Hierarchical recursive text chunking method for power equipment operation and maintenance knowledge management
Patent Information
- Application Number
- CN202611403678.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-10
- Publication Date
- 2026-11-03
AI Technical Summary
[0008]本发明的目的在于现有文本分块技术粒度难以控制、容易导致语义不完整和上下文割裂的缺陷，特别是电力设备运维文档天然具有多层级语义结构、而现有方法无法输出多粒度匹配结果的问题，提出了面向电力设备运维知识管理的层次化递归文本分块方法
本发明提出的面向电力设备运维知识管理的层次化递归文本分块方法，针对电力设备运维文档多层级、强逻辑关联的特点，通过层次化递归合并的分块架构，实现分块粒度灵活可控、语义完整内聚与处理效率的兼顾，适配电力运维知识管理的多场景业务需求。通过至少一个层级的递归合并处理，可输出不同层级的语义单元序列，实现分块粒度的灵活调控。细粒度分块结果可支撑单条故障处理措施、单项试验参数的精准定位检索；粗粒度分块结果可聚合同设备多份运维记录、同型号设备家族性缺陷资料，满足设备全生命周期分析、家族缺陷研判等高层级知识管理需求，解决现有技术分块粒度单一、无法适配多业务场景的问题。基于语义相似度识别语义波谷并执行单元合并，以语义关联的自然边界作为分块依据，可避免固定尺寸分块造成的语义割裂。针对变压器故障报告、检修作业指导书等逻辑链条完整的运维文档，能够完整保留故障现象、原因分析、处理措施等关联内容，确保各层级文本块内部语义高度内聚，有效提升后续知识检索与调用的准确性。基于当前层级语义单元的嵌入向量计算下一层级嵌入向量，无需对合并后的长文本重复执行编码计算，在保障向量语义代表性的前提下，显著降低海量运维文档批量预处理的计算开销，提升电力设备大规模运维文档库的处理效率。综上，本方法为电力设备运维知识的检索增强生成、知识图谱构建等下游应用提供高质量数据基础，可有效支撑运维知识资产的智能化管理与高效复用。
Smart Images
Figure CN123030753A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a hierarchical recursive text segmentation method for knowledge management of power equipment operation and maintenance. Background Technology
[0002] During the entire lifecycle management of power equipment, enterprises accumulate a massive amount of technical documents, covering equipment technical manuals and operating procedures, maintenance operation instructions, historical fault reports and maintenance records, equipment test and inspection reports, as well as text data obtained from scanned copies of professional books and technical standards through optical character recognition (OCR). To achieve intelligent management and efficient utilization of document assets, retrieval augmented generation (RAG) technology based on large language models (LLM) has become the mainstream technical approach. In the RAG system, text segmentation is a key step in preprocessing long documents into data units that can be processed by the model, and the quality of the segmentation results directly affects the retrieval accuracy and generation effect.
[0003] Existing text segmentation techniques can be mainly divided into two categories. The first category is fixed-size segmentation, which cuts the text according to a fixed number of characters or tokens. This type of method is simple to implement and fast to process, but it will cut off complete semantic units and destroy the internal logic and contextual connections of the text. In the context of power equipment documents, the impact of this limitation is particularly prominent: for example, a transformer fault report usually contains multiple related components such as fault phenomena, test data, cause analysis, handling measures, and prevention suggestions. Fixed-size segmentation will arbitrarily interrupt the above logical links, resulting in only fragmented fragments being obtained during retrieval, making it impossible to reconstruct the complete chain of fault information.
[0004] The second category is semantic segmentation techniques based on similarity thresholds. This type of technique calculates the similarity of the embedding vectors of adjacent sentences or text segments and performs segmentation when the similarity is below a preset fixed threshold. This method improves upon fixed-size segmentation and can preserve the integrity of semantic units to some extent. However, in the context of power equipment documents, this type of technique still has the following limitations: (1) Limited granularity and controllability of segmentation. This type of method typically outputs independent segments of natural paragraphs or even shorter granularities. However, the power equipment operation and maintenance scenario has multi-granularity business requirements: when maintenance personnel locate the handling measures for a single fault, they need to adapt to fine-grained segments at the paragraph level; when technicians analyze the long-term fault evolution patterns of equipment, they need to aggregate coarse-grained segments from multiple documents. Existing semantic segmentation technology can only output results of a single granularity in a single processing, which cannot adapt to the differentiated retrieval needs of multiple scenarios.
[0005] (2) Inability to adapt to the multi-level semantic structure of documents. Power equipment technical documents naturally possess multi-level semantic relationships: a single maintenance record and other records of the same maintenance task constitute a task-level relationship; multiple maintenance records of the same equipment constitute an equipment-level knowledge cluster; multiple equipment records within the same substation constitute a station-level knowledge cluster; and equipment records of the same voltage level and model constitute a system-level or family-level knowledge cluster. Existing semantic segmentation technology only performs segmentation at a single level, which will disperse content belonging to the same upper-level topic into multiple independent small blocks, severing the semantic relationships across levels, and failing to support high-level retrieval needs such as equipment life cycle analysis and defect analysis of the same model of equipment family.
[0006] (3) Insufficient adaptability to different scenarios. The power enterprise document library covers a variety of equipment types such as transformers, circuit breakers, and instrument transformers, as well as multiple business areas such as operation, maintenance, and testing. The internal hierarchical structure of different categories of documents varies significantly. The single-level block method with a fixed threshold has poor adaptability to documents in different fields and with different structures, making it difficult to achieve flexible control of block granularity while ensuring semantic integrity.
[0007] Therefore, how to provide a text segmentation method that can flexibly control the granularity of segmentation, aggregate multi-level semantic content, and ensure the semantic integrity of each level of segmentation is a technical problem that urgently needs to be solved in the field of power equipment operation and maintenance knowledge management. Summary of the Invention
[0008] The purpose of this invention is to address the shortcomings of existing text segmentation techniques, such as difficulty in controlling the granularity, which can easily lead to semantic incompleteness and contextual fragmentation. In particular, existing methods cannot output multi-granularity matching results when power equipment operation and maintenance documents naturally have a multi-level semantic structure. Therefore, this invention proposes a hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management.
[0009] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a hierarchical recursive text segmentation method for knowledge management of power equipment operation and maintenance, comprising the following steps: Determine the initial semantic unit sequence of the power equipment operation and maintenance text to be processed, and determine the initial embedding vector of each initial semantic unit in the initial semantic unit sequence; Based on the initial semantic unit sequence and the initial embedding vector, perform at least one level of recursive merging processing, and output the final level of semantic unit sequence as the power equipment operation and maintenance text block result. The recursive merging process at each level includes: Based on the embedding vectors of each semantic unit in the current hierarchical semantic unit sequence, calculate the semantic similarity between adjacent semantic units and generate a semantic similarity sequence; Semantic valleys are identified based on semantic similarity sequences, and adjacent semantic units at the current level are merged based on these semantic valleys to obtain the semantic unit sequence at the next level. For each semantic unit in the next-level semantic unit sequence, the next-level embedding vector of the semantic unit is calculated based on the embedding vectors of all current-level semantic units contained in the semantic unit.
[0010] Furthermore, the next-level embedding vector of the semantic unit is calculated, including: Perform average pooling on the embedding vectors of all current-level semantic units contained in the semantic unit.
[0011] Furthermore, identifying semantic valleys based on semantic similarity sequences includes: extracting all local minima in the semantic similarity sequence; for each local minima, when the difference between the local minima and the adjacent local maxima is greater than a preset valley depth threshold, the location of the local minima is determined as a semantic valley.
[0012] Furthermore, adjacent semantic units at the current level are merged based on semantic valleys, including: following preset minimum unit size constraints and preset maximum unit size constraints.
[0013] Furthermore, the initial semantic unit is a sentence unit obtained by segmenting from the power equipment operation and maintenance text to be processed through preset sentence segmentation rules.
[0014] Furthermore, the semantic similarity between adjacent semantic units is calculated using cosine similarity.
[0015] Furthermore, each level of recursive merging processing is configured with independent operating parameters.
[0016] Secondly, this invention provides a hierarchical recursive text segmentation system for knowledge management of power equipment operation and maintenance, comprising: The determination module is used to determine the initial semantic unit sequence of the power equipment operation and maintenance text to be processed, and to determine the initial embedding vector of each initial semantic unit in the initial semantic unit sequence. The processing module is used to perform at least one level of recursive merging processing based on the initial semantic unit sequence and the initial embedding vector, and output the final level of semantic unit sequence as the power equipment operation and maintenance text block result. The recursive merging process at each level includes: Based on the embedding vectors of each semantic unit in the current hierarchical semantic unit sequence, calculate the semantic similarity between adjacent semantic units and generate a semantic similarity sequence; Semantic valleys are identified based on semantic similarity sequences, and adjacent semantic units at the current level are merged based on these semantic valleys to obtain the semantic unit sequence at the next level. For each semantic unit in the next-level semantic unit sequence, the next-level embedding vector of the semantic unit is calculated based on the embedding vectors of all current-level semantic units contained in the semantic unit.
[0017] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a hierarchical recursive text segmentation method for knowledge management of power equipment operation and maintenance.
[0018] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a hierarchical recursive text segmentation method for knowledge management of power equipment operation and maintenance.
[0019] Compared with the prior art, the present invention has the following beneficial technical effects: This invention proposes a hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management. Addressing the multi-level and strongly logically related characteristics of power equipment operation and maintenance documents, this method employs a hierarchical recursive merging segmentation architecture to achieve a balance between flexible and controllable segmentation granularity, semantic integrity and cohesion, and processing efficiency, adapting to the diverse business needs of power operation and maintenance knowledge management. Through recursive merging processing at at least one level, it can output semantic unit sequences at different levels, enabling flexible adjustment of segmentation granularity. Fine-grained segmentation results can support precise location and retrieval of single fault handling measures and single test parameters; coarse-grained segmentation results can aggregate multiple operation and maintenance records of contracted equipment and family defect data of the same model of equipment, meeting high-level knowledge management needs such as equipment lifecycle analysis and family defect assessment, solving the problem of existing technologies having a single segmentation granularity and being unable to adapt to multiple business scenarios. Based on semantic similarity, it identifies semantic valleys and performs unit merging, using the natural boundaries of semantic association as the segmentation basis, avoiding semantic fragmentation caused by fixed-size segmentation. For maintenance documents with complete logical chains, such as transformer fault reports and maintenance work instructions, this method can fully preserve related content such as fault phenomena, cause analysis, and handling measures, ensuring high semantic cohesion within text blocks at each level and effectively improving the accuracy of subsequent knowledge retrieval and retrieval. Based on the embedding vector of the current level semantic unit, the method calculates the embedding vector of the next level, eliminating the need to repeatedly perform encoding calculations on the merged long text. While ensuring the semantic representativeness of the vectors, this significantly reduces the computational overhead of batch preprocessing massive maintenance documents, improving the processing efficiency of large-scale power equipment maintenance document libraries. In summary, this method provides a high-quality data foundation for downstream applications such as enhanced generation of power equipment maintenance knowledge retrieval and knowledge graph construction, effectively supporting the intelligent management and efficient reuse of maintenance knowledge assets. Attached Figure Description
[0020] The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of the invention in any way. Furthermore, the shapes and proportions of the components in the drawings are merely schematic to aid in understanding the invention and do not specifically limit the shapes and proportions of the components. In the drawings: Figure 1 This is a flowchart of the hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management according to the present invention.
[0021] Figure 2 This is a simplified structural diagram of the hierarchical recursive text segmentation system for power equipment operation and maintenance knowledge management according to the present invention.
[0022] Figure 3 This is an electronic device diagram illustrating the hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management according to the present invention.
[0023] Figure 4 This is a schematic diagram of the functional modules of the text segmentation device provided in an embodiment of the present invention.
[0024] Figure 5 A flowchart of a text segmentation method provided in an embodiment of the present invention.
[0025] Figure 6 This is a schematic diagram of the hardware structure of the computer system in which the device of this embodiment of the invention is located. Detailed Implementation
[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.
[0027] Example 1 See Figure 1 A hierarchical recursive text segmentation method for knowledge management of power equipment operation and maintenance includes the following steps: The process involves determining the initial semantic unit sequence of the power equipment operation and maintenance text to be processed, and determining the initial embedding vector of each initial semantic unit in the initial semantic unit sequence. Based on the initial semantic unit sequence and the initial embedding vector, at least one level of recursive merging processing is performed, and the final level of semantic unit sequence is output as the power equipment operation and maintenance text segmentation result. The recursive merging processing at each level includes: calculating the semantic similarity between adjacent semantic units based on the embedding vector of each semantic unit in the current level semantic unit sequence, and generating a semantic similarity sequence; identifying semantic valleys based on the semantic similarity sequence, and merging adjacent current level semantic units based on the semantic valleys to obtain the next level semantic unit sequence; and for each semantic unit in the next level semantic unit sequence, calculating the next level embedding vector of the semantic unit based on the embedding vectors of all current level semantic units contained in the semantic unit.
[0028] The next-level embedding vector of the semantic unit is calculated by performing average pooling on the embedding vectors of all current-level semantic units contained in the semantic unit.
[0029] Identifying semantic valleys based on semantic similarity sequences includes: extracting all local minima in the semantic similarity sequence; for each local minima, when the difference between the local minima and the adjacent local maxima is greater than a preset valley depth threshold, the location of the local minima is determined as a semantic valley.
[0030] Merge adjacent semantic units at the current level based on semantic valleys, including: following preset minimum unit size constraints and preset maximum unit size constraints.
[0031] The initial semantic units are sentence units obtained by segmenting from the power equipment operation and maintenance text to be processed using preset sentence segmentation rules.
[0032] The semantic similarity between adjacent semantic units is calculated using cosine similarity.
[0033] Each level of recursive merging is configured with independent running parameters.
[0034] Reference Figure 5 The method flow of a specific embodiment of the present invention will be described in detail below.
[0035] Steps S1 and S2: Determine the initial semantic units and embedding vectors. The determination module receives the text to be processed and segments it according to preset sentence segmentation rules using its internal segmentation unit to obtain a sequence of initial semantic units at the sentence level. Those skilled in the art will understand that the initial semantic units are not limited to sentences; they can also be natural paragraphs segmented based on line breaks, etc., and this invention does not limit this. Subsequently, the embedding unit calls the embedding model to convert the text units into initial embedding vectors. The embedding model can be any model capable of mapping text to a high-dimensional vector space, such as, but not limited to, BERT, RoBERTA, and GPT series models based on the Transformer architecture, or models called through API interfaces provided by cloud service providers.
[0036] Steps S3-S6: Perform recursive merging. This is the core step performed by the processing module.
[0037] Step S4: The similarity calculation unit calculates the semantic similarity between adjacent units. In this preferred embodiment, cosine similarity is used as the metric. In other embodiments, the reciprocal of Euclidean distance, dot product, or other metrics that can characterize vector similarity may also be used.
[0038] Step S5: Valley Identification and Merging. Semantic valleys are identified and merged. A semantic valley is a local minimum point in a similarity sequence. Preferably, to increase robustness, a valid valley point not only needs to be a local minimum, but its difference from its adjacent local maximum (peak) must also be greater than a configurable valley depth threshold, valleyDelta. After identifying all valid valleys, continuous units between valleys are merged, using the valleys as boundaries. Furthermore, this merging operation can be constrained by the minimum / maximum size constraints of the level configuration to ensure that the generated block length is within a reasonable range.
[0039] Step S6: The high-level vector generation unit calculates the embedding vector for the newly merged unit. As claimed in claim 2, this embodiment preferably employs a low-computational-cost and robust average pooling method. In other possible embodiments, max pooling or weighted average pooling with attention mechanisms can also be used to generate high-level vectors.
[0040] Recursive Loop: After completing the merging process of one level, the system determines whether the preset final level has been reached. If not, the currently generated next-level unit sequence and its embedding vector are used as the new "current level" input, and the system returns to step S3 to begin the merging process of the next level. It is worth noting that each recursive level can use an independent set of different merging parameters (such as valleyDelta, size constraints, etc.) to adapt to semantic aggregation characteristics at different granularities.
[0041] Step S7: Output Results. Once the final level is reached, the processing module will output the results.
[0042] The present invention has the following advantages: Flexible and controllable granularity of block segmentation: Through a multi-level recursive merging framework, users can set the number of recursive levels and the merging parameters for each level according to actual business needs, thereby generating power equipment operation and maintenance text blocks of different granularities from paragraph level, chapter level to topic level. For example, in a transformer operation and maintenance knowledge management system, the first level can output fine-grained power equipment operation and maintenance text blocks (such as independent accident cause paragraphs) for "quick retrieval of single faults," the second level can output medium-grained power equipment operation and maintenance text blocks (containing a complete description of the same fault case) for "equipment health assessment," and the third level can output coarse-grained power equipment operation and maintenance text blocks (aggregating all historical records of equipment of the same model) for "family defect analysis." This feature greatly improves the flexibility and adaptability of block segmentation in multiple business scenarios of power operation and maintenance.
[0043] Strong semantic integrity and coherence: This invention merges based on semantic valleys, ensuring that the most semantically relevant continuous units are merged at each level. This results in highly aggregated and coherent internal semantics in the generated power equipment operation and maintenance text blocks. Taking transformer fault reports as an example, this method can correctly merge description paragraphs and cause analysis paragraphs of the same fault at lower levels; at higher levels, it can aggregate multiple fault reports of the same equipment within a similar time period into a single topic cluster, effectively avoiding the fragmentation of contextual information. This is particularly suitable for operation and maintenance retrieval scenarios requiring a complete contextual chain, such as "fault tracing."
[0044] High computational efficiency: For the embedding vectors of high-level units, this invention adopts an average pooling strategy for the low-level vectors, avoiding repetitive and time-consuming encoding calculations on the merged long text, thus significantly improving processing efficiency while ensuring semantic representativeness. This advantage is particularly crucial in batch processing scenarios of large-scale power equipment document libraries (such as document libraries containing tens of thousands of maintenance reports and test records from decades ago).
[0045] Structured and Hierarchical: This invention reflects the inherent hierarchical structure of documents, particularly aligning with the natural hierarchical organization of power equipment technical documents, from "single record → equipment file → station-level summary → systematic analysis." The output power equipment operation and maintenance text blocks are not only content blocks but also include hierarchical information, serving as structured input for downstream tasks such as equipment knowledge graph construction and cross-document topic association analysis, providing fundamental support for the intelligent upgrade of the power equipment operation and maintenance knowledge management system.
[0046] Example 2 See Figure 2 A hierarchical recursive text segmentation system for knowledge management of power equipment operation and maintenance includes: The determination module is used to determine the initial semantic unit sequence of the power equipment operation and maintenance text to be processed, and to determine the initial embedding vector of each initial semantic unit in the initial semantic unit sequence. The processing module is used to perform at least one level of recursive merging processing based on the initial semantic unit sequence and the initial embedding vector, and output the final level of semantic unit sequence as the power equipment operation and maintenance text block result. The processing module is configured as follows when performing recursive merging processing at each level: Based on the embedding vectors of each semantic unit in the current hierarchical semantic unit sequence, calculate the semantic similarity between adjacent semantic units and generate a semantic similarity sequence; Semantic valleys are identified based on semantic similarity sequences, and adjacent semantic units at the current level are merged based on these semantic valleys to obtain the semantic unit sequence at the next level. For each semantic unit in the next-level semantic unit sequence, the next-level embedding vector of the semantic unit is calculated based on the embedding vectors of all current-level semantic units contained in the semantic unit.
[0047] Furthermore, the processing module is further configured to perform average pooling on the embedding vectors of all current-level semantic units contained in the next-level semantic unit to obtain the next-level embedding vector.
[0048] Example 3 See Figure 3 An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a hierarchical recursive text segmentation method for knowledge management of power equipment operation and maintenance.
[0049] Reference Figure 4 and Figure 6This invention provides a text segmentation device. The text segmentation device receives text to be processed and outputs text blocks and metadata. The device internally includes a determination module and a processing module. The determination module internally includes a segmentation unit and an embedding unit. The segmentation unit receives the text to be processed and outputs an initial semantic unit sequence. The embedding unit is cascaded with the segmentation unit to generate initial embedding vectors for each initial semantic unit in the initial semantic unit sequence. The processing module is connected to the determination module and receives the initial semantic unit sequence and the initial embedding vectors, performing at least one level of recursive merging processing. The processing module internally includes multi-level processing units to sequentially complete semantic similarity calculation, semantic valley identification and unit merging within a single level, and embedding vector calculation for the next level. The processing module includes a recursive processing loop to feed back the semantic unit sequence and embedding vectors output from the current level to the input side of the processing module as input for the next level of merging processing. The processing module finally outputs the semantic unit sequence and corresponding metadata for the final level. At the hardware level, this device can be a general-purpose computer system, such as... Figure 6 As shown, it includes a processor, a memory, and input and output interfaces for data interaction. The memory stores computer program instructions that implement the invention; when the processor executes these instructions, it will implement… Figure 4 The functional modules shown are as follows. At the software level, these modules can be independent software units or different functional parts integrated into an application.
[0050] Example 4 A computer-readable storage medium storing a computer program that, when executed by a processor, implements a hierarchical recursive text segmentation method for knowledge management of power equipment operation and maintenance.
[0051] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, read-only optical discs, optical storage, etc.) containing computer-usable program code.
[0052] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0053] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0054] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.
Claims
1. A hierarchical recursive text segmentation method for knowledge management of power equipment operation and maintenance, characterized in that, Includes the following steps: Determine the initial semantic unit sequence of the power equipment operation and maintenance text to be processed, and determine the initial embedding vector of each initial semantic unit in the initial semantic unit sequence; Based on the initial semantic unit sequence and the initial embedding vector, perform at least one level of recursive merging processing to output the final level of semantic unit sequence as the power equipment operation and maintenance text segmentation result. The recursive merging process at each level includes: Based on the embedding vectors of each semantic unit in the current hierarchical semantic unit sequence, calculate the semantic similarity between adjacent semantic units and generate a semantic similarity sequence; Based on the semantic similarity sequence, semantic valleys are identified, and adjacent semantic units at the current level are merged according to the semantic valleys to obtain the semantic unit sequence at the next level. For each semantic unit in the next-level semantic unit sequence, the next-level embedding vector of the semantic unit is calculated based on the embedding vectors of all current-level semantic units contained in the semantic unit.
2. The hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management according to claim 1, characterized in that, The calculation to obtain the next-level embedding vector of the semantic unit includes: Average pooling is performed on the embedding vectors of all current-level semantic units contained in the semantic unit.
3. The hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management according to claim 1, characterized in that, The step of identifying semantic valleys based on the semantic similarity sequence includes: extracting all local minima in the semantic similarity sequence; and for each local minima, determining the location of the local minima as a semantic valley when the difference between the local minima and its adjacent local maxima is greater than a preset valley depth threshold.
4. The hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management according to claim 1, characterized in that, The merging of adjacent current-level semantic units based on the semantic valley includes: following preset minimum unit size constraints and preset maximum unit size constraints.
5. The hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management according to claim 1, characterized in that, The initial semantic unit is a sentence unit obtained by segmenting the power equipment operation and maintenance text to be processed using preset sentence segmentation rules.
6. The hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management according to claim 1, characterized in that, The semantic similarity between adjacent semantic units is calculated using cosine similarity.
7. The hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management according to claim 1, characterized in that, Each level of recursive merging is configured with independent running parameters.
8. A hierarchical recursive text segmentation system for knowledge management of power equipment operation and maintenance, characterized in that: include: The determination module is used to determine the initial semantic unit sequence of the power equipment operation and maintenance text to be processed, and to determine the initial embedding vector of each initial semantic unit in the initial semantic unit sequence. The processing module is used to perform at least one level of recursive merging processing based on the initial semantic unit sequence and the initial embedding vector, and output the final level of semantic unit sequence as the power equipment operation and maintenance text block result. The recursive merging process at each level includes: Based on the embedding vectors of each semantic unit in the current hierarchical semantic unit sequence, calculate the semantic similarity between adjacent semantic units and generate a semantic similarity sequence; Based on the semantic similarity sequence, semantic valleys are identified, and adjacent semantic units at the current level are merged according to the semantic valleys to obtain the semantic unit sequence at the next level. For each semantic unit in the next-level semantic unit sequence, the next-level embedding vector of the semantic unit is calculated based on the embedding vectors of all current-level semantic units contained in the semantic unit.
9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the hierarchical recursive text segmentation method for power equipment operation and maintenance knowledge management as described in any one of claims 1-7.