain and constructing a complete brain-like intelligent document processing framework.

[0047] The dual-path parallel processing mechanism simulates the ventral and dorsal pathways of the human brain, processing semantic content and spatial-temporal information respectively; a multi-level document knowledge base is constructed based on the characteristics of the power transmission and transformation industry to achieve accurate and efficient document retrieval; the brain-like evidence accumulation decision-making mechanism simulates the decision-making mechanism of the prefrontal cortex, providing interpretable reasoning processes and confidence assessments; and the brain-like memory mechanism simulates the hippocampal memory system to achieve continuous learning and optimization of knowledge.

[0048] Through the coordinated work of the above modules and mechanisms, the system can understand, reason, and make decisions like a human expert, significantly improving the level of intelligence in document processing.

[0049] Includes the following steps:

[0050] S1. Receive the user's query request for the target document, generate the query intent based on the query request using a pre-trained large language model in the power transmission and transformation field, and generate a query plan based on the query intent.

[0051] The query intent refers to the structured intent representation output by the model, specifically a multi-field vector in JSON format (example fields include: task type, time range, organizational unit, expected output format, etc.).

[0052] A query plan is an executable plan (examples of which include retrieval strategy, recall candidate document set, and subsequent processing order) generated based on query intent and used for retrieval and processing. It can be expressed as a serialized instruction set or a schedule table.

[0053] Pre-trained large language models for the power transmission and transformation field refer to models that are fine-tuned or pre-trained using power transmission and transformation field corpora on the basis of general large language models. They are used to generate query intent and support query plan generation and text understanding.

[0054] In step S1, after receiving a user's query request for the target document, the system first retrieves historical query cases from the brain-like memory mechanism and calculates their confidence scores. If the confidence score of a retrieved similar case is not lower than a preset confidence threshold, the system calls the memory result corresponding to the similar case as output and terminates subsequent steps; otherwise, step S2 is executed. The preset confidence threshold is a configurable numerical parameter set or adjusted by the system administrator or the online learning mechanism. A successful match is considered achieved when the confidence score of a retrieved candidate case is greater than or equal to this threshold.

[0055] In S1, after the query plan is executed, a reflective checking mechanism is introduced for optimization. Specifically, through logical consistency verification and cross-validation, inconsistencies or obvious logical errors are identified and corrected.

[0056] The logical consistency verification module is used to detect semantic and format contradictions or inconsistencies within the inference chain or output. The cross-validation module is used to compare the current result with other independent evidence in the knowledge base or external data sources to identify deviations. When inconsistencies or deviations are detected, the query plan or execution parameters are corrected and the corresponding processing steps are re-executed.

[0057] S2. Based on the query plan, relevant documents are retrieved from the hierarchically modeled power transmission and transformation document knowledge base, and information extraction and reasoning are performed on the retrieved relevant documents using a dual-path parallel processing mechanism and a hierarchical abstract processing mechanism.

[0058] In S2, based on the query plan, the specific steps for retrieving target documents in a hierarchically modeled power transmission and transformation document knowledge base are as follows:

[0059] S21. Construct a multi-level knowledge base structure for power transmission and transformation documents based on organizational structure, functional classification, and time range.

[0060] S22. First, perform coarse-grained retrieval at a high level to determine relevant document categories, and then perform fine-grained document retrieval within the determined categories; coarse-grained retrieval uses organization / time-based, topic classifiers (such as SVM / NN); fine-grained retrieval is based on vector retrieval (cosine similarity) or inverted index (BM25).

[0061] S23. Identify the reference relationships, dependencies, and hierarchical relationships between documents, and construct a document relationship graph;

[0062] The graph construction employs a graph neural network-based approach: nodes represent documents or document fragments, and node feature vectors... Edges represent relationships between documents, including references, dependencies, and hierarchies; graph updates: ;in, For the first The original data content of each document; For the first The feature vector of each node; For document embedding functions, For graph neural network models; For the entire document's relational network; For connecting nodes and nodes The edge; To represent the graph state after the graph neural network update, it contains the latest representations of nodes and edges in the feature space, which are used for subsequent retrieval and reasoning tasks;

[0063] S24. Based on the document association graph, perform association expansion retrieval on the documents retrieved through fine-grained retrieval, and also add related documents that have a dependency relationship with the recalled documents to the recall scope.

[0064] The hierarchical design significantly narrows the search scope, improves recall accuracy and efficiency, and supports the extraction and verification of cross-document related information.

[0065] In S2, the dual-path parallel processing mechanism includes a ventral pathway processing sub-step, a dorsal pathway processing sub-step, and a pathway fusion sub-step.

[0066] Among them, the ventral path processing sub-step is used to specifically process the "what is" information, and is responsible for equipment identification, fault type classification, and semantic content extraction of operating procedures in power transmission and transformation scenarios.

[0067] The back-side pathway processing sub-step is specifically used to handle "where" and "how" information, and is responsible for spatial location positioning, time series analysis, operation step planning, and spatial-temporal information processing.

[0068] The pathway fusion sub-step is used to integrate the semantic content information extracted from the ventral pathway with the spatial-temporal information extracted from the dorsal pathway at a higher cognitive level, forming a complete document understanding context and decision support foundation.

[0069] The hierarchical abstraction processing mechanism includes primary feature extraction, intermediate pattern recognition, and advanced semantic understanding;

[0070] Among them, primary feature extraction is used to simulate the V1 region of the visual cortex to extract basic features of documents such as keywords, entities, and relationships; intermediate pattern recognition is used to simulate the V2-V4 regions of the visual cortex to identify complex business patterns and rule structures (such as template matching and event / process extraction); and advanced semantic understanding is used to simulate the temporal cortex to perform abstract concept understanding and business logic reasoning (such as causal relationship recognition, norm / standard determination, and reasoning chain construction).

[0071] S3. The information obtained from extraction and reasoning is structured and filled according to the predefined power transmission and transformation industry document template to generate preliminary structured output. At the same time, a brain-like evidence accumulation decision-making mechanism is adopted to transform the information obtained from extraction and reasoning into an interpretable reasoning chain record.

[0072] In S3, the specific steps for filling in the extracted and inferred information according to the predefined power transmission and transformation industry document template to generate preliminary structured output are as follows:

[0073] Select a structured output template based on the query type and target document format. Organize and populate the extracted and inferred information on the template according to the predefined structured format. After population, verify the completeness of the output results, identify missing information and provide supplementary suggestions to ensure that the output format meets the power transmission and transformation industry standards and business process requirements.

[0074] In S3, a brain-like evidence accumulation decision-making mechanism is adopted. The specific steps for transforming the information obtained from extraction and reasoning into interpretable reasoning chain records are as follows:

[0075] S31. Extract and collect evidence to verify the preliminary structured output, and label the information source for each piece of evidence; specifically, collect evidence from the target document content, historical cases in brain-like memory mechanisms, and knowledge bases in the power transmission and transformation field; evidence includes, but is not limited to, document text paragraphs, table items, illustration annotations, and log entries; information sources include, but are not limited to, document ID, paragraph offset, timestamp, and source system.

[0076] S32. Based on the reliability, timeliness, and relevance of the evidence source to the current query, assign dynamic weights to each piece of evidence collected; at the same time, identify and label evidence conflicts from different information sources.

[0077] Evidence is the basic unit that supports or refutes a conclusion (document fragments, table values, drawings, log entries, etc.), and each piece of evidence is assigned a weight.

[0078] The dynamic weights are assigned according to predefined weighting rules or learned weighting models. These weighting rules consider at least the evidence source category (e.g., normative documents, on-site records, historical cases), evidence timestamps, and the semantic similarity between the evidence and the query intent. Evidence conflict identification can be based on semantic consistency verification or numerical differences exceeding a tolerance threshold.

[0079] S33. Starting with the initial structured output, construct a logical skeleton composed of multiple reasoning nodes to generate a reasoning chain, and attach weighted evidence and evidence conflicts to the corresponding nodes of the reasoning chain.

[0080] Weighted evidence and detected evidence conflicts can serve as the basis for supporting or challenging inference nodes. The inference chain consists of several nodes, including fact nodes, hypothesis nodes, evidence nodes, and decision nodes. Each node records its conclusion, a list of cited evidence IDs, and a node confidence score. When constructing the inference chain, the evidence in S32 is attached to the evidence list of relevant nodes based on semantic and factual matching, and the evidence weights are used to calculate the node confidence.

[0081] S34. Identify the risk level and problem nodes of the inference chain decision, quantify the uncertainty of the problem nodes in the inference chain, and mark and record the quantified problem nodes in the inference chain.

[0082] The specific steps for performing uncertainty quantification include:

[0083] Identify the potential risk types and levels of the power transmission and transformation business scenarios involved in the current decision-making process;

[0084] Based on the risk level, the evidence accumulation threshold required for this decision will be automatically adjusted up or down.

[0085] Problem nodes are those with weak, missing, or seriously conflicting evidence. Uncertainty is quantified and scored for problem nodes, and a risk warning text description is provided.

[0086] S35. Integrate all the information from steps S31 to S34 to generate a structured, interpretable reasoning chain record.

[0087] This brain-based evidence accumulation decision-making mechanism achieves dynamic speed-accuracy trade-off optimization, and possesses risk perception and adaptive adjustment capabilities, supporting multi-level decision strategy selection.

[0088] S4. A multi-source evidence accumulation mechanism is used to assess the confidence of the initial structured output, thereby optimizing the inference chain record. Memory cases are generated based on the inference chain record and the confidence assessment results. A brain-like memory mechanism is used to store and optimize the memory cases. The memory cases specifically include the query plan, recall path, inference chain, retrieval results, and confidence.

[0089] In S4, a multi-source evidence accumulation mechanism is used to assess the confidence level of the initial structured output, specifically as follows:

[0090] Based on the evidence weights attached to the nodes, the cumulative evidence strength is calculated. Depending on whether the cumulative evidence strength reaches the confidence threshold preset for the decision, it is determined whether to implement an early stop strategy or a gradual deep analysis strategy. The adopted decision strategy, the cumulative evidence strength, and the reasoning for the decision are recorded in the inference chain. The confidence threshold preset for the decision is 0.8.

[0091] Therefore, an optimized structured interpretable reasoning chain record is generated, which includes the final conclusion, logical derivation path, all cited evidence and their weights, explanation of evidence conflicts, decision-making strategy adopted, cumulative evidence strength and decision rationale, risk assessment and uncertainty labeling.

[0092] Furthermore, based on whether the accumulated evidence strength reaches the preset confidence threshold for decision-making, the decision is made to determine whether to implement an early cessation strategy or a gradual in-depth analysis strategy.

[0093] Real-time calculation of the cumulative strength of evidence supporting the core conclusions;

[0094] If the strength reaches or exceeds the high confidence threshold set for rapid response within a preset time, the early stop mechanism is triggered to terminate further search, and the accumulated evidence strength and the computing resources saved when the early stop is triggered are recorded in the inference chain.

[0095] If the strength does not reach the high confidence threshold, a progressive decision-making process is initiated, involving a more in-depth evidence retrieval and analysis, and this in-depth analysis process is recorded as an extension of the reasoning chain.

[0096] The early stop strategy and the progressive deep analysis strategy are adaptively selected by the system based on the strength of accumulated evidence, the potential risk level of the current decision and resource constraints, and the selected strategy and its triggering conditions are recorded in the inference chain log.

[0097] In S4, the brain-like memory mechanism simulates the hippocampal indexing mechanism, neuroplasticity mechanism, and memory retrieval and matching to store and optimize memory cases. The specific steps are as follows:

[0098] S41. Simulate the hippocampal indexing mechanism to generate a memory case for each successful historical query case, store the memory cases in a brain-like memory mechanism, and build a multi-dimensional index for the memory cases; the brain-like memory mechanism includes a temporary memory cache and a long-term memory area. The memory cases are first stored in the temporary memory cache of the brain-like memory mechanism; when the confidence assessment result of its association is higher than the preset standard, it is then transferred to the long-term memory area of ​​the brain-like memory mechanism; the multi-dimensional index includes semantic index, structural index, temporal index, and association index;

[0099] S42. For new query requests, multi-dimensional indexes are used to perform memory retrieval through multi-modal similarity matching to obtain highly similar memory cases, and then the results of these cases are reused; multi-modal similarity matching adopts a weighted fusion strategy;

[0100] S43. Simulate the neural plasticity mechanism and dynamically adjust the connection weights of memory cases based on user feedback and query confidence; perform long-term enhancement on frequently used and highly confident memory paths, and perform long-term inhibition on memory paths with low usage frequency or persistently low confidence.

[0101] S44. Extract the query logic for complex query paths, generate a general query template, and update it to the brain-like memory mechanism.

[0102] The hippocampal indexing mechanism specifically works as follows: a memory case is created for each successful query case, and a multi-dimensional index is created for the memory case. The new memory case is first stored in a temporary memory cache. When the confidence assessment result of its association is higher than a preset standard and after optimization, it is transferred from the temporary memory cache to the long-term memory storage area. When a new query request is received, the multi-dimensional index is used to perform association retrieval, one or more related memory fragments are retrieved from the long-term memory storage area, and recombined through the association index to reconstruct a complete query solution.

[0103] Semantic indexes are built based on the core semantic content of the query case; structural indexes are built based on the document hierarchy and relationships involved in the query case; time indexes are built based on the timestamps of the query case; and association indexes are built based on the logical relationships between the case and other memory cases.

[0104] The specific mechanisms of neural plasticity are as follows: for memory paths that are frequently used and have high confidence, their connection strength in the memory network is enhanced, i.e., long-term reinforcement is performed; for memory paths that are used infrequently or have consistently low confidence, their connection weight is reduced, i.e., long-term inhibition is performed.

[0105] The implementation steps for memory retrieval and matching are as follows:

[0106] For the current query request, based on its semantic content, the structure of the document involved and the context information, the semantic similarity, structural similarity and contextual similarity with the cases in the memory case library are calculated respectively, and the comprehensive similarity score is calculated accordingly.

[0107] Based on the comprehensive similarity matching, associated retrieval is performed through the constructed associative memory network. The associative memory network consists of memory cases as nodes and logical associations between cases as edges, supporting the retrieval of semantically or structurally indirectly related memory cases along the associated edges starting from the current query.

[0108] For complex query paths identified through frequent use, their general query logic is extracted through abstraction and generalization to form a reusable general query template, which is then stored in the brain-like memory mechanism.

[0109] The update rule for memory connection weights is as follows: when a memory path is frequently invoked and the corresponding output confidence level is consistently higher than the high confidence threshold, the corresponding connection weight is increased by a predefined ratio (similar to long-term enhancement); when a memory path is invoked infrequently and the confidence level is consistently lower than the low confidence threshold, the corresponding connection weight is decreased by a predefined ratio (similar to long-term suppression); the threshold and enhancement / decrease ratio are configurable or learnable parameters.

[0110] It deeply simulates the working mechanism of the human brain's hippocampus-neocortical memory system, possessing plasticity and adaptability similar to biological neural networks, enabling dynamic management, optimization, and evolution of memory, and supporting multi-level memory retrieval and associative reasoning.

[0111] Example 2: This example provides a power transmission and transformation scenario document processing system based on a brain-like memory mechanism, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the power transmission and transformation scenario document processing method based on any one of the above-described brain-like memory mechanisms.

[0112] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A document processing method for power transmission and transformation scenarios based on a brain-like memory mechanism, characterized in that, Includes the following steps: S1. Receive the user's query request for the target document, generate the query intent based on the query request using a pre-trained large language model in the power transmission and transformation field, and generate a query plan based on the query intent. S2. Based on the query plan, relevant documents are retrieved from the hierarchically modeled power transmission and transformation document knowledge base, and information extraction and reasoning are performed on the retrieved relevant documents using a dual-path parallel processing mechanism and a hierarchical abstract processing mechanism. S3. The information obtained from extraction and reasoning is structured and filled according to the predefined power transmission and transformation industry document template to generate preliminary structured output. At the same time, a brain-like evidence accumulation decision-making mechanism is adopted to transform the information obtained from extraction and reasoning into an interpretable reasoning chain record. S4. A multi-source evidence accumulation mechanism is used to assess the confidence of the initial structured output, thereby optimizing the inference chain record. Memory cases are generated based on the inference chain record and the confidence assessment results. A brain-like memory mechanism is used to store and optimize the memory cases.

2. The document processing method for power transmission and transformation scenarios based on brain-like memory mechanism according to claim 1, characterized in that: In step S1, after receiving the user's query request for the target document, the system first retrieves historical query cases from the brain-like memory mechanism and calculates the confidence level. If the confidence level of the retrieved similar cases is not lower than the preset confidence threshold, the system calls the memory result corresponding to the similar cases as the output and terminates the subsequent steps. Otherwise, step S2 is executed.

3. The document processing method for power transmission and transformation scenarios based on brain-like memory mechanism according to claim 2, characterized in that: In S1, after the query plan is executed, a reflection and checking mechanism is introduced for optimization. Specifically, through logical consistency verification and cross-validation, inconsistencies or obvious logical errors are identified and corrected.

4. The document processing method for power transmission and transformation scenarios based on brain-like memory mechanism according to claim 3, characterized in that: In step S2, the specific steps for retrieving target documents based on the query plan in a hierarchically modeled power transmission and transformation document knowledge base are as follows: S21. Construct a multi-level knowledge base structure for power transmission and transformation documents based on organizational structure, functional classification, and time range. S22. First, perform coarse-grained retrieval at a high level to determine the relevant document categories, and then perform fine-grained document retrieval within the determined categories. S23. Identify the reference relationships, dependencies, and hierarchical relationships between documents, and construct a document relationship graph; S24. Based on the document association graph, perform association expansion retrieval on the documents retrieved through fine-grained retrieval, and add related documents that have a dependency relationship with the recalled documents to the recall scope.

5. The document processing method for power transmission and transformation scenarios based on brain-like memory mechanism according to claim 4, characterized in that: In S2, the dual-path parallel processing mechanism includes a ventral path processing sub-step, a dorsal path processing sub-step, and a path fusion sub-step; the hierarchical abstract processing mechanism includes primary feature extraction, intermediate pattern recognition, and advanced semantic understanding.

6. The document processing method for power transmission and transformation scenarios based on brain-like memory mechanism according to claim 5, characterized in that: In step S3, the specific steps for filling in the extracted and inferred information according to a predefined power transmission and transformation industry document template to generate preliminary structured output are as follows: Select a structured output template based on the query type and target document format. Organize and populate the extracted and inferred information on the template according to the predefined structured format. After filling, verify the completeness of the output results.

7. The document processing method for power transmission and transformation scenarios based on brain-like memory mechanism according to claim 6, characterized in that: In S3, the brain-like evidence accumulation decision-making mechanism is used to transform the extracted and reasoned information into an interpretable reasoning chain record. The specific steps are as follows: S31. Extract and collect evidence to verify the preliminary structured output, and label the information source for each piece of evidence; S32. Based on the reliability, timeliness, and relevance of the evidence source to the current query, assign dynamic weights to each piece of evidence collected; at the same time, identify and label evidence conflicts from different information sources. S33. Starting with the initial structured output, construct a logical skeleton composed of multiple reasoning nodes to generate a reasoning chain, and attach weighted evidence and evidence conflicts to the corresponding nodes of the reasoning chain. S34. Identify the risk level and problem nodes of the inference chain decision, quantify the uncertainty of the problem nodes in the inference chain, and mark and record the quantified problem nodes in the inference chain. S35. Integrate all the information from steps S31 to S34 to generate a structured, interpretable reasoning chain record.

8. The document processing method for power transmission and transformation scenarios based on brain-like memory mechanism according to claim 7, characterized in that: In step S4, a multi-source evidence accumulation mechanism is used to assess the confidence level of the preliminary structured output, specifically as follows: Based on the evidence weights attached to the nodes, the cumulative evidence strength is calculated. Depending on whether the cumulative evidence strength reaches the confidence threshold preset for decision-making, it is determined whether to implement an early stop strategy or a progressive deep analysis strategy. The adopted decision strategy, the cumulative evidence strength, and the reasoning for the decision are recorded in the inference chain.

9. The document processing method for power transmission and transformation scenarios based on brain-like memory mechanism according to claim 8, characterized in that: In step S4, the brain-like memory mechanism simulates the hippocampal indexing mechanism, neuroplasticity mechanism, and memory retrieval and matching to store and optimize memory cases. The specific steps are as follows: S41. Simulate the hippocampal indexing mechanism to generate a memory case for each successful historical query case, store the memory cases in a brain-like memory mechanism, and establish a multi-dimensional index for the memory cases; S42. For new query requests, use multi-dimensional indexes and multi-modal similarity matching to perform memory retrieval, obtain highly similar memory cases, and then reuse the results of those cases. S43. Simulate the neural plasticity mechanism and dynamically adjust the connection weights of memory cases based on user feedback and query confidence. S44. Extract the query logic for complex query paths, generate a general query template, and update it to the brain-like memory mechanism.

10. A document processing system for power transmission and transformation scenarios based on a neuromorphic memory mechanism, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the document processing method for power transmission and transformation scenarios based on a brain-like memory mechanism as described in any one of claims 1-9.

Images