Power document generation method, system, device and medium based on sentence segmentation retrieval

Abstract

The application discloses a power document generation method and system based on sentence division retrieval, a device and a medium, and is applied to the field of document generation, and comprises the following steps: encoding power equipment images through a visual encoder based on a VLLM model structure to obtain high-dimensional dense visual features, performing dimension reduction processing on the high-dimensional dense visual features by using an abstract module to extract low-dimensional visual embedding, performing theme recognition on the low-dimensional visual embedding by a theme division encoder to generate common theme query vectors and rare theme query vectors, retrieving sentences in a power report training set, extracting a sentence set matched with the query vectors, and selecting relevant sentences to generate a power document according to the sentence set and power document generation requirements through a preset sentence screening rule. The application realizes efficient monitoring and fault diagnosis of power equipment states, automatically generates power documents with strong pertinence and clear logic, and significantly improves the intelligent level and work efficiency of power equipment operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of document generation technology, and in particular to a method, system, device and medium for generating power documents based on sentence segmentation retrieval. Background Technology

[0002] With the digital transformation of the power industry, the operation and maintenance data of power equipment is growing explosively, generating a large amount of structured and unstructured document resources. These documents contain rich information on equipment status, faults, and operation and maintenance records, which are of great value for the intelligent management of power systems.

[0003] Existing methods for generating power system documentation suffer from numerous inconveniences when handling such massive amounts of data. Traditional methods typically rely on manual writing or simple template filling, which is not only time-consuming and labor-intensive but also struggles to guarantee document accuracy and consistency. Furthermore, existing technologies lack effective integration mechanisms when processing images and text information from power equipment, resulting in documents that often fail to fully reflect the actual condition of the equipment. These limitations become even more pronounced when facing complex and ever-changing power equipment fault and maintenance scenarios, hindering the rapid generation of high-quality document reports.

[0004] Therefore, how to efficiently and accurately generate power documents and realize intelligent processing of images and text information of power equipment has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides a method, system, device, and medium for generating power documents based on sentence segmentation retrieval, so as to achieve accurate generation of power documents.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a method for generating power documents based on sentence segmentation retrieval, comprising:

[0007] The power equipment image is encoded by a visual encoder based on the VLLM model structure to obtain the high-dimensional dense visual features of the power equipment image.

[0008] The high-dimensional dense visual features are subjected to dimensionality reduction processing to extract low-dimensional visual embedding information.

[0009] Field recognition is performed on the low-dimensional visual embedded information to generate a first field query vector and a second field query vector.

[0010] In the power report training set, the first field query vector and the second field query vector are used as benchmarks for retrieval to extract a first set of statements corresponding to the first field query vector and a second set of statements corresponding to the second field query vector.

[0011] Based on the selected power document generation requirements, power documents corresponding to the power equipment images are generated according to the first set of statements and the second set of statements.

[0012] Furthermore, the process of encoding the power equipment image using a visual encoder based on a VLLM model structure to obtain the high-dimensional dense visual features of the power equipment image includes:

[0013] Images of power equipment are input into a visual encoder based on a VLLM model structure.

[0014] The feature extraction layer of the visual encoder extracts features from the image of the power equipment to obtain an initial feature representation.

[0015] The initial feature representation is encoded by the encoding layer of the visual encoder to obtain high-dimensional dense visual features.

[0016] Further, the step of performing dimensionality reduction processing on the high-dimensional dense visual features to extract low-dimensional visual embedding information includes:

[0017] The high-dimensional dense visual features are input into the abstraction module.

[0018] The high-dimensional dense visual features are processed by the multi-layer Transformer decoder of the abstract module to generate intermediate feature representations.

[0019] The intermediate feature representation is projected onto a low-dimensional space through the projection layer of the abstract module, thereby extracting low-dimensional visual embedding information.

[0020] Further, the step of performing field recognition on the low-dimensional visual embedding information to generate a first field query vector and a second field query vector includes:

[0021] The low-dimensional visual embedding information is used to extract features through the multi-head self-attention mechanism of the topic segmentation encoder to generate topic feature representations.

[0022] The topic feature representation is classified by the classification layer of the topic partitioning encoder to generate a first field query vector and a second field query vector.

[0023] Further, in the power report training set, the retrieval is performed using the first field query vector and the second field query vector as benchmarks, respectively, to extract a first set of statements corresponding to the first field query vector and a second set of statements corresponding to the second field query vector, including:

[0024] The similarity between the query vector of the first field and the query vector of the second field and the sentence vector of the statement in the power report training set is calculated.

[0025] Based on the similarity calculation results, statements with a matching degree higher than a preset threshold with the first field query vector and the second field query vector are selected and combined to form a first statement set corresponding to the first field query vector and a second statement set corresponding to the second field query vector.

[0026] Further, the step of generating an electrical document corresponding to the electrical equipment image based on the selected electrical document generation requirements and according to the first statement set and the second statement set includes:

[0027] Determine the set of key information keywords based on the requirements for generating power documents.

[0028] Select statements from the first set of statements and the second set of statements that contain any keyword from the set of key information keywords.

[0029] Based on the preset statement combination template, the selected statements are combined to generate power documents.

[0030] Another embodiment of the present invention provides a power document generation system based on sentence segmentation retrieval, comprising:

[0031] The image encoding module is used to encode the power equipment image using a visual encoder based on the VLLM model structure to obtain the high-dimensional dense visual features of the power equipment image.

[0032] The feature dimensionality reduction module is used to perform dimensionality reduction processing on the high-dimensional dense visual features and extract low-dimensional visual embedding information.

[0033] The vector generation module is used to perform field recognition on the low-dimensional visual embedded information and generate a first field query vector and a second field query vector.

[0034] The statement retrieval module is used to retrieve statements from the power report training set by using the first field query vector and the second field query vector as benchmarks, respectively, and to extract a first set of statements corresponding to the first field query vector and a second set of statements corresponding to the second field query vector.

[0035] The document generation module is used to generate power documents corresponding to the power equipment images based on the selected power document generation requirements, according to the first statement set and the second statement set.

[0036] Furthermore, the vector generation module is used for:

[0037] The low-dimensional visual embedding information is used to extract features through the multi-head self-attention mechanism of the topic segmentation encoder to generate topic feature representations.

[0038] The topic feature representation is classified by the classification layer of the topic partitioning encoder to generate a first field query vector and a second field query vector.

[0039] Another embodiment of the present invention provides a computer device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the power document generation method based on sentence segmentation retrieval as described above.

[0040] Another embodiment of the present invention provides a computer-readable storage medium storing a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the power document generation method based on sentence segmentation retrieval as described above.

[0041] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following:

[0042] This system efficiently processes power equipment images through a visual encoder and abstraction module, extracting low-dimensional visual embeddings to significantly reduce redundant information and enhance feature representation capabilities. Combined with a topic segmentation encoder, it accurately generates query vectors for common and rare topics, enabling differentiated retrieval and improving retrieval accuracy and efficiency. For rare topics in the power field where samples are scarce but crucial, the topic segmentation encoder accurately identifies relevant sentences, effectively compensating for the shortcomings of traditional methods with long-tail data. This comprehensively reflects the operating status of the power system, enhancing the professionalism and completeness of reports. By using preset sentence filtering rules to select relevant sentences from the retrieved sentence set and combining them using preset sentence combination templates, it generates high-quality power documents, ensuring clear document structure, semantic coherence, and compliance with industry standards. This significantly reduces manual writing workload, improves work efficiency, and lowers labor costs. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the steps of a power document generation method based on sentence segmentation retrieval in one embodiment of the present invention.

[0044] Figure 2 This is a structural block diagram of a power document generation system based on sentence segmentation retrieval in one embodiment of the present invention;

[0045] Figure 3 A structural diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0046] 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. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0047] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0048] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0049] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0050] One embodiment of the present invention provides a method for generating power documents based on sentence segmentation retrieval. For details, please refer to [link to relevant documentation]. Figure 1 , Figure 1 The flowchart shown is a step diagram of a power document generation method based on sentence segmentation retrieval in one embodiment of the present invention, including steps S11-S15:

[0051] S11. The power equipment image is encoded by a visual encoder based on the VLLM model structure to obtain the high-dimensional dense visual features of the power equipment image.

[0052] With the continuous development of power systems and the acceleration of intelligentization, the operation and maintenance management of power equipment faces higher requirements. The health status of power equipment is directly related to the stable operation of the power system, but traditional equipment inspection methods often rely on manual inspection, which is not only inefficient but also easily affected by human factors. Therefore, using advanced technologies for automatic inspection and condition assessment of power equipment has become an important research direction. In recent years, computer vision technology has made significant progress in many fields, and its application in power equipment inspection has gradually attracted attention. By encoding power equipment images using a visual encoder based on the VLLM (Vision-Language Large Model) model structure, high-dimensional dense visual features can be effectively extracted, providing strong support for the condition monitoring and fault diagnosis of power equipment. The specific implementation process is as follows:

[0053] In practical applications of power equipment inspection, the first step is to input the image data of the power equipment into a visual encoder based on the VLLM model structure. The VLLM model is a large model that integrates visual and language processing capabilities, with its visual encoder portion specifically dedicated to processing image data. This encoder can perform preliminary processing and analysis on the input power equipment images, laying the foundation for subsequent feature extraction and encoding. Specifically, the power equipment images in this embodiment contain rich information, including the equipment's appearance, the state of its components, and the surrounding environment. Inputting these images into the visual encoder transforms the complex information in the images into processable feature data.

[0054] The feature extraction layer inside the visual encoder is the key part for image feature extraction. When an image of electrical equipment enters the visual encoder, the feature extraction layer performs a series of processes to extract important features from the image.

[0055] This process is similar to the initial perception of objects by the human visual system, capable of recognizing basic visual elements such as edges, textures, and shapes in an image. Through the combination and analysis of these basic elements, an initial feature representation of the image can be obtained. Feature extraction in this stage primarily aims to filter out key information related to the state of electrical equipment from the original image, removing irrelevant noise and redundant data, thereby providing a clearer and more valuable data foundation for subsequent encoding processes. For example, for an image of a transformer, the feature extraction layer can identify features such as the shape of the transformer's casing, the texture of the heat sink, and the location of connecting wires.

[0056] After obtaining the initial feature representations, the encoding layer of the visual encoder further encodes these features. The role of the encoding layer is to transform the initial feature representations into more abstract and denser high-dimensional feature vectors. This process can be seen as a deep processing and compression of the features. Through complex neural network structures and algorithms, the information in the initial features is recombinated and optimized to better reflect the state characteristics of the power equipment. High-dimensional dense visual features have stronger expressive and discriminative capabilities, and can more accurately describe the differences in image features of power equipment under different states. For example, the high-dimensional dense visual features of normally operating power equipment and faulty power equipment will have significant numerical differences. These differences can be utilized by subsequent analysis models to achieve accurate judgment of the power equipment's state and fault diagnosis.

[0057] S12. Perform dimensionality reduction processing on the high-dimensional dense visual features to extract low-dimensional visual embedding information.

[0058] In the image processing workflow of power equipment, the high-dimensional dense visual features processed by the visual encoder need further processing to extract key information. Therefore, in this embodiment, these high-dimensional features are input into a specially designed abstraction module. The role of the abstraction module is to remove redundant information from the high-dimensional features through a series of complex processing steps, while retaining the information most critical to the state assessment of power equipment. The abstraction module can be regarded as an information filter, capable of identifying which features are useful for subsequent analysis and which can be ignored.

[0059] Within the abstract module, the multi-layer Transformer decoder is the core component for feature processing. The Transformer architecture is widely adopted due to its superior performance in natural language processing and computer vision. In this stage, high-dimensional, dense visual features are input into the multi-layer Transformer decoder, and each layer processes the features through a self-attention mechanism. This self-attention mechanism allows the model to dynamically pay attention to other relevant features while processing each feature, thereby better capturing the relationships and dependencies between features.

[0060] In this way, multi-layer Transformer decoders can progressively extract more abstract and high-level feature representations. These intermediate feature representations not only contain information from the original high-dimensional features but also enhance the semantic relationships between features through decoder processing. For example, in images of power equipment, some features may be related to the normal operating status of the equipment, while others may be related to potential fault modes.

[0061] After generating intermediate feature representations, the next step is to project these features into a low-dimensional space to extract low-dimensional visual embedding information. This is achieved through a projection layer in the abstraction module. The projection layer maps the high-dimensional intermediate feature representations to a low-dimensional space while preserving as much of the semantic and structural information of the original features as possible. Although the dimensionality of the features is reduced in the low-dimensional space, the low-dimensional embedding can still effectively represent the state of power equipment through the designed projection method. The projection layer typically uses linear or nonlinear transformation methods to achieve this goal. For example, principal component analysis (PCA) or other advanced dimensionality reduction techniques can be used to reduce the dimensionality of the features while maximizing the preservation of variance information in the features. In this way, the low-dimensional visual embedding information can represent the key features of power equipment images in a more compact form, enabling more efficient computation and analysis in subsequent applications, such as equipment condition assessment and fault diagnosis.

[0062] S13. Perform field recognition on the low-dimensional visual embedded information to generate a first field query vector and a second field query vector.

[0063] While low-dimensional visual embedding has effectively compressed feature dimensions, in order to further improve the understanding and analysis of the status of power equipment in power equipment image analysis, it is still necessary to perform more in-depth field recognition on these low-dimensional features to achieve topic recognition.

[0064] Multi-head self-attention (MSA) is a powerful feature extraction tool capable of capturing complex relationships and dependencies between features. In the topic segmentation encoder of this embodiment, MSA generates richer and more semantic topic feature representations by processing low-dimensional visual embeddings. Specifically, MSA allows the model to independently focus on different aspects of features in different subspaces, thereby enabling a more comprehensive understanding of the semantic information of the features. For example, in an image of power equipment, some features may be related to the normal operating status of the equipment, while others may be related to potential fault modes.

[0065] After generating the topic feature representations, the next step is to classify these features using the classification layer of the topic segmentation encoder. The role of the classification layer is to further subdivide the topic feature representations into common topics and rare topics.

[0066] Common themes typically refer to feature patterns that frequently appear in power equipment images and are related to the normal operating status of the equipment. Rare themes, on the other hand, are related to abnormal conditions or potential faults in the equipment. Through classification layer processing, two types of query vectors can be generated: a first-field query vector and a second-field query vector. The common-theme query vector serves as the first-field query vector, and the rare-theme query vector serves as the second-field query vector.

[0067] These two types of query vectors are used for subsequent retrieval and analysis, helping the system to more accurately identify and process the status of power equipment. For example, common topic query vectors can be used to quickly identify the normal operating status of equipment, while rare topic query vectors can be used to detect and warn of potential faults.

[0068] S14. In the power report training set, the first field query vector and the second field query vector are used as benchmarks to perform retrieval and extract the first statement set corresponding to the first field query vector and the second statement set corresponding to the second field query vector.

[0069] To fully utilize the information in the text data, the generated common topic query vectors and rare topic query vectors need to be matched with sentences in the power report training set, respectively.

[0070] The power report training set contains a large amount of textual information related to the operating status of power equipment. This information exists in the form of sentences, and each sentence can be converted into a sentence vector using natural language processing techniques. A sentence vector is a numerical representation of a sentence, reflecting its semantic content. By calculating the similarity between common topic query vectors, rare topic query vectors, and sentence vectors, the degree of matching between sentences and query vectors can be quantified. In this embodiment, similarity calculation can employ cosine similarity, Euclidean distance, or other suitable distance metrics. For example, cosine similarity measures the angle between two vectors; the smaller the angle, the higher the similarity.

[0071] The results of similarity calculations are used for filtering. In order to ensure that the extracted statements are highly relevant to the query vector, a preset threshold needs to be set during the filtering process.

[0072] The preset threshold is a standard used to determine whether the similarity is high enough. Only when the similarity between the sentence vector and the query vector is higher than this threshold will the sentence be selected. In this way, the system can filter out sentences from the power report training set that have a high degree of matching with query vectors of common and rare topics.

[0073] The selected statements contain detailed information related to the status of power equipment, such as operating parameters, fault descriptions, and maintenance records. These selected statements are combined into a statement set. The set of statements corresponding to common topic query vectors forms the first statement set, and the set of statements corresponding to rare topic query vectors forms the second statement set. This provides more comprehensive textual information support for the maintenance and fault diagnosis of power equipment. For example, if a rare topic query vector has a high match with a sentence describing a equipment fault, that sentence can be extracted, providing technicians with important clues for fault diagnosis.

[0074] S15. Based on the selected power document generation requirements, generate a power document corresponding to the power equipment image according to the first statement set and the second statement set.

[0075] Before generating electrical documentation, it's crucial to define the specific requirements. Different electrical documents focus on different information, such as equipment operating status, fault diagnosis, and maintenance records. Therefore, based on the document generation requirements, a set of key information keywords needs to be determined. These keywords are essential for selecting statements during document generation. If the document's purpose is to record the equipment's fault diagnosis process, the keyword set might include terms like "fault," "abnormality," "repair," and "testing." By identifying these keywords, it's ensured that the generated document accurately reflects the required information.

[0076] After determining the key information keyword set, the next step is to filter out statements containing these keywords from the extracted first and second statement sets. Each statement in the statement set may contain information related to the status of power equipment, but not all statements are directly relevant to the document generation requirements. Keyword filtering ensures that only statements highly relevant to the document requirements are selected. For example, if the keyword set contains "fault," then all statements containing the word "fault" will be filtered out.

[0077] After filtering out statements related to key information keywords, the next step is to combine these statements into a complete power document. To ensure the document's logic and readability, preset statement combination templates are used. These templates define the document's structure and format, such as how the beginning, body, and end should be organized. By combining the filtered statements according to the template structure, a complete and logically clear power document can be generated. For example, the template might stipulate that the beginning of the document should contain basic equipment information, the body should describe the equipment's operating status and fault conditions in detail, and the end should provide maintenance suggestions and follow-up measures. In this way, the system can automatically generate power documents that meet specific needs, improving the efficiency and quality of document generation. The final generated power document not only contains key information related to the status of power equipment but also presents it to users in an easy-to-understand and easy-to-use format, providing strong support for the maintenance and management of power equipment.

[0078] The present invention provides a power document generation method based on sentence segmentation retrieval. This method efficiently processes power equipment images through a visual encoder and abstraction module, extracting low-dimensional visual embeddings to significantly reduce redundant information and enhance feature expression capabilities. Combined with a topic segmentation encoder, it accurately generates query vectors for common and rare topics, enabling differentiated retrieval and improving retrieval accuracy and efficiency. For rare topics in the power field where samples are scarce but crucial, the method uses a topic segmentation encoder to accurately identify relevant sentences, effectively compensating for the shortcomings of traditional methods with long-tail data. This comprehensively reflects the operating status of the power system, enhancing the professionalism and completeness of reports. By selecting relevant sentences from the retrieved sentence set through preset sentence filtering rules and combining them using preset sentence combination templates, high-quality power documents are generated. This ensures clear document structure, semantic coherence, compliance with industry standards, significantly reduces manual writing workload, improves work efficiency, and lowers labor costs.

[0079] This invention also provides a power document generation system based on sentence segmentation retrieval, used to execute the power document generation method based on sentence segmentation retrieval described above. Figure 2 This is a structural block diagram of a power document generation system based on sentence segmentation retrieval according to an embodiment of the present invention. The system includes:

[0080] Image encoding module 21 is used to encode the power equipment image through a visual encoder based on the VLLM model structure to obtain the high-dimensional dense visual features of the power equipment image.

[0081] The feature dimensionality reduction module 22 is used to perform dimensionality reduction processing on the high-dimensional dense visual features and extract low-dimensional visual embedding information.

[0082] The vector generation module 23 is used to perform field recognition on the low-dimensional visual embedded information and generate a first field query vector and a second field query vector.

[0083] The statement retrieval module 24 is used to retrieve statements from the power report training set by using the first field query vector and the second field query vector as benchmarks, respectively, and to extract a first set of statements corresponding to the first field query vector and a second set of statements corresponding to the second field query vector.

[0084] The document generation module 25 is used to generate an electrical document corresponding to the image of the electrical equipment based on the selected electrical document generation requirements and according to the first statement set and the second statement set.

[0085] The vector generation module is used for:

[0086] The low-dimensional visual embedding information is used to extract features through the multi-head self-attention mechanism of the topic segmentation encoder to generate topic feature representations.

[0087] The topic feature representation is classified by the classification layer of the topic partitioning encoder to generate a first field query vector and a second field query vector.

[0088] The technical features and effects of the system proposed in the embodiments of the present invention are the same as those of the method proposed in the embodiments of the present invention, and will not be repeated here. Each module in the above system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0089] See Figure 3 This is a structural block diagram of a computer device provided in an embodiment of the present invention. The computer device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps described in the above embodiment of the power document generation method based on sentence segmentation retrieval. Figure 1 Steps S11 to S15 as described above; or, when the processor executes the computer program, it implements the functions of each module in the above-described device embodiments, such as modules 21 to 25 of the power document generation system based on sentence segmentation retrieval.

[0090] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the computer device.

[0091] The computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the schematic diagram is merely an example of a computer device and does not constitute a limitation on the computer device. It may include more or fewer components than illustrated, or combine certain components, or different components. For example, the computer device may also include input / output devices, network access devices, buses, etc.

[0092] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the computer device, connecting various parts of the computer device via various interfaces and lines.

[0093] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0094] If the modules integrated into the computer device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0095] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0096] Accordingly, embodiments of the present invention provide a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform steps in the power document generation method based on sentence segmentation retrieval as described in the above embodiments, for example... Figure 1 Steps S11 to S15 as described above.

[0097] In summary, compared with the prior art, the power document generation method, system, computer device, and computer-readable storage medium based on sentence segmentation retrieval provided by the embodiments of the present invention have the following beneficial effects:

[0098] This system efficiently processes power equipment images through a visual encoder and abstraction module, extracting low-dimensional visual embeddings to significantly reduce redundant information and enhance feature representation capabilities. Combined with a topic segmentation encoder, it accurately generates query vectors for common and rare topics, enabling differentiated retrieval and improving retrieval accuracy and efficiency. For rare topics in the power field where samples are scarce but crucial, the topic segmentation encoder accurately identifies relevant sentences, effectively compensating for the shortcomings of traditional methods with long-tail data. This comprehensively reflects the operating status of the power system, enhancing the professionalism and completeness of reports. By using preset sentence filtering rules to select relevant sentences from the retrieved sentence set and combining them using preset sentence combination templates, it generates high-quality power documents, ensuring clear document structure, semantic coherence, and compliance with industry standards. This significantly reduces manual writing workload, improves work efficiency, and lowers labor costs.

[0099] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A power document generation method based on sentence division search, characterized by, The method comprises the following steps: encoding the power equipment image through a visual encoder based on a VLLM model structure to obtain high-dimensional dense visual features of the power equipment image; dimension reduction processing is performed on the high-dimensional dense visual features to extract low-dimensional visual embedding information; field recognition is performed on the low-dimensional visual embedding information to generate a first field query vector and a second field query vector, including: feature extraction is performed on the low-dimensional visual embedding information through a multi-head self-attention mechanism of a topic division encoder to generate topic feature representation; the topic feature representation is classified through a classification layer of the topic division encoder to generate the first field query vector and the second field query vector; in the power report training set, the first field query vector and the second field query vector are respectively taken as the benchmark for retrieval to extract a first sentence set corresponding to the first field query vector and a second sentence set corresponding to the second field query vector; based on the selected power document generation requirement, the first sentence set and the second sentence set are used to generate a power document corresponding to the power equipment image.

2. The power document generation method based on sentence division search according to claim 1, wherein, The method comprises the following steps: inputting the power equipment image into the visual encoder based on the VLLM model structure; feature extraction is performed on the power equipment image through a feature extraction layer of the visual encoder to obtain initial feature representation; the initial feature representation is encoded through an encoding layer of the visual encoder to obtain high-dimensional dense visual features.

3. The power document generation method based on sentence division search according to claim 1, wherein, The method comprises the following steps: inputting the high-dimensional dense visual features into an abstraction module; the high-dimensional dense visual features are processed through a multi-layer Transformer decoder of the abstraction module to generate intermediate feature representation; the intermediate feature representation is projected to a low-dimensional space through a projection layer of the abstraction module to extract low-dimensional visual embedding information.

4. The power document generation method based on sentence division search according to claim 1, wherein, The method comprises the following steps: similarity calculation is performed between the first field query vector and the second field query vector and the sentence vector of the sentences in the power report training set; according to the similarity calculation result, sentences with a matching degree higher than a preset threshold value are screened out from the first field query vector and the second field query vector to form a first sentence set corresponding to the first field query vector and a second sentence set corresponding to the second field query vector.

5. The power document generation method based on sentence division search according to claim 1, wherein, The method comprises the following steps: determining a key information keyword set according to the power document generation requirement; screening a sentence containing any keyword in the set of keywords of the key information from the first and second sets of sentences; combining the screened sentences according to a preset sentence combination template to generate the power document.

6. A power document generation system based on sentence division search, characterized by, The method comprises: an image encoding module configured to encode the power equipment image through a visual encoder based on a VLLM model structure to obtain high-dimensional dense visual features of the power equipment image; a feature dimension reduction module configured to perform dimension reduction processing on the high-dimensional dense visual features to extract low-dimensional visual embedding information; a vector generation module configured to perform field identification on the low-dimensional visual embedding information to generate a first field query vector and a second field query vector; wherein the vector generation module is configured to: perform feature extraction on the low-dimensional visual embedding information through a multi-head self-attention mechanism of a topic division encoder to generate topic feature representation; and perform classification on the topic feature representation through a classification layer of the topic division encoder to generate the first field query vector and the second field query vector; a sentence retrieval module configured to perform retrieval in a power report training set respectively using the first field query vector and the second field query vector as a reference to extract a first set of sentences corresponding to the first field query vector and a second set of sentences corresponding to the second field query vector; a document generation module configured to generate a power document corresponding to the power equipment image according to the first set of sentences and the second set of sentences based on selected power document generation requirements.

7. A computer device, comprising: The computer readable storage medium stores a computer program, wherein when a device in which the computer readable storage medium is located executes the computer program, the power document generation method based on sentence retrieval as claimed in any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, wherein when a device in which the computer readable storage medium is located executes the computer program, the power document generation method based on sentence retrieval as claimed in any one of claims 1 to 5 is implemented.

Images