Field source labeling method, control device and cloud server

Abstract

The application provides a field source labeling method, a control device and a cloud server, relates to the technical field of fault diagnosis, and the field source labeling method comprises the following steps: obtaining a structured field template and multi-source data; constructing a source labeling prompt word based on the structured field template and the multi-source data, in the source labeling prompt word, the multi-source data is marked and distinguished by a source identifier; inputting the source labeling prompt word into a large language model, triggering the large language model to extract field values of each field from the multi-source data, labeling the source types of each information segment in the field values, and generating a field extraction result with segment source labeling; and outputting the field extraction result with segment source labeling, so that the field values of each field are associated with the corresponding source type labeling and are displayed. The application realizes the credibility classification of direct extraction content and model reasoning content, and improves the accuracy of blockchain server fault maintenance decision and the information traceability.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis technology, and in particular to a field source annotation method, control device and cloud server. Background Technology

[0002] In the fault repair process of blockchain servers and their supporting computing boards, fans, power supplies, and other equipment, repair reports need to extract information from multiple data sources to fill in various fields. These data sources include at least the multi-round repair dialogue text between repair personnel and the diagnostic system, the recognized text obtained from optical character recognition of images of faulty equipment taken by repair personnel, and the query data returned by the manufacturing execution system for this repair operation. The query data may include equipment model, historical repair records, and inventory spare parts information. Currently, a large language model is typically used to simultaneously extract field values ​​from the above multi-source data to achieve automated filling of repair reports.

[0003] However, when large language models jointly extract data from multiple sources, the output field value is a complete text value. For example, the extracted value for the fault cause field might be "Continuous overcurrent at the output terminal caused thermal breakdown of the switching transistor, and the decrease in cooling fan speed exacerbated the temperature rise." When reviewing this field value, maintenance personnel cannot distinguish that the first half, "Continuous overcurrent at the output terminal caused thermal breakdown of the switching transistor," comes from the maintenance personnel's verbal description of the fault phenomenon in the maintenance dialogue, while the second half, "The decrease in cooling fan speed exacerbated the temperature rise," comes from the causal inference made by the large language model based on the context of the circuit board's thermal distribution. Because the reliability of model inferences is lower than that of content directly extracted from the raw data, indiscriminately mixing information fragments from different sources in the same field value can cause confusion regarding the reliability of information: maintenance personnel may treat the model's predictions as actual measured data to guide subsequent maintenance operations, or they may doubt all the information in the entire field value due to distrust of the model's inferences, affecting the accuracy of maintenance decisions. Summary of the Invention

[0004] The main purpose of this application is to provide a field source annotation method, which aims to achieve credibility classification of directly extracted content and model inference content, thereby improving the accuracy of blockchain server fault repair decisions and information traceability.

[0005] To achieve the above objectives, this application provides a field source annotation method, which includes: Retrieve structured field templates and multi-source data; Source labeling prompts are constructed based on the structured field template and the multi-source data, wherein the multi-source data is marked and distinguished by source identifiers in the source labeling prompts; The source-labeled prompt words are input into the large language model, which triggers the large language model to extract the field values ​​of each field from the multi-source data, and labels the source type of each information fragment in the field value, generating field extraction results with fragment source labels; Output the field extraction results with fragment source labels, so that the field values ​​of each field are associated with the corresponding source type labels and displayed.

[0006] Optionally, obtaining the structured field template and multi-source data specifically involves: Obtain the text of the troubleshooting dialogue; Perform optical character recognition on the fault image to generate optical character recognition text; Retrieve query data returned by the Manufacturing Execution System; The fault repair dialogue text, the optical character recognition text, and the query data are used as the multi-source data.

[0007] Optionally, the multi-source data is marked and distinguished by a source identifier, specifically: Configure a dialogue identifier for the fault repair dialogue text, configure an image identifier for the optical character recognition text, and configure a system data identifier for the query data.

[0008] Optionally, the source type of each information fragment in the labeled field value is specifically as follows: When the information fragment originates from the fault repair dialogue text, the source type of the information fragment is marked as the dialogue extraction type; When the information fragment originates from the optical character recognition text, the source type of the information fragment is labeled as image recognition type; When the information fragment originates from the queried data, the source type of the information fragment is labeled as the system data extraction type; If the information fragment does not originate directly from the multi-source data, the source type of the information fragment is labeled as the model inference type.

[0009] Optionally, configuring a dialogue identifier for the fault repair dialogue text, an image identifier for the optical character recognition text, and a system data identifier for the query data includes: An open tag is configured at the beginning of the corresponding data text, and a closed tag is configured at the end of the corresponding data text.

[0010] Optionally, obtaining the structured field template specifically involves: Retrieve multiple field definitions, each of which includes a field identifier and a semantic description of the field. The structured field template is generated based on the multiple field definitions.

[0011] Optionally, the step of constructing source annotation prompts based on the structured field template and the multi-source data specifically involves: The role setting section, field definition list section, multi-source data section, and output format constraint section are written sequentially in the source label prompt text; The field definition list segment is generated based on the structured field template.

[0012] Optionally, the source type of each information fragment in the labeled field value is specifically as follows: Determine the source attribution of the information fragment within the multi-source data; A source type label corresponding to the source attribution is appended after the information fragment.

[0013] Optionally, the step of associating the field values ​​of each field with the corresponding source type annotation for display specifically involves: The field value of each field in the target report is associated with the source type of each information fragment in that field value and displayed in the same display area.

[0014] Optionally, the step of associating the field values ​​of each field with the corresponding source type annotation for display specifically involves: When the field value of the target field consists of multiple information fragments from different source types, the multiple information fragments are arranged alternately with their respective source type labels and displayed in association.

[0015] Optionally, triggering the large language model to extract field values ​​from the multi-source data specifically involves: If the large language model does not extract any information fragment that matches the target field, an empty field value is output for the target field, and a source missing marker is attached.

[0016] Optionally, the step of associating the field values ​​of each field with the corresponding source type annotation for display specifically involves: Retrieve source filtering instructions that carry the target source type; Fields whose values ​​contain source type annotations corresponding to the target source type are identified as fields to be displayed; The field value of the field to be displayed is associated with the corresponding source type label and displayed.

[0017] Optionally, the step of associating the field values ​​of each field with the corresponding source type annotation for display specifically involves: Obtain the source tracing operation instructions to be executed on the target source type annotation; Locate the original data position of the information fragment corresponding to the target source type label in the multi-source data; The display context will jump to the original data location, where the display context includes the current view and the display location.

[0018] Optionally, locating the original data location of the information fragment corresponding to the target source type label in the multi-source data specifically involves: Based on the source type labeled with the target source type, retrieve data segments in the multi-source data that semantically match the information fragment; Determine the start and end positions of the information segment within the data segment; The step of jumping the display context to the original data location specifically involves: The display context will be switched from the report view to the data view of the data source to which the information fragment belongs, and the display will be positioned at the start and end points.

[0019] In addition, to achieve the above objectives, this application also provides a control device, which includes: a memory, a processor, and a field source annotation program stored in the memory and executable on the processor, the field source annotation program being configured to implement the field source annotation method as described above.

[0020] In addition, to achieve the above objectives, this application also provides a cloud server, including the control device described above.

[0021] This application embodiment obtains structured field templates and multi-source data, then constructs source annotation prompts based on these templates and data. In these prompts, multi-source data is distinguished by source identifiers. The source annotation prompts are then input into a large language model, triggering the model to extract field values ​​from the multi-source data and annotate the source type of each information fragment within those values. This generates field extraction results with fragment source annotations and outputs the results, linking each field value to its corresponding source type annotation. Thus, by labeling and partitioning multi-source data using source identifiers, the large language model simultaneously completes field extraction and fragment source annotation in a single inference iteration. This allows for the distinguishable display of information fragments from different sources within field values, enabling maintenance personnel to assess the credibility of each information fragment based on the source type annotations. This achieves credibility grading between directly extracted content and model-inferred content, improving the accuracy and traceability of blockchain server fault repair decisions. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0023] Figure 1 This is a schematic flowchart of a field source annotation method according to an embodiment of this application; Figure 2 for Figure 1 A schematic diagram of the annotation method for step S100; Figure 3 A flowchart illustrating the method for configuring the source identifier; Figure 4 This is a flowchart illustrating the method for determining the source type. Figure 5 for Figure 3 A flowchart illustrating the field source annotation method in step S210; Figure 6 A flowchart illustrating the method for generating structured field templates; Figure 7 for Figure 1 A schematic diagram of the annotation method for step S200; Figure 8 for Figure 1 A schematic diagram of the annotation method for step S300; Figure 9 A flowchart illustrating the method for displaying related data within the same region; Figure 10 A flowchart illustrating the method of alternating multiple source fragments; Figure 11 A flowchart illustrating the method for handling empty field values ​​and missing source markers; Figure 12 for Figure 1 A flowchart illustrating the annotation method for step S400. Figure 13 This is a flowchart illustrating the method of source tracing and redirection; Figure 14 This is a flowchart illustrating the method for tracing the source, locating the location, and switching views.

[0024] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Well-known modules, units, and their connections, links, communications, or operations are not shown or described in detail. Furthermore, the described features, architectures, or functions can be combined in any way in one or more embodiments. Those skilled in the art should understand that the various embodiments described below are only for illustration and not for limiting the scope of protection of this application. It is also readily understood that the modules or units or processing methods in the various embodiments shown herein and in the accompanying drawings can be combined and designed in various different configurations. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Because in the existing technology, in the process of filling out maintenance reports for blockchain servers and their supporting computing boards, fans, power supplies, etc., the large language model extracts field values ​​from multiple sources of data and mixes information fragments from different sources into a single text value output. Maintenance personnel cannot distinguish whether the information in a field value comes from the original data extraction or the model's own reasoning.

[0027] The main solution of this application embodiment is: when constructing input prompt words for a large language model, mark and distinguish multi-source data with source identifiers, so that the large language model can trace the source of each information fragment in the multi-source data while extracting the values ​​of each field, and add corresponding source type labels to the information fragments during output, so that each information fragment of the field value can be traced back to the original data source.

[0028] In this embodiment, for ease of description, the following description will focus on the control device as the executing entity.

[0029] This application provides a solution that uses source identifiers to label and partition multi-source data and drives a large language model to complete field extraction and fragment source labeling in a single inference. This enables the differentiation and traceability of information fragments from different sources in field values, allowing maintenance personnel to distinguish the credibility level of directly extracted content and model-inferred content based on source type, thereby improving the accuracy of maintenance decisions.

[0030] Therefore, this application proposes a field source annotation method; it is understood that the cloud server is equipped with a control device for storing and executing the following method. The control device can be implemented using a main controller, such as MCU (Microcontroller Unit), DSP (Digital Signal Processor), FPGA (Field Programmable Gate Array), SOC (System On Chip), etc.

[0031] In existing technologies, fault repair reports for blockchain servers and their associated computing boards, fans, power supplies, and other equipment typically use large language models to jointly extract field values ​​from multiple data sources for automatic filling. However, the field values ​​output by the large language model are complete text values, mixed with different information fragments from repair dialogues, image recognition, system queries, and model inference. Repair personnel cannot distinguish the source of each fragment and are prone to misinterpreting the model's inference as actual test data, affecting the accuracy of repair decisions.

[0032] Therefore, referring to Figure 1 In one embodiment of this application, the field source annotation method includes steps S100-S400, wherein: S100, Obtain structured field templates and multi-source data; S200. Construct source labeling prompts based on structured field templates and multi-source data. In the source labeling prompts, multi-source data are marked and distinguished by source identifiers. S300. Input the source annotation prompt words into the large language model, trigger the large language model to extract the field values ​​of each field from the multi-source data, and annotate the source type of each information fragment in the field value, generating field extraction results with fragment source annotations; S400 Output the field extraction results with fragment source labels, so that the field values ​​of each field are associated with the corresponding source type labels and displayed.

[0033] Here, the structured field template refers to the set of definitions for each target field that needs to be extracted and populated from multi-source data in the maintenance report; multi-source data refers to data from multiple different data sources, including fault maintenance dialogue text, optical character recognition (OCR) text obtained by performing optical character recognition on fault images, and query data returned by the manufacturing execution system; source identifier refers to the distinguishing mark attached to each data source in the source annotation prompt, used to enable the large language model to distinguish the data source of each text segment when reading multi-source data; source annotation prompt refers to the prompt text input to the large language model, which includes the structured field template, multi-source data, and the source identifier corresponding to each data source; field value refers to the content corresponding to the field extracted by the large language model from multi-source data; information fragment refers to the component in the field value that can independently identify the source attribution; source type refers to the classification of the source attribution of the information fragment in multi-source data, including dialogue extraction type, image recognition type, model inference type, and system data extraction type; source type annotation refers to the mark attached to the information fragment to indicate the source type of the information fragment; fragment source annotation refers to the combined annotation form of the information fragment and its source type annotation.

[0034] Secondly, the control device acquires a structured field template, which includes seven field definitions: equipment model, fault type, fault location, fault cause, repair plan, required spare parts, and estimated repair time. Each field definition includes a field identifier and a semantic description text. Simultaneously, it receives fault repair dialogue text (e.g., a repairman says, "There is a burning smell at the P2 position of the power module. After disconnecting the output, it was found that the switching transistor has been shorted"), optical character recognition text obtained from fault images through optical character recognition (e.g., "FAN SPEED: 3600RPM" recognized from a circuit board photo), and query data returned by the manufacturing execution system (e.g., the part number TH-PSU-800W of the power module corresponding to the P2 slot in the bill of materials for equipment model TH-5000), which together constitute multi-source data. Subsequently, the control device constructs source-annotated prompts based on structured field templates and multi-source data. Within the prompt text, a role setting section, a field definition list section, a multi-source data section, and an output format constraint section are sequentially written. In the multi-source data section, dialogue identifiers, image identifiers, and system data identifiers are used to distinguish fault repair dialogue text, optical character recognition text, and query data, respectively. Dialogue identifiers use paired open / closed markers “[dialogue]…[ / dialogue]” to enclose dialogue text; image identifiers use “[image]…[ / image]” to enclose recognition text; and system data identifiers use “[system]…[ / system]” to enclose query data. Finally, the control device inputs the source-annotated prompts into a large language model, triggering the model to extract field values ​​from the multi-source data and annotate the source type of each information fragment within the field values. For example, the extracted value for the fault cause field is “Continuous overcurrent at the output terminal caused thermal breakdown of the switching transistor [dialogue], and the decrease in cooling fan speed exacerbated the temperature rise [inference]”. The model generates and outputs field extraction results with fragment source annotations, thus associating the field values ​​of each field with their corresponding source type annotations.

[0035] Through the above technical solution, this embodiment marks and partitions multi-source data with source identifiers, enabling the large language model to simultaneously complete field extraction and fragment source labeling in a single inference. This transforms field values ​​from indivisible plain text into labeled text whose source can be traced fragment by fragment, achieving the distinguishable display of information fragments from different sources. This allows maintenance personnel to judge the credibility of each information fragment based on the source type labeling, avoiding mistaking the model's inferred content for actual measured data, thereby improving the accuracy of maintenance decisions.

[0036] This embodiment drives a large language model to simultaneously complete field extraction and fragment source labeling in a single inference by marking and partitioning multi-source data with source identifiers. In existing technologies, when large language models extract report field values ​​from multi-source data, they typically output the field values ​​as indivisible plain text. The source credibility issue is either ignored or resolved by constructing independent verification prompts to trigger a second model call, resulting in additional token overhead and inference latency. Furthermore, the outputs of the two calls are prone to misalignment in the correspondence between field values ​​and source labels. Existing technologies follow the two-stage processing convention of "extract first, verify later," without considering completing both tasks simultaneously in a single call. This often leads to the failure to consider marking and partitioning multi-source data with source identifiers at the prompt level, making the source information part of the model's inference context. Compared with existing technologies, this embodiment replaces the two-stage processing method of extraction followed by independent verification with source partitioning at the prompt word level. This allows the large language model to naturally know which marked region each information fragment comes from while extracting it. This enables field extraction and fragment-level source labeling to be completed in a single call without adding extra token overhead. Furthermore, the field values ​​and source labels are naturally aligned, and there is no misalignment problem between the two calls.

[0037] This application embodiment obtains structured field templates and multi-source data, then constructs source annotation prompts based on these templates and data. In these prompts, multi-source data is distinguished by source identifiers. The source annotation prompts are then input into a large language model, triggering the model to extract field values ​​from the multi-source data and annotate the source type of each information fragment within those values. This generates field extraction results with fragment source annotations and outputs the results, linking each field value to its corresponding source type annotation. Thus, by labeling and partitioning multi-source data using source identifiers, the large language model simultaneously completes field extraction and fragment source annotation in a single inference iteration. This allows for the distinguishable display of information fragments from different sources within field values, enabling maintenance personnel to assess the credibility of each information fragment based on the source type annotations. This achieves credibility grading between directly extracted content and model-inferred content, improving the accuracy and traceability of blockchain server fault repair decisions.

[0038] There are no specific limitations on the composition and acquisition methods of multi-source data. For example, how to acquire and organize fault repair dialogue text, fault image recognition text, and query data from manufacturing execution systems into multi-source data? If the specific source and acquisition path of each data source are not clear, the source semantics of each data segment cannot be determined when constructing source labeling prompts, and the accuracy of source labeling lacks a data foundation.

[0039] Therefore, referring to Figure 2Another embodiment of this application provides a field source annotation method, based on the above. Figure 1 The embodiment shown obtains the structured field template and multi-source data, specifically through steps S110-S140, wherein: S110, Obtain the fault repair dialogue text; S120. Perform optical character recognition on the fault image to generate optical character recognition text; S130. Obtain the query data returned by the manufacturing execution system; S140. Use fault repair dialogue text, optical character recognition text, and query data as multi-source data.

[0040] Among them, fault repair dialogue text refers to the textual representation of all interaction records generated between the maintenance personnel and the fault diagnosis system regarding the current fault repair process; fault images refer to the appearance images of the faulty equipment taken by the maintenance personnel during the maintenance process, including circuit board photos and screenshots of the fault diagnosis interface; optical character recognition text refers to the text content obtained after performing optical character recognition processing on the fault images, including the component model silkscreen identified from the circuit board photos, the cooling fan speed display value, and the diagnostic prompt text in the fault diagnosis interface screenshots; query data refers to the structured data related to this maintenance obtained by the control device through the interface provided by the manufacturing execution system, including the equipment model, the equipment's historical maintenance records, the spare parts part numbers and inventory information corresponding to each component in the equipment's bill of materials.

[0041] Secondly, the control device acquires the fault repair dialogue text, such as a repairman stating, "There is a burning smell at the P2 position of the power module. After disconnecting the output, it was found that the switching transistor is short-circuited, and the fan speed label at the same position shows only 60% of the rated value." Subsequently, the control device performs optical character recognition (OCR) on the fault image, recognizing "FAN SPEED: 3600RPM (rated 6000RPM)" and the parameter text on the cooling fan nameplate from the circuit board photo, generating OCR text. Simultaneously, it retrieves query data by calling the manufacturing execution system interface; for example, in the bill of materials for equipment model TH-5000, the part number of the power module corresponding to the P2 slot is TH-PSU-800W. Finally, the control device uses the fault repair dialogue text, OCR text, and query data together as multi-source data for subsequent construction of source annotation prompts.

[0042] Through the above technical solution, this embodiment clarifies the specific acquisition methods of the three data sources and organizes them into multi-source data, so that the subsequent prompt word construction steps have a clear source semantic basis, realizes the distinguishable organization of multi-source data source types, and provides a data source basis for fragment-level source annotation.

[0043] This embodiment concretizes multi-source data into three data sources: fault repair dialogue text, optical character recognition text, and query data, and clarifies the acquisition path of each, providing a definite semantic basis for source labeling. In existing technologies, when large language models jointly extract field values ​​from multi-source data, the composition of the multi-source data is usually not distinguished. The model treats all input as homogeneous text, and even if field values ​​are output, it is impossible to trace the source of each information fragment. When processing automatic report filling, existing technologies often directly concatenate multi-source data and input it into the model at once, without acquiring and organizing the data sources in a typological manner, resulting in the model lacking an input basis for distinguishing data sources. Compared with existing technologies, this embodiment acquires dialogue text, recognition text, and query data separately and organizes them into multi-source data, so that each data source has a definite type semantic on the input side, providing a data basis for subsequent labeling of partitions with source identifiers, and realizing a closed correspondence between data source type and source labeling semantics.

[0044] If the source identifiers corresponding to different data sources are not consistent, the large language model cannot distinguish the source of each text segment when reading multi-source data, and the input-side basis of source labeling is unclear.

[0045] Therefore, referring to Figure 3 Another embodiment of this application provides a method for labeling the source of a field, based on the above. Figure 2 In the illustrated embodiment, multi-source data is distinguished by a source identifier, specifically in step S210, wherein: S210. Configure dialogue identifiers for fault repair dialogue text, configure image identifiers for optical character recognition text, and configure system data identifiers for query data.

[0046] Among them, dialogue identifier refers to the source identifier used to mark fault repair dialogue text; image identifier refers to the source identifier used to mark optical character recognition text; system data identifier refers to the source identifier used to mark query data; source identifier refers to the distinguishing mark added to each data source in the source annotation prompt, so that the large language model can distinguish which data source each piece of text belongs to when reading multi-source data.

[0047] Secondly, when constructing source labeling prompts, the control device configures dialogue identifiers for fault repair dialogue text, image identifiers for optical character recognition text, and system data identifiers for manufacturing execution system query data. Subsequently, each data source embeds its corresponding source identifier into the multi-source data segment of the source labeling prompts, enabling the large language model to use these identifier-separated multi-source data as the basis for determining the source of each extracted segment. Finally, the identifier configuration for the three data sources is completed, forming multi-source data input marked with source identifiers.

[0048] Through the above technical solution, this embodiment establishes a one-to-one correspondence between the input source identifier and the data source by configuring dialogue identifiers, image identifiers and system data identifiers for the three data sources respectively, realizing the distinguishable marking of multi-source data in prompt words, and providing input-side basis for subsequent fragment-level source annotation.

[0049] This embodiment configures corresponding source identifiers for each of the three data sources, enabling multi-source data to carry source type semantics when inputting prompts. In existing technologies, large language models typically concatenate data segments directly when processing multi-source data, lacking source markers between segments. This prevents the model from distinguishing whether text originates from dialogue, image recognition, or system queries, resulting in extracted field values ​​lacking source information. Existing technologies often focus on task instructions and output formats when constructing input for large language models, neglecting to differentiate source identifiers for each data source at the input data level. Compared to existing technologies, this embodiment configures dedicated source identifiers for each data source, allowing the model to obtain clues about data source distinctions during the reading phase. This achieves semantic correspondence between input-side source markers and output-side source annotations, providing an identifier foundation for simultaneous extraction and annotation in a single inference iteration.

[0050] The specific classification system for the source type of information fragments in field values ​​has not yet been defined. If the source type classification is ambiguous, the large language model may produce ambiguous annotation results when determining the source attribution of fragments.

[0051] Therefore, referring to Figure 4 In another embodiment of this application, a method for annotating the source of a field is provided, based on the above. Figure 2 The illustrated embodiment annotates the source type of each information fragment in the field value, including steps S310-S340, wherein: S310. When the information fragment originates from the fault repair dialogue text, the source type of the information fragment is marked as the dialogue extraction type. S320. When the information fragment originates from optical character recognition text, the source type of the information fragment is labeled as image recognition type; S330. When the information fragment originates from the queried data, the source type of the information fragment is marked as the system data extraction type. S340. When an information fragment does not directly originate from multi-source data, the source type of the information fragment is labeled as the model inference type.

[0052] Among them, the dialogue extraction type refers to information fragments that have a clear corresponding original text description in the fault repair dialogue text, and the large language model extracts the information fragments by directly extracting or summarizing them from the dialogue text; the image recognition type refers to information fragments that come from the optical character recognition text of fault images; the system data extraction type refers to information fragments that come directly from a field value in the query data returned by the manufacturing execution system, and the extraction method is to directly copy them from the structured query results; the model inference type refers to information fragments that are not explicitly contained in the fault repair dialogue text, optical character recognition text, and query data, and the large language model performs cross-data source causal inference or semantic combination based on the context information in the multi-source data.

[0053] Secondly, during the extraction of field values ​​from each field, the large language model performs source attribution determination for each information fragment. Subsequently, when an information fragment finds a corresponding original text expression in the fault repair dialogue text, the information fragment is labeled as a dialogue extraction type; when the information fragment originates from optical character recognition text, it is labeled as an image recognition type; and when the information fragment is directly taken from field values ​​in the query data, it is labeled as a system data extraction type. Finally, when an information fragment has no direct correspondence in any of the three data sources and is generated by the model based on contextual inference, it is labeled as a model inference type. For example, the fault cause field value "the decrease in cooling fan speed exacerbated the temperature rise" has no direct expression in the dialogue text, recognition text, or query data, and is generated by the model based on the contextual inference of fan speed data and circuit board thermal distribution, and is labeled as a model inference type.

[0054] Through the above technical solution, this embodiment establishes four types of source type determination rules, enabling the large language model to have a clear classification basis for the source attribution of each information fragment, realizing the standardization of source annotation semantics, and enabling maintenance personnel to clearly distinguish between directly extracted content and model-inferred content.

[0055] This embodiment establishes four source type determination rules for dialogue extraction, image recognition, system data extraction, and model inference, enabling fragment-level source labeling to have definite classification semantics. In existing technologies, field values ​​extracted by large language models do not distinguish information sources; model-inferred content and directly extracted content are mixed in the output, making it impossible for maintenance personnel to determine the credibility level of the field values. Existing technologies typically only require the model to output the field values ​​themselves in information extraction tasks, without requiring the model to explicitly classify and label the source type of the output content, and the model itself does not actively distinguish the source attribution of the output fragments. Compared to existing technologies, this embodiment defines four source type determination rules for information fragments, particularly explicitly labeling model-inferred content as the model inference type. This allows maintenance personnel to distinguish the credibility level of directly extracted information from inferred information based on the source type, avoiding the use of inferred content as actual test data, and achieving standardization and credibility grading of source labeling semantics.

[0056] There are no specific restrictions on the marking format of source identifiers. If the identifier format is not uniform, the large language model may not be able to accurately identify the boundaries of each identifier when parsing multi-source data, affecting the effectiveness of the marking partition.

[0057] Therefore, referring to Figure 5 This application also provides a method for labeling the source of a field, based on the above. Figure 3 The illustrated embodiment configures a dialogue identifier for fault repair dialogue text, an image identifier for optical character recognition text, and a system data identifier for query data, including step S211, wherein: S211. Configure an open tag at the beginning of the corresponding data text and a closed tag at the end of the corresponding data text.

[0058] Among them, the opening marker refers to the identifier start marker configured at the beginning of the data text; the closing marker refers to the identifier end marker configured at the end of the data text; and the corresponding data text refers to the fault repair dialogue text, optical character recognition text, and query data corresponding to each source identifier.

[0059] Secondly, when configuring dialogue identifiers for fault repair dialogue text, the control device configures a dialogue open marker "[dialogue]" at the beginning of the dialogue text and a dialogue close marker "[ / dialogue]" at the end, forming a paired enclosed form of "[dialogue]...[ / dialogue]". Subsequently, for optical character recognition text, an image open marker "[image]" and an image close marker "[ / image]" are configured; for query data, a system data open marker "[system]" and a system data close marker "[ / system]" are configured. Finally, all three data sources are enclosed in paired open and close markers and then embedded with source annotation prompts, enabling the large language model to accurately identify the boundaries and sources of each data segment.

[0060] Through the above technical solution, this embodiment uses paired opening and closing tags to wrap the beginning and end of the data text, giving the source identifier a clear boundary semantics, thus achieving accurate partitioning of multi-source data in prompt words and avoiding ambiguity in model parsing.

[0061] This embodiment uses paired open and closed tags to wrap the beginning and end of the data text, making the boundaries of the source identifiers clearly identifiable. In existing technologies, multi-source data in large language model prompts is usually separated by ordinary text concatenation or simple delimiters, lacking clear paired boundary markers between data segments. This makes it easy for the model to confuse the start and end positions of data segments in long text inputs. Compared to existing technologies, this embodiment uses paired open and closed tags to wrap each data text, enabling the model to accurately identify the boundaries and source attribution of each data segment, providing a clear input-side partitioning structure for fragment-level source annotation.

[0062] The specific structure of the structured field template has not yet been specified. If the field definition lacks clear field identifiers and semantic descriptions, the large language model cannot accurately understand the extraction target of each field, affecting the accuracy of field extraction.

[0063] Therefore, referring to Figure 6 Another embodiment of this application provides a field source annotation method, based on the above. Figure 1 The embodiment shown obtains the structured field template, specifically through steps S150-S160, wherein: S150. Obtain multiple field definitions, each field definition containing a field identifier and a field semantic description text; S160. Generate structured field templates based on multiple field definitions.

[0064] Among them, field definition refers to the data structure that describes the extraction requirements of a target field in the maintenance report; field identifier refers to the unique identifier name of the field, such as "fault cause"; field semantic description text refers to the natural language description of the meaning and extraction requirements of the field, which is used to enable the large language model to understand the scope of content that needs to be extracted from the field.

[0065] Secondly, the control device acquires multiple field definitions, such as seven fields: equipment model, fault type, fault location, fault cause, maintenance plan, required spare parts, and estimated maintenance time. Each field definition includes a field identifier and a semantic description text; for example, the semantic description text for the fault cause field is "extract the root cause description of the equipment failure." Subsequently, the control device organizes these field definitions according to a report layout and generates a structured field template based on all field definitions. Finally, the structured field template is used to construct the field definition list segment in the subsequent source annotation prompts.

[0066] Through the above technical solution, this embodiment uses field identifiers and field semantic description text to construct field definitions and generate structured field templates, enabling the large language model to accurately understand the extraction targets of each field and realizing the semantic clarification of the field extraction task.

[0067] This embodiment organizes field definitions using field identifiers and semantic description text to generate structured field templates, enabling the model to accurately understand the extraction semantics of each field. In existing technologies, large language models typically only indicate the extraction target with a simple list of field names when extracting information. The model's understanding of the field meaning relies on the name itself, and when the field name cannot fully express the extraction scope, omissions or errors in extraction are likely. Compared to existing technologies, this embodiment defines additional semantic description text for each field, making the extraction requirements explicit. This allows the model to understand the extraction scope based on the semantic description, achieving semantic clarity of the field extraction target and providing an accurate field definition foundation for subsequent construction of prompt words based on structured field templates.

[0068] The internal structure of source annotation prompts has not been defined. If the functional segments of the prompts are not clearly organized, the large language model cannot accurately perform field extraction and source annotation tasks.

[0069] Therefore, referring to Figure 7 Another embodiment of this application provides a method for labeling the source of a field, based on the above. Figure 1 The illustrated embodiment constructs source tagging prompts based on structured field templates and multi-source data, specifically in step S220, where: S220. Write the role setting section, field definition list section, multi-source data section and output format constraint section in sequence in the source annotation prompt text; The field definition list section is generated based on a structured field template.

[0070] Among them, the role setting segment refers to the text segment in the prompt words used to set the role and task objectives of the large language model; the field definition list segment refers to the text segment generated based on the structured field template, which lists the definitions of each field and their semantic descriptions one by one; the multi-source data segment refers to the text segment containing multi-source data distinguished by source identifiers; and the output format constraint segment refers to the text segment in the prompt words used to constrain the output format of the large language model, requiring the model to label the source type of each information fragment in the field value when extracting each field value.

[0071] Secondly, the control device constructs source annotation prompts, sequentially writing a role setting segment, a field definition list segment, a multi-source data segment, and an output format constraint segment into the prompt text. Subsequently, the role setting segment sets the model to perform report field extraction tasks as an equipment maintenance expert; the field definition list segment lists the definitions and semantic descriptions of seven fields based on a structured field template; the multi-source data segment embeds three data sources distinguished by source identifiers; and the output format constraint segment requires the model to label the source type for each information fragment when extracting field values. Finally, the four organized source annotation prompts serve as input to the large language model.

[0072] Through the above technical solution, this embodiment organizes the source annotation prompts into a four-segment structure: role setting segment, field definition list segment, multi-source data segment, and output format constraint segment. This enables the model to perform extraction and annotation under the premise of clearly defining the task role, field target, data source, and output requirements, thus realizing the modular organization of the prompt function.

[0073] This embodiment organizes source annotation prompts into a four-segment structure: role setting segment, field definition list segment, multi-source data, and output format constraints. This provides complete functional support for field extraction and source annotation tasks at the prompt word level. In existing technologies, prompts for information extraction in large language models typically only contain task instructions and data inputs, and output format constraints often do not involve source annotation requirements; the model output is only the field values ​​themselves. Compared to existing technologies, this embodiment explicitly requires the model to label the source type for each information fragment by setting an output format constraint segment. This integrates the source annotation task into the functional structure of the prompt words, enabling the model to complete extraction and annotation simultaneously in a single inference iteration, thus achieving complete support for dual-output tasks through prompt words.

[0074] There are no specific restrictions on the annotation method for the source of the fragment. If the annotation method is an additional data structure independent of the field value, the field value and the annotation need to be rendered separately and then combined when displaying, which increases the complexity of the front-end display logic.

[0075] Therefore, referring to Figure 8 In another embodiment of this application, a method for annotating the source of a field is provided, based on the above. Figure 1 The embodiment shown indicates that the source type of each information fragment in the labeled field value is determined in steps S350-S360, where: S350. Determine the source and attribution of information fragments in multi-source data; S360. Add a source type label corresponding to the source attribution after the information fragment.

[0076] Among them, source attribution refers to the source correspondence determined by the large language model when extracting the information fragment by comparing the semantics of the information fragment with the text content within the area enclosed by each source identifier; source type labeling refers to the visual mark attached to the end of the information fragment to indicate the source type of the information fragment.

[0077] Secondly, during the field value extraction process, the large language model performs source attribution determination for each information fragment. For example, the semantics of the information fragment "TH-5000" perfectly matches "Device Model: TH-5000" within the "[System]" tag area of ​​the multi-source data, thus determining its source attribution to the system data extraction type. Subsequently, the source type label is directly appended to the information fragment text in the model's output field value text. For example, if the information fragment is written as "TH-5000[System]", the source type label "[System]" is directly appended to the information fragment text, forming a complete labeled field value string. Finally, the labeled field values ​​of all fields constitute the field extraction result with fragment source labels.

[0078] Through the above technical solution, this embodiment implements fragment source labeling in the form of "information fragment + source type label" appended, so that the field value text itself carries source labeling information, and the output is a complete string that can be displayed. No additional labeling and field value concatenation logic is required when displaying it on the front end.

[0079] This embodiment determines the source attribution of an information fragment and appends a source type label to the fragment, enabling the field value text to carry the source label information itself. In existing technologies, if the source label is stored as an additional data structure independent of the field value, the field value and label need to be rendered separately and then combined during display, resulting in complex front-end display logic and potential misalignment of the label and field value during storage and transmission. Compared to existing technologies, this embodiment directly embeds the source type label into the field value text using an appended append method, outputting a complete, displayable string. This achieves a natural binding between the label information and the field value, simplifies the display logic, and avoids the misalignment problem caused by the separation of the label and field value.

[0080] The specific display format for related information has not yet been specified. If field values ​​and labels are displayed in different areas of the report, maintenance personnel need to visually align information fragments with source labels across different areas, which reduces reading efficiency.

[0081] Therefore, referring to Figure 9 This application also provides a method for labeling the source of a field, based on the above. Figure 1 The illustrated embodiment associates the field values ​​of each field with their corresponding source type labels for display, specifically in step S410, where: S410. Associate the field value of each field in the target report with the source type label of each information fragment in that field value and display them in the same display area.

[0082] The target report refers to the maintenance report that needs to display the field extraction results; the display area refers to a continuous display space allocated for a field in the maintenance report, such as a row or a cell range in the maintenance report.

[0083] Secondly, the control device renders the extracted field results with source tagging on the target maintenance report interface. Then, in the display cell of the fault cause field in the maintenance report, the entire string of "Continuous overcurrent at the output terminal caused thermal breakdown of the switching transistor [dialogue], and the decrease in cooling fan speed exacerbated the temperature rise [inference]" is displayed, with the entire labeled field value string within the same cell area. Finally, the field values ​​of each field and their source type tags for the information fragments are displayed in association within the same display area, eliminating the need for maintenance personnel to search for corresponding relationships across different areas.

[0084] Through the above technical solution, this embodiment converges the display space of field values ​​and source annotations into the same display area, and the layout of trailing annotations allows the correspondence between fragments and sources to be directly perceived visually without cross-region scanning.

[0085] This embodiment displays field values ​​and source type labels together within the same display area, making the correspondence between fragments and their sources visually direct. In existing technologies, when field values ​​and labels are laid out separately, maintenance personnel need to align information across different areas, resulting in low reading efficiency. Compared to existing technologies, this embodiment uses the same display area to carry both field values ​​and their source labels, shortening the visual correspondence distance and enabling direct perception of the fragment source relationship.

[0086] When a single field value is composed of multiple information fragments from different sources, simply labeling it with a general source type will lose the distinction between fragment-level sources, while displaying each fragment independently will disrupt the coherence and readability of the field value.

[0087] Therefore, referring to Figure 10 Another embodiment of this application provides a field source annotation method, based on the above. Figure 1 The illustrated embodiment associates the field values ​​of each field with their corresponding source type labels for display, specifically in step S420, where: S420. When the field value of the target field consists of multiple information fragments of different source types, the multiple information fragments are arranged alternately with their respective source type labels and displayed in association.

[0088] Alternating arrangement refers to a combination structure in which information fragments from different sources and their source type labels are alternately arranged within the display area of ​​the same field value. For example, the field value is displayed as "Continuous overcurrent at the output end caused thermal breakdown of the switching tube [dialogue], and the decrease in the speed of the cooling fan aggravated the temperature rise [inference]". The whole is presented as a coherent text, but the source labels naturally separate the different fragments.

[0089] Secondly, when the control device renders the field extraction results with fragment source labels to the maintenance report interface, for field values ​​containing multi-source information fragments, the fragments are arranged sequentially according to their appearance order in the field value text. Then, after each information fragment is displayed, its source type label is shown, achieving a natural alternation between fragments and labels. Finally, multi-source field values ​​are presented in a complete structure with alternating "fragment + label" arrangements. For example, the fault cause field value "Continuous overcurrent at the output terminal caused thermal breakdown of the switching transistor [dialogue], and the decrease in cooling fan speed exacerbated the temperature rise [inference]" is presented coherently.

[0090] Through the above technical solution, this embodiment balances the fragment-level source differentiation and the overall readability of field values ​​in the display of multi-source field values, so that the source annotation information is integrated into the field value text flow without destroying the coherent semantics.

[0091] This embodiment achieves a balance between fragment-level source differentiation and overall readability of field values ​​by alternating and associating multi-source information fragments with their source type labels. In existing technologies, multi-source field values ​​either only label the total source type, losing fragment differentiation, or display each fragment independently, disrupting coherence. Compared to existing technologies, this embodiment displays multi-source field values ​​using a combined structure of alternating fragments and labels, achieving a unity of source differentiation and semantic coherence.

[0092] If the large language model fails to find any information fragment matching a certain field in the multi-source data, the extraction result of the large language model for that field will be null. In this case, it is necessary to clearly communicate to the maintenance personnel that the field has no information source in the current multi-source data.

[0093] Therefore, referring to Figure 11 Another embodiment of this application provides a method for labeling the source of a field, based on the above. Figure 1 The illustrated embodiment triggers the large language model to extract field values ​​from multi-source data, specifically in step S370, where: S370. If the large language model does not extract any information fragment that matches the target field, output an empty field value for the target field and attach a source missing marker.

[0094] Among them, the source missing marker is a special marker that is different from the four normal source type markers. It is used to indicate that the information source of the field was not found in the multi-source data, for example, it is displayed as "[No source]".

[0095] Secondly, when extracting the maintenance time estimation field, the large language model traversed the dialogue text in the "[Dialogue]" marked area and found no mention of maintenance time by the maintenance personnel; it traversed the optical character recognition (OCR) text in the "[Image]" marked area and found no numerical or time-related expressions related to maintenance time; it traversed the manufacturing execution system returned data in the "[System]" marked area, but the maintenance time for different fault types in the equipment's historical maintenance records varied too much to be used as a reference. Subsequently, the large language model determined that this field had no information source in the current multi-source data. Finally, it output an empty field value for this field and added a "[No Source]" missing source marker to inform the maintenance personnel that this field needs to be manually filled in.

[0096] Through the above technical solution, this embodiment replaces the silent failure with a source missing marker when the information source of a field is completely missing, and transforms the fact of "no information found" into a status prompt visible to maintenance personnel, so that maintenance personnel know clearly which fields need to be manually supplemented.

[0097] This embodiment explicitly presents the "no information source" status by outputting an empty field value and attaching a source-missing flag when no matching information fragment is extracted. In existing technologies, when a large language model fails to extract a field value, it typically outputs an empty value or fabricated content. Maintenance personnel cannot distinguish between "empty field value" and "no information source for the field," and silent failure with an empty value easily leads to omissions or errors. Compared to existing technologies, this embodiment explicitly informs maintenance personnel of the "no information source found" status through a source-missing flag, replacing the silent failure with a blank value. This allows maintenance personnel to clearly know which fields need manual input, avoiding omissions caused by silent empty values.

[0098] When there are many fields in a maintenance report, maintenance personnel may want to focus on fields that come from a specific source type. In this case, displaying all source types together will reduce the efficiency of focused reading.

[0099] Therefore, referring to Figure 12 In another embodiment of this application, a method for annotating the source of a field is provided, based on the above. Figure 1 The illustrated embodiment associates the field values ​​of each field with their corresponding source type labels for display, specifically in steps S430-S450, where: S430, Obtain source filtering instructions carrying the target source type; S440. Fields whose values ​​contain source type annotations corresponding to the target source type are identified as fields to be displayed. S450. Associate the field value of the field to be displayed with the corresponding source type label.

[0100] Among them, the source filtering instruction refers to the filtering operation triggered by maintenance personnel on the maintenance report display interface, which specifies a certain source type as the target source type. For example, clicking the "[Inference]" filter button at the top of the interface intends to only view fields and fragments marked with "[Inference]" in order to focus on reviewing the reasoning content of the large language model.

[0101] Secondly, when maintenance personnel review the maintenance report, they click the "[Inference]" button in the interface filter bar. The control device receives a source filtering instruction carrying the target source type, "Model Inference Type." Subsequently, the control device scans all seven fields in the report. The "Fault Cause" and "Required Spare Parts" fields contain the "[Inference]" label, while the remaining five fields do not. The "Fault Cause" and "Required Spare Parts" fields are selected as the fields to be displayed. Finally, the report interface only displays the field values ​​of the fields to be displayed along with their corresponding source type labels, hiding the other five fields. This allows maintenance personnel to focus on reviewing the model's inference content.

[0102] Through the above technical solution, this embodiment introduces an interactive mechanism for filtering by source type on the basis of source labeling, enabling maintenance personnel to upgrade source labeling from passively displaying information to actively filtering view criteria, and focus on viewing the field content corresponding to a specific source type according to the data credibility requirements.

[0103] This embodiment upgrades source labeling to a view filtering criterion by receiving source filtering instructions and displaying only fields containing the target source type annotation. In existing technologies, source annotations (if any) are merely passively displayed information; maintenance personnel cannot focus on filtering report fields based on source type. When there are many fields, manual identification of source annotations is required for each field, resulting in low efficiency in focused reading. Compared to existing technologies, this embodiment drives the display of only fields with the target source type annotation through source filtering instructions, enabling maintenance personnel to focus on viewing field content of specific source types according to data reliability requirements. This upgrades the function of source annotation from simply displaying information to providing an interactive filtering criterion.

[0104] When maintenance personnel have doubts about the source label of a certain information segment, the source label only indicates "from the conversation" and cannot help maintenance personnel quickly locate the corresponding specific location in the original text of the conversation for verification.

[0105] Therefore, referring to Figure 13 This application also provides a method for labeling the source of a field, based on the above. Figure 1 The illustrated embodiment associates the field values ​​of each field with their corresponding source type labels for display, specifically in steps S460-S480, where: S460, Obtain the source tracing operation instructions to be performed on the target source type annotation; S470. Locate the original data location of the information fragment corresponding to the target source type label in multi-source data; S480, Jump the display context to the original data location, where the display context includes the current view and the display location.

[0106] Among them, the traceability operation instruction refers to the interactive operation performed by maintenance personnel on the maintenance report interface to mark the source type of a certain information fragment, such as clicking or long-pressing the "[Dialogue]" mark after the information fragment; the original data location refers to the position of the specific data segment that semantically matches the information fragment in the original data file, such as lines 28 to 30 in the maintenance dialogue text.

[0107] Secondly, when the maintenance personnel were reviewing the fault cause field value, they clicked the "[Dialogue]" marker following "Continuous overcurrent at the output terminal caused thermal breakdown of the switching transistor," and the control device received a tracing operation command. Subsequently, the control device analyzed the source type corresponding to this marker as a dialogue extraction type, searched the fault maintenance dialogue text for a dialogue segment that semantically matched this information fragment, and located line 28 of the dialogue text: "Maintenance personnel: The switching transistor at position P2 was found to be broken down; the output terminal was previously in a continuous overcurrent state." Finally, the control device switched the display context of the maintenance report from the report view to the dialogue text view and automatically scrolled to the position corresponding to line 28.

[0108] Through the above technical solution, this embodiment upgrades the source labeling from passive tag information to an interactive traceability entry point. Maintenance personnel can click on the source labeling to directly jump to the corresponding original data location in the multi-source data for verification, so that the source labeling not only tells "where it comes from" but also provides an interactive channel to "go there".

[0109] This embodiment upgrades source labeling into an interactive source tracing entry point by receiving a source tracing operation command and jumping to the original data location. In existing technologies, source labeling (if any) only indicates the source of the information fragment; maintenance personnel need to manually search for the corresponding content in the original data, resulting in low verification efficiency and a high risk of mislocation. Compared to existing technologies, this embodiment uses a source tracing operation command to drive the location of the information fragment in the original data of multi-source data and jump to its display. This allows maintenance personnel to directly access the corresponding location in the original data for verification by clicking on the source label, realizing a functional upgrade of source labeling from a passive tag to an active source tracing channel, and enhancing the verifiability of the credibility of information fragments from different sources.

[0110] There is no specific definition of how to locate the original data in multi-source data. If all data is traversed and compared segment by segment, the retrieval efficiency is low and the location is inaccurate. At the same time, the presentation of the view after the jump is not clear.

[0111] Therefore, referring to Figure 14 Another embodiment of this application provides a field source annotation method, based on the above. Figure 13 The embodiment shown locates the original data position of the information fragment corresponding to the target source type label in multi-source data, specifically in steps S471-S472, wherein: S471. Based on the source type of the target source type, retrieve the data segment that semantically matches the information fragment in the multi-source data; S472. Determine the start and end positions of the information segments within the data paragraphs; The display context will be redirected to the original data location, specifically in step S481, where: S481. Switch the display context from the report view to the data view of the data source to which the information fragment belongs, and position it at the start and end points.

[0112] Semantic matching refers to retrieving data segments that are semantically consistent with the information fragments from the data sources corresponding to the source type by using vector similarity or keyword weighted matching; the start and end positions are determined by character offsets or word segmentation boundaries.

[0113] Secondly, the control device parses the source type of the traceability operation command as a dialogue extraction type, limiting the search to the fault maintenance dialogue text to avoid traversing all data sources. Then, it searches the dialogue text for a dialogue segment semantically matching the information fragment "Continuous overcurrent at the output terminal caused thermal breakdown of the switching transistor," locating line 28 of the dialogue text: "Maintenance personnel: The switching transistor at position P2 was found to be broken down; the output terminal was previously in a continuous overcurrent state." The start and end positions corresponding to this information fragment are determined to be lines 28 to 30. Finally, the display context is switched from report view to dialogue text view, and the device automatically scrolls to the positions corresponding to lines 28 to 30.

[0114] Through the above technical solution, this embodiment improves the accuracy of original data positioning and retrieval efficiency by limiting the search range based on source type and locating the start and end positions by semantic matching. At the same time, it achieves seamless connection of context after jump through view switching.

[0115] This embodiment achieves precise and efficient source tracing by limiting the search scope based on source type, locating the start and end positions through semantic matching, and combining view switching. In existing technologies, locating the original position of information fragments in multi-source data typically requires traversing all data sources and comparing each segment individually. This results in a large search scope, low efficiency, and a lack of precise start and end positions in the location results. Users still need to manually scroll through the search after switching. Compared to existing technologies, this embodiment first limits the search scope to a single corresponding data source based on the source type, then uses semantic matching to determine the data segment and its start and end positions, and finally automatically locates the start and end positions by switching from the report view to the data view. This achieves a complete source tracing chain with narrowed search scope, precise location, and automatic context connection after switching.

[0116] This application also proposes a control device, which includes: a memory, a processor, and a field source annotation program stored in the memory and executable on the processor, the field source annotation program being configured to implement the field source annotation method described above.

[0117] It is worth noting that since the control device of this application is based on the above-mentioned field source labeling method, the embodiments of the control device of this application include all the technical solutions of all embodiments of the above-mentioned field source labeling method, and the technical effects achieved are exactly the same, so they will not be repeated here.

[0118] This application also proposes a cloud server, which includes the control device as described in the above embodiments.

[0119] It is worth noting that since the cloud server of this application is based on the control device described above, the embodiments of the cloud server of this application include all the technical solutions of all the embodiments of the control device described above, and the technical effects achieved are exactly the same, so they will not be repeated here.

[0120] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for labeling the source of a field, characterized in that, The method for annotating the source of the field includes: Retrieve structured field templates and multi-source data; Source labeling prompts are constructed based on the structured field template and the multi-source data, wherein the multi-source data is marked and distinguished by source identifiers in the source labeling prompts; The source-labeled prompt words are input into the large language model, which triggers the large language model to extract the field values ​​of each field from the multi-source data, and labels the source type of each information fragment in the field value, generating field extraction results with fragment source labels; Output the field extraction results with fragment source labels, so that the field values ​​of each field are associated with the corresponding source type labels and displayed.

2. The field source annotation method as described in claim 1, characterized in that, The specific steps for obtaining structured field templates and multi-source data are as follows: Obtain the text of the troubleshooting dialogue; Perform optical character recognition on the fault image to generate optical character recognition text; Retrieve query data returned by the Manufacturing Execution System; The fault repair dialogue text, the optical character recognition text, and the query data are used as the multi-source data.

3. The field source annotation method as described in claim 2, characterized in that, The multi-source data is distinguished by source identifiers, specifically: Configure a dialogue identifier for the fault repair dialogue text, configure an image identifier for the optical character recognition text, and configure a system data identifier for the query data.

4. The field source annotation method as described in claim 2, characterized in that, The source type of each information fragment in the labeled field value is specifically as follows: When the information fragment originates from the fault repair dialogue text, the source type of the information fragment is marked as the dialogue extraction type; When the information fragment originates from the optical character recognition text, the source type of the information fragment is labeled as image recognition type; When the information fragment originates from the queried data, the source type of the information fragment is labeled as the system data extraction type; If the information fragment does not originate directly from the multi-source data, the source type of the information fragment is labeled as the model inference type.

5. The field source annotation method as described in claim 3, characterized in that, The steps of configuring dialogue identifiers for the fault repair dialogue text, configuring image identifiers for the optical character recognition text, and configuring system data identifiers for the query data include: An open tag is configured at the beginning of the corresponding data text, and a closed tag is configured at the end of the corresponding data text.

6. The field source annotation method as described in claim 1, characterized in that, The specific steps for obtaining the structured field template are as follows: Retrieve multiple field definitions, each of which includes a field identifier and a semantic description of the field. The structured field template is generated based on the multiple field definitions.

7. The field source annotation method as described in claim 1, characterized in that, The construction of source annotation prompts based on the structured field template and the multi-source data specifically involves: The role setting section, field definition list section, multi-source data section, and output format constraint section are written sequentially in the source label prompt text; The field definition list segment is generated based on the structured field template.

8. The field source annotation method as described in claim 1, characterized in that, The source type of each information fragment in the labeled field value is specifically as follows: Determine the source attribution of the information fragment within the multi-source data; A source type label corresponding to the source attribution is appended after the information fragment.

9. The field source annotation method as described in claim 1, characterized in that, The method of associating the field values ​​of each field with their corresponding source type labels for display is as follows: The field value of each field in the target report is associated with the source type of each information fragment in that field value and displayed in the same display area.

10. The field source annotation method as described in claim 1, characterized in that, The method of associating the field values ​​of each field with their corresponding source type labels for display is as follows: When the field value of the target field consists of multiple information fragments from different source types, the multiple information fragments are arranged alternately with their respective source type labels and displayed in association.

11. The field source annotation method as described in claim 1, characterized in that, The specific steps to trigger the large language model to extract field values ​​from the multi-source data are as follows: If the large language model does not extract any information fragment that matches the target field, an empty field value is output for the target field, and a source missing marker is attached.

12. The field source annotation method as described in claim 1, characterized in that, The method of associating the field values ​​of each field with their corresponding source type labels for display is as follows: Retrieve source filtering instructions that carry the target source type; Fields whose values ​​contain source type annotations corresponding to the target source type are identified as fields to be displayed; The field value of the field to be displayed is associated with the corresponding source type label and displayed.

13. The field source annotation method as described in claim 1, characterized in that, The method of associating the field values ​​of each field with their corresponding source type labels for display is as follows: Obtain the source tracing operation instructions to be executed on the target source type annotation; Locate the original data position of the information fragment corresponding to the target source type label in the multi-source data; The display context will jump to the original data location, where the display context includes the current view and the display location.

14. The field source annotation method as described in claim 13, characterized in that, The step of locating the original data location of the information fragment corresponding to the target source type label in the multi-source data specifically involves: Based on the source type labeled with the target source type, retrieve data segments in the multi-source data that semantically match the information fragment; Determine the start and end positions of the information segment within the data segment; The step of jumping the display context to the original data location specifically involves: The display context will be switched from the report view to the data view of the data source to which the information fragment belongs, and the display will be positioned at the start and end points.

15. A control device, characterized in that, The control device includes: a memory, a processor, and a field source annotation program stored in the memory and executable on the processor, the field source annotation program being configured to implement the field source annotation method as described in any one of claims 1 to 14.

16. A cloud server, characterized in that, Includes the control device as described in claim 15.

Images