Enterprise carbon inventory and auditable evidence generation method and system based on multi-modal ai

Abstract

This invention discloses a method and system for enterprise carbon inventory and auditable evidence generation based on multimodal AI, relating to the fields of carbon emission accounting and artificial intelligence technology. The invention acquires multimodal evidence data from enterprise carbon inventory, uses a multimodal AI model to analyze and extract relevant carbon inventory information, generates candidate emission activities and completes intelligent attribution of multi-source, multimodal evidence, constructs an evidence conflict graph to identify six types of conflicts, including numerical and temporal conflicts, and performs comprehensive credibility assessment and conflict resolution based on six dimensions such as source reliability and equipment operational capability constraints, matches emission factors to calculate carbon emissions, and generates an auditable evidence package containing full-chain traceability information. This invention significantly reduces the workload of manual review, significantly improves the accuracy and credibility of accounting, achieves full-process auditability of carbon inventory, and meets regulatory and third-party verification requirements.

Description

Technical Field

[0001] This invention relates to the fields of carbon emission accounting and artificial intelligence technology, specifically to a method and system for enterprise carbon inventory and auditable evidence generation based on multimodal AI. Background Technology

[0002] Enterprise carbon audits require the identification, collection, and accounting of energy consumption, material usage, and related emissions activities involved in production and operations. Data sources typically include invoices, contracts, ledgers, energy bills, equipment photos, video inspection data, IoT time-series data, and records from enterprise business systems. Due to significant differences in format, source, semantic expression, and time granularity among these data sources, enterprise carbon audit data processing exhibits characteristics of being multi-source, heterogeneous, and multimodal. Existing technologies often rely on manual data entry, structured form collection, or rule extraction from a single data source. This makes it difficult to uniformly analyze different modalities of evidence, such as images, text, tables, scanned documents, and sensor data, leading to problems such as incomplete field extraction, inconsistent semantic understanding, and errors in identifying units, boundaries, and facility affiliations.

[0003] Furthermore, the same emission activity often corresponds to candidate evidence from multiple sources, and these different pieces of evidence may conflict in terms of numerical values, time, units, attribution, boundaries, and the applicable conditions of emission factors. Existing technologies typically lack a unified mechanism for associating, expressing, and systematically resolving conflicts among multi-source candidate evidence, often relying on manual judgment or simple priority rules, leading to insufficient accuracy in determining activity data and high costs for manual review. Simultaneously, existing solutions lack complete records of original evidence, analytical results, conflict resolution basis, emission factor versions, and calculation paths during the carbon inventory process, making it difficult to establish a complete traceability chain from original evidence to accounting results. Therefore, it is difficult to meet the credibility and traceability requirements of audit review and regulatory verification. Thus, it is necessary to propose a multimodal AI-based method and system for enterprise carbon inventory and the generation of auditable evidence to address the aforementioned technical problems. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for enterprise carbon inventory and auditable evidence generation based on multimodal AI, so as to solve the technical problems in the prior art, such as difficulty in multimodal evidence parsing, insufficient attribution and conflict resolution of multi-source evidence, and poor traceability and auditability of the accounting process.

[0005] In a first aspect, the present invention provides a method for enterprise carbon inventory and auditable evidence generation based on multimodal AI, including:

[0006] Obtain multimodal evidence data for corporate carbon inventory;

[0007] The multimodal evidence data is analyzed using a multimodal AI model to extract information related to carbon inventory, including activity data, emission sources, units of measurement, time intervals, facility affiliation, organizational boundaries, and applicable conditions for emission factors.

[0008] Based on the analysis results, candidate emission activities are generated, and evidence source identification, data confidence level, and activity attribute information are established for each candidate emission activity.

[0009] Establish the correlation between evidence and emission activity candidates, and attribute evidence from different sources and modalities to the corresponding emission activity candidates;

[0010] For multiple candidate evidences corresponding to the same emission activity candidate, construct an evidence conflict map and identify numerical conflicts, temporal conflicts, unit conflicts, attribution conflicts, boundary conflicts and emission factor applicability condition conflicts.

[0011] Based on source reliability, resolution confidence, spatiotemporal consistency, equipment operating capability constraints, boundary matching degree, and evidence tampering risk, a comprehensive credibility assessment is performed on the multi-source candidate evidence, and conflict resolution is performed on the evidence conflict graph according to the comprehensive credibility to determine the target activity data;

[0012] Calculate the company's carbon emissions based on the target activity data and the matched emission factors;

[0013] Generate an auditable evidence package, which includes an original evidence index, parsing results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results.

[0014] Secondly, a multimodal AI-based enterprise carbon inventory and auditable evidence generation system is also provided, including:

[0015] The acquisition module is used to acquire multimodal evidence data for enterprise carbon inventory checks;

[0016] The analysis module is used to analyze the multimodal evidence data using a multimodal AI model and extract information related to the carbon inventory, including activity data, emission sources, units of measurement, time intervals, facility affiliation, organizational boundaries, and applicable conditions for emission factors.

[0017] The candidate object generation module is used to generate emission activity candidate objects based on the analysis results, and to establish evidence source identifiers, data confidence levels and activity attribute information for each emission activity candidate object;

[0018] The association module is used to build the association between evidence and emission activity candidates, attributing evidence from different sources and modalities to the corresponding emission activity candidates.

[0019] The module is used to construct evidence conflict diagrams for multiple source candidate evidence corresponding to the same emission activity candidate, and to identify numerical conflicts, temporal conflicts, unit conflicts, attribution conflicts, boundary conflicts, and emission factor applicability condition conflicts.

[0020] The conflict resolution module is used to comprehensively evaluate the credibility of the multi-source candidate evidence based on source reliability, resolution confidence, spatiotemporal consistency, equipment operating capability constraints, boundary matching degree and evidence tampering risk, and to perform conflict resolution on the evidence conflict graph based on the comprehensive credibility to determine the target activity data.

[0021] The calculation module is used to calculate the enterprise's carbon emissions based on the target activity data and the matched emission factors;

[0022] The evidence package generation module is used to generate an auditable evidence package, which includes the original evidence index, parsing results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results.

[0023] Thirdly, the present invention also provides an electronic device, comprising: a memory for storing computer software programs; and a processor for reading and executing the computer software programs, thereby realizing the enterprise carbon inventory and auditable evidence generation method based on multimodal AI as described above.

[0024] Fourthly, the present invention also provides a non-transitory computer-readable storage medium storing a computer software program, which, when executed by a processor, implements the enterprise carbon inventory and auditable evidence generation method based on multimodal AI as described above.

[0025] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the enterprise carbon inventory and auditable evidence generation method based on multimodal AI as described above.

[0026] Compared with existing technologies, this invention improves the accuracy of extracting relevant fields and the semantic standardization capability of carbon inventory by uniformly accessing and multimodal AI parsing text, tables, images and other heterogeneous evidence in corporate carbon inventory.

[0027] By generating candidate emission activities and establishing the correlation between evidence and candidate activities, evidence data scattered across different sources and modalities can be attributed to the corresponding emission activities.

[0028] By constructing an evidence conflict diagram, the conflict, support, derivation, and temporal inheritance relationships among multi-source candidate evidence are expressed in a structured manner. Combined with the reliability of the source, the confidence level of the resolution, the spatiotemporal consistency, the constraints of equipment operation capability, the boundary matching degree, and the risk of evidence tampering, a comprehensive credibility assessment and conflict resolution can be carried out, which can improve the accuracy and stability of the determination of target activity data and reduce the pressure of manual review.

[0029] Meanwhile, by generating an auditable evidence package that includes an original evidence index, parsing results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results, and by performing versioning, timestamp recording, digest hash verification, and signature processing on the relevant content, a complete traceability chain from original evidence to final carbon emission results can be formed, thereby improving the credibility, traceability, and auditability of corporate carbon inventory results. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the enterprise carbon inventory and auditable evidence generation method based on multimodal AI provided in this embodiment of the invention.

[0031] Figure 2 This is a schematic diagram of the structure of the enterprise carbon inventory and auditable evidence generation system based on multimodal AI provided in an embodiment of the present invention;

[0032] Figure 3 An embodiment diagram of the electronic device provided in this invention;

[0033] Figure 4 An embodiment diagram of a computer-readable medium provided for an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] In the description of this invention, the terms "first" and "second" 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. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0036] In the description of this invention, the term "such as" is used to mean "used as an example, illustration, or description." Any embodiment described "such as" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to implement and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be implemented without using these specific details. In other instances, well-known structures and processes are not described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0037] See Figure 1 , Figure 1 This is a flowchart illustrating the enterprise carbon inventory and auditable evidence generation method based on multimodal AI provided by the present invention. In this embodiment of the invention, the executing entity of the enterprise carbon inventory and auditable evidence generation method based on multimodal AI is the enterprise carbon inventory and auditable evidence generation system based on multimodal AI. Therefore, the enterprise carbon inventory and auditable evidence generation method based on multimodal AI includes...

[0038] Step 10: Obtain multimodal evidence data for enterprise carbon inventory.

[0039] In step 10, multimodal evidence data related to the enterprise's carbon inventory is acquired. When conducting a carbon inventory, the basic data involved is typically scattered across multiple business processes and management systems, and the corresponding forms of evidence are not uniform. These may include documents or forms such as contracts, invoices, reports, and ledgers, as well as equipment photos, scanned copies, on-site recorded images, video screenshots, IoT data collection records, and data exported from business systems. Therefore, in this step, various types of original evidence are uniformly received and organized around the carbon inventory task to form the original input set required for subsequent processing.

[0040] Optionally, for evidence data from different sources and with significant structural differences, a unified data access method can be used for integration, so that data originally scattered across different media, systems, and expressions can enter the same processing chain.

[0041] Furthermore, while acquiring evidence data, basic information related to the evidence, such as source identification, collection method, access time, original storage location, and uploading entity, can be recorded to facilitate subsequent evidence tracing, source analysis, and evidence encapsulation.

[0042] Step 10 provides the initial evidentiary basis for subsequent multimodal analysis and activity attribution. See steps 101-104 for details.

[0043] Step 20: Use a multimodal AI model to analyze the multimodal evidence data and extract information related to the carbon inventory.

[0044] In step 20, a multimodal AI model is used to parse the multimodal evidence data obtained in step 10 to extract information related to the carbon inventory. Because corporate carbon inventory evidence varies in its content organization, structuring, and semantic expression, the same field may appear differently in different pieces of evidence. For example, the same activity data may appear as text descriptions, table items, image content, or system records. Therefore, it is necessary to use parsing methods adapted to different modal characteristics to identify, extract, and organize the effective information from various types of evidence.

[0045] In this step, the information extracted may include carbon inventory-related content such as activity data, emission sources, units of measurement, time intervals, facility affiliation, organizational boundaries, and applicable conditions for emission factors.

[0046] Optionally, for evidence of different layout types or different modalities, corresponding parsing pipelines can be used for processing, and the extraction results can be semantically standardized to form a unified representation of fields with similar meanings but inconsistent expressions.

[0047] Furthermore, by structuring the analysis results, the original evidence can be transformed into data objects that can be used for activity identification and relationship establishment, thereby improving the data usability of subsequent steps. See steps 201-204 for details.

[0048] Step 30: Generate candidate emission activities based on the analysis results, and establish evidence source identification, data confidence level and activity attribute information for each candidate emission activity.

[0049] In step 30, emission activity candidate objects are generated based on the analysis results obtained in step 20. Since the various fields extracted from multimodal evidence are typically discretely distributed, and different pieces of evidence may provide partial information about the same emission activity, these analysis results need to be aggregated according to predetermined activity identification rules to form emission activity candidate objects that can be used for subsequent carbon accounting processing. The emission activity candidate object is a structured representation unit established for potential emission activities, used to carry activity data, time information, facility information, and organizational affiliation information related to the activity.

[0050] Optionally, in the process of generating candidate emission activities, analytical results from different evidence but pointing to the same activity can be merged and organized, and evidence source identifiers, data confidence levels and activity attribute information can be established for the corresponding candidate activities.

[0051] Furthermore, the activity attribute information can be used to characterize the activity type, scope, and accounting conditions of the candidate object, and the data confidence level can provide a reference for subsequent conflict identification and credibility assessment.

[0052] Step 30 transforms the fragmented parsed fields into data objects organized around emission activities, providing a foundation for subsequent evidence attribution and conflict resolution. See steps 301-303 for details.

[0053] Step 40: Construct the association between evidence and emission activity candidates, attributing evidence from different sources and modalities to the corresponding emission activity candidates.

[0054] In step 40, a correlation is established between evidence and candidate emission activities, attributing evidence from different sources and modalities to the corresponding candidate emission activities. Since the same emission activity in a corporate carbon inventory is often not fully described by a single piece of evidence, but requires the combined identification and verification of documents, images, tables, or system records from multiple sources, it is necessary to establish a mapping relationship between evidence and candidate emission activities to determine the scope and degree of support that the evidence provides for the candidates.

[0055] Optionally, when constructing the association, the correspondence between the evidence and the candidate objects in terms of semantic content, time range, facility objects, units of measurement, and business process relationships can be comprehensively considered to improve the rationality of the attribution results.

[0056] Furthermore, by establishing correlations between evidence and candidate emission activities, information scattered across different pieces of evidence can be organized around a specific emission activity, enabling subsequent conflict identification to focus on the same activity unit. Step 40 provides processing boundaries for subsequent evidence comparison, relationship analysis, and conflict resolution within the same activity. This is specifically described in steps 401-402.

[0057] Step 50: Construct an evidence conflict map for multi-source candidate evidence corresponding to the same emission activity candidate, and identify numerical conflicts, time conflicts, unit conflicts, attribution conflicts, boundary conflicts, and emission factor applicability condition conflicts.

[0058] In step 50, an evidence conflict diagram is constructed for multiple candidate evidences corresponding to the same emission activity candidate, and potential conflicts between different candidate evidences are identified. In corporate carbon inventory scenarios, even if multiple pieces of evidence all point to the same emission activity, the activity data, occurrence time, unit of measurement, attribution entity, boundary range, or applicable factor conditions they reflect may be inconsistent, leading to differences in accounting inputs. Therefore, it is necessary to organize multiple source evidences under the same candidate into a unified relational structure to analyze the supporting relationships, conflicting relationships, derivation relationships, or temporal inheritance relationships between the evidences.

[0059] Optionally, the evidence conflict diagram can be used to characterize the relationship between candidate evidence and candidate emission activities, as well as the relationship status between different candidate pieces of evidence.

[0060] Furthermore, after constructing the evidence conflict diagram, different types of conflicts can be classified and identified, such as numerical conflicts, temporal conflicts, unit conflicts, attribution conflicts, boundary conflicts, and conflicts of application conditions for emission factors, thereby transforming the problem of multiple pieces of evidence inconsistency into a structured conflict problem. Step 50 provides a relational and judgmental basis for subsequent comprehensive credibility assessment and conflict resolution decisions. This is specifically described in steps 501-505.

[0061] Step 60: Based on source reliability, resolution confidence, spatiotemporal consistency, equipment operating capability constraints, boundary matching degree and evidence tampering risk, a comprehensive credibility assessment is performed on multi-source candidate evidence, and conflict resolution is performed on the evidence conflict graph according to the comprehensive credibility to determine the target activity data.

[0062] In step 60, a comprehensive credibility assessment is performed on multiple candidate evidence sources under the same emission activity candidate, and conflict resolution is performed on the evidence conflict map based on the assessment results to determine the target activity data for carbon emission accounting. During corporate carbon inventory, the credibility of different evidence sources may vary. Some evidence sources are relatively stable, some evidence analysis results are more complete, and some evidence is more closely matched to facility status or time scenarios. Therefore, it is not appropriate to determine the final activity data based solely on a single field value or simple priority; rather, a comprehensive judgment of candidate evidence from multiple dimensions is necessary.

[0063] Optionally, in this step, factors such as source reliability, resolution confidence, spatiotemporal consistency, equipment operating capacity constraints, boundary matching degree, and evidence tampering risk can be comprehensively considered to conduct a credibility analysis on candidate evidence. Specifically, equipment operating capacity constraints can be determined based on equipment rated parameters, historical operating records, production plan data, or preset capacity thresholds; evidence tampering risk can be characterized based on the consistency of the original evidence content summary, metadata anomalies, version change records, timestamp verification results, or signature verification results.

[0064] Furthermore, corresponding processing strategies can be adopted for different types of conflicts to screen or determine data items more suitable as accounting basis within the conflicting relationships; for cases with insufficient credibility or difficult-to-resolve conflicts, supplementary evidence, manual review, or conservative accounting-related markers can also be generated. Through step 60, multi-source evidence with discrepancies under the same activity can be converged into target activity data that can be used for subsequent emission calculations. This is specifically described in steps 601-605.

[0065] Step 70: Calculate the company's carbon emissions based on the target activity data and the matching emission factors.

[0066] In step 70, the enterprise's carbon emissions are calculated based on the target activity data determined in step 60 and the matched emission factors. The target activity data is the accounting input formed after previous multimodal analysis, activity attribution, conflict identification, and credibility assessment. Its matching result with the emission factors jointly affects the final carbon emission calculation result. Therefore, in this step, it is necessary to match the corresponding emission factors based on the activity type, time range, regional attributes, energy type, supplier information, and factor version corresponding to the target activity data, and generate the carbon emission calculation result accordingly.

[0067] Optionally, while performing carbon emission calculations, the matching basis between target activity data and emission factors, as well as the corresponding calculation path, can be recorded so that when conducting subsequent reviews or audits, the source of activity data on which a certain calculation result depends and the applicable source of the factors can be clearly identified.

[0068] Furthermore, the calculation results generated in this step can not only serve as part of the enterprise's carbon inventory output, but also as an important component of subsequent evidence packaging and audit traceability. Step 70 transforms the results of the preceding data processing into the enterprise's carbon emission accounting output, as described in steps 701-702.

[0069] Step 80: Generate an auditable evidence package.

[0070] In step 80, an auditable evidence package is generated. This auditable evidence package is used to organize and encapsulate key data and processing results generated during the execution of the method of this invention. Its content may include original evidence indexes, parsing results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results. By uniformly organizing the above content, a structured evidence carrier oriented towards result review, audit inspection, and process traceability can be formed, enabling carbon inventory results to be traced back, interpreted, and verified.

[0071] Optionally, when generating an auditable evidence package, the relevant content can be structurally encapsulated according to predetermined rules, and version associations and audit traceability evidence dependencies can be established.

[0072] Furthermore, a link can be established between the evidence package and the original evidence, the analysis results, and the calculation results. A timestamp, digest hash, and signature information can be generated for the auditable evidence package to allow for gradual backtracking from the final accounting results to the corresponding original evidence during subsequent auditing, and to verify the sealed status of the evidence package. Step 80 enables enterprises to achieve better process interpretability and result traceability in their carbon inventory results. See steps 801-803 for details.

[0073] In this embodiment, by uniformly acquiring and processing multimodal evidence such as documents, tables, images, scanned copies, and system records generated during the enterprise carbon inventory process, it can better adapt to application scenarios where enterprise carbon inventory data sources are diverse and their forms are complex. By using a multimodal AI model for analysis and generating candidate emission activities and evidence associations based on the analysis results, evidence scattered across different sources and modalities can be organized around specific emission activities. By constructing an evidence conflict diagram and conducting comprehensive credibility assessment and conflict resolution for multi-source candidate evidence, it can provide technical support for determining target activity data. Furthermore, by performing emission factor matching and emission calculation on target activity data and generating an auditable evidence package, a processing link from original evidence to final accounting results can be formed, thereby improving the credibility, traceability, and audit review convenience of enterprise carbon inventory results.

[0074] In one embodiment, steps 101-104 are described as follows:

[0075] Step 101: Receive original evidence from multiple sources related to the enterprise's carbon inventory.

[0076] In step 101, multiple sources of original evidence related to the enterprise's carbon inventory are received. This original evidence can originate from multiple business processes involved in the enterprise's carbon inventory, including procurement, production, warehousing, transportation, energy management, equipment operation and maintenance, and financial settlement. Correspondingly, the original evidence can be in the form of text files, spreadsheets, invoice images, scanned copies, on-site photos, ledger pages, records exported from business systems, and sensor data collection records. Since the same emission activity may be recorded by multiple business processes simultaneously, this step is used to complete the receipt of evidence related to the carbon inventory task.

[0077] Optionally, the receiving method may include one or more of the following: manual upload, interface synchronization, batch import, or system periodic retrieval. For data received through different methods, a unified access task identifier can be assigned to facilitate subsequent tracking of the processing status of the same batch of evidence.

[0078] Furthermore, upon receipt, the integrity, reading status, and basic format availability of the evidence files can be verified. For evidence objects that are damaged, empty, or unreadable, an abnormal status can be marked, and they can be entered into a queue to be supplemented or repaired, so as to reduce their impact on subsequent structured processing.

[0079] Step 102: Perform format recognition and encapsulation on the original evidence.

[0080] In step 102, the original evidence received in step 101 undergoes format recognition and is encapsulated according to a unified data access standard. Format recognition distinguishes whether the original evidence is structured, semi-structured, or unstructured data, and further identifies its corresponding file type, page type, or record type, so that the matching parsing path can be invoked in step 20. For example, spreadsheets and database exported records can enter the table parsing or field mapping path, while scanned tickets, photos, and image pages can enter the visual recognition and layout parsing path.

[0081] Furthermore, during the access encapsulation process, information such as the original file content, access time, source channel, file type, business system identifier, and storage location can be packaged into a unified evidence object. This allows evidence from different modalities and sources to have a unified processing entry point in subsequent stages. For cases where the same evidence contains multiple pages, tables, or fragments, page-level or fragment-level sub-objects can be created during encapsulation to enable finer-grained information extraction, correlation, and conflict identification later. Through this step, the original evidence can be transformed into a uniformly schedulable data processing object.

[0082] Step 103: Extract and record the original evidence metadata.

[0083] In step 103, original evidence metadata is extracted and recorded from the original evidence ontology and its access process. This metadata may include source identifier, acquisition time, uploading entity, file type, access method, original storage location, business system identifier, and unique evidence identifier. The metadata describes the background of the evidence's formation and access, and can serve as the foundational data for subsequent source reliability analysis, evidence tracing, and evidence sealing.

[0084] Optionally, the unique evidence identifier can be generated by combining the original evidence content summary, source information, and access attributes, and is used to identify the uniqueness of the same evidence object in the system. The original evidence content summary can be used to generate evidence summary identifiers later, or to characterize the risk of evidence tampering during the credibility assessment stage. For evidence from different channels but with the same or similar content, aggregation management can be achieved in subsequent processing through unique identifiers or similar identifiers.

[0085] Furthermore, if the accessed object is fragmented evidence or multi-page evidence, page-level identifiers and fragment-level identifiers can be established under the unique identifier to support precise targeting during subsequent field location, fragment referencing, and audit backtracking.

[0086] Step 104: Establish the original evidence index set.

[0087] In step 104, an original evidence index set is established based on the evidence objects and their metadata obtained in steps 101 to 103. This original evidence index set records the mapping relationship between the evidence objects and their unique identifiers, source information, access information, page fragment information, and original storage locations, thereby establishing a referential basis for subsequent analysis results and the original evidence.

[0088] Furthermore, the original evidence index set can be organized using a hierarchical indexing approach, including evidence-level indexes, page-level indexes, and fragment-level indexes. Evidence-level indexes describe the entire original evidence object, page-level indexes identify specific pages or tables, and fragment-level indexes identify text blocks, table areas, image areas, or time-series fragments. This way, when subsequently parsing specific activity data, time fields, or unit fields, the index can directly point back to its corresponding original evidence location. Through this step, the subsequently generated parsing results, relationships, conflict relationships, and calculation paths can all be linked to the original evidence index set, thus providing underlying support for the generation of auditable evidence packages.

[0089] In one embodiment, steps 201-204 are described as follows:

[0090] Step 201: Identify the layout type and modality type of the evidence and determine the corresponding parsing pipeline.

[0091] In step 201, the evidence object acquired and encapsulated in step 10 is subjected to layout type identification and modality type identification to determine the corresponding parsing pipeline. The layout type identification result may include one or more of the following: plain text page, table page, mixed page, or scanned page. Different page layout types are suitable for different data extraction methods. For example, plain text pages may be suitable for text parsing pipelines, table pages may be suitable for table parsing pipelines, scanned pages may be suitable for image parsing pipelines, and mixed pages may be suitable for at least two combinations of text parsing pipelines, table parsing pipelines, and image parsing pipelines.

[0092] Furthermore, modality recognition can be used to distinguish whether evidence is presented as text, table, document, image, scanned page, or system record. Even among image or document evidence, different pages may have different layouts; for example, some may be primarily regular tables, while others may be primarily mixed text and document fields. Therefore, layout recognition can also be used to refine analysis strategies.

[0093] Optionally, the layout structure recognition results may include page area division information, field candidate area information, table area information, and text-image mixed layout relationship information. For evidence objects containing multiple pages of content, layout type recognition and modal type recognition can be performed page by page, and a page parsing task sequence can be generated, allowing subsequent steps to be processed by page or by area.

[0094] Step 202: Perform fusion analysis on the multimodal evidence and extract relevant fields for carbon inventory.

[0095] In step 202, the multimodal evidence identified in step 201 is fused and analyzed to extract carbon inventory-related fields. These carbon inventory-related fields may include activity data fields, emission source fields, material or energy category fields, time fields, unit of measurement fields, facility name fields, organization entity fields, location fields, and auxiliary fields related to the applicable conditions of emission factors. For text-based, tabular, and image-based evidence, corresponding features can be extracted separately, followed by fusion encoding and joint field extraction to reduce information fragmentation caused by synonymous and heterogeneous expressions.

[0096] In one embodiment, candidate regions can be generated based on page layout segmentation results, detection box results, table cell segmentation results, or page structure segmentation results. To improve the adaptability of field extraction for mixed text and image evidence, invoice pages, and complex forms, a multimodal fusion representation can be constructed and field candidate scores can be calculated. Specifically, the following fusion expression can be used:

[0097]

[0098] in, Indicates the first The fused feature vector of each candidate region or candidate field;

[0099] This represents the textual semantic features corresponding to the candidate region;

[0100] This indicates the visual features corresponding to the candidate region;

[0101] This indicates the layout structure features corresponding to the candidate region;

[0102] , , These represent mapping parameters for textual, visual, and layout features, respectively.

[0103] Indicates the bias term;

[0104] This represents a nonlinear mapping function.

[0105] Among them, text semantic features It can be derived from text recognition results and context encoding results, visual features It can be derived from image region coding results and page layout features. This can be derived from the layout structure recognition result in step 201. This fusion feature... This is used for subsequent field classification, entity recognition, and field attribution determination, thereby improving the adaptability of carbon inventory field extraction in complex evidence.

[0106] Furthermore, after obtaining the fusion features, the candidate fields can be classified and extracted by combining the carbon accounting terminology dictionary, domain labeling system, and field templates, and the field values, field types, field positions, and field confidence information can be output. This field confidence information can be used in subsequent steps 303 and 602.

[0107] Step 203: Perform semantic disambiguation, standardization, and unit normalization on the extracted fields.

[0108] In step 203, the fields extracted in step 202 are subjected to semantic disambiguation, standardization, and unit normalization. Since the expression methods in different pieces of evidence may differ—for example, the same activity may be represented by abbreviations, alternative names, business codes, or natural language descriptions, and the same measurement value may be represented by different units or different writing styles—it is necessary to perform unified mapping and standardization processing on the extracted results so that subsequent activity object generation can organize data with a relatively consistent standard.

[0109] Optionally, the semantic disambiguation and standardization process can be carried out based on the carbon accounting terminology database, the synonym mapping table, activity type mapping rules, and organization or facility name standardization table; the unit standardization process can be carried out based on preset unit conversion rules and unit legality rules.

[0110] Furthermore, for fields with contextual dependencies, the page containing the field, adjacent fields, and the type of source evidence can be used for auxiliary judgment. For example, when a numerical field itself does not explicitly indicate the meaning of the activity, the page title, table header, or adjacent fields can be consulted to determine whether it belongs to consumption, purchase, inventory, or equipment parameters. After standardization, semantic ambiguity in subsequent evidence merging and conflict identification can be reduced.

[0111] Step 204: Generate structured parsing results and establish the reference relationship between fields and original evidence.

[0112] In step 204, the various fields processed in steps 202 and 203 are organized into a structured parsing result. This structured parsing result may include field names, standard field values, original values, field types, unit information, time information, field confidence levels, page positions, and corresponding original evidence citation information. This structured organization allows subsequent steps to generate activity objects and determine relationships based on a unified data structure.

[0113] Furthermore, when generating structured parsing results, a referential relationship can be established between fields and the original evidence. This referential relationship includes at least the unique identifier of the original evidence, the page identifier, and the fragment location identifier, enabling each parsed field to be traced back to its source location. In this way, if subsequent issues arise such as conflicting evidence, credibility disputes, or audit review requirements, the original evidence content can be located based on the referential relationship.

[0114] In one embodiment, steps 301-303 are described as follows:

[0115] Step 301: Identify candidate emission activity units based on the structured analysis results.

[0116] In step 301, candidate emission activity units are identified based on the structured analysis results generated in step 204. Since the parsed fields are scattered as fragments of evidence, and carbon inventory calculations are usually carried out around specific emission activities, it is necessary to aggregate the sets of fields with potential similarities based on information such as activity type, time interval, facility object, organizational entity, and business attributes to form candidate emission activity units.

[0117] Optionally, when identifying candidate emission activity units, the activity type or activity keywords can be used as the initial clustering entry point, and then the selection can be carried out by combining constraints such as time, facilities, organization and location.

[0118] Furthermore, for records with incomplete field information but possessing some activity characteristics, candidate units to be confirmed can be formed first, and their attributes can be supplemented in subsequent association and conflict resolution. Through this step, the discrete fields obtained in the parsing stage can be initially organized into an activity data framework oriented towards carbon accounting.

[0119] Step 302: Merge the fields within the same candidate emission activity unit to generate emission activity candidate objects.

[0120] In step 302, the fields within the candidate emission activity units identified in step 301 are merged to generate emission activity candidate objects. The merging is based on semantic consistency, temporal proximity, facility consistency, and source correlation among fields within the same activity unit, combining and organizing multiple fields to form an objectified representation with clear activity boundaries.

[0121] Optionally, the emission activity candidate objects generated after merging may include fields such as activity type, candidate activity data value, candidate unit, target time interval, facility identifier, organization affiliation, organization boundary, spatial location, and auxiliary attributes.

[0122] Furthermore, in cases where multiple pieces of evidence provide different values ​​or descriptions of the same activity, these can be retained as parallel candidates in this step, rather than being decided upon directly at this stage. This preserves information for subsequent conflict graph construction and credibility assessment. This step elevates disparate fields to candidate emission activities with business and accounting significance.

[0123] Step 303: Establish evidence source identification, data confidence level, and activity attribute information for candidate emission activities.

[0124] In step 303, evidence source identifiers, data confidence levels, and activity attribute information are established for the emission activity candidate objects generated in step 302. The evidence source identifier records which original evidence, pages, or fragments support the candidate object; the data confidence level describes the reliability of each field within the candidate object during the parsing phase; and the activity attribute information characterizes the activity category, applicable boundaries, facility affiliation, and accounting auxiliary conditions of the candidate object.

[0125] Optionally, the activity attribute information may include activity type, activity data value, unit of measurement, time interval, facility identifier, organization affiliation, spatial location, data source, candidate emission factor conditions, and evidence summary identifier. The data source characterizes the business source category or source channel corresponding to the activity attribute, and the evidence source identifier characterizes the original evidence citation relationship supporting the candidate object. The evidence summary identifier can be generated from the original evidence content summary, unique evidence identifier, or page fragment identifier in step 103, and is used for integrity verification and citation of the corresponding evidence during subsequent credibility assessment and auditable evidence package generation.

[0126] Optionally, data confidence can be obtained by aggregating the confidence scores of fields within an object, retaining both field-level and object-level levels. Field-level confidence primarily reflects the reliability of the extraction results from step 202, while object-level confidence reflects whether the overall structure of the object is complete and whether the fields are consistent.

[0127] Furthermore, the activity attribute information may also include indicators such as whether the object needs to match specific emission factor conditions, whether it belongs to a specific organization's boundary, and whether it relies on subsequent supplementary evidence. Through this step, subsequent steps 402, 503, and 602 can perform evidence association, conflict identification, and credibility assessment around a unified object.

[0128] In one embodiment, steps 401-402 are described as follows:

[0129] Step 401: Generate candidate associations between evidence and candidate emission activities based on multidimensional matching factors.

[0130] In step 401, candidate associations between structured evidence fragments and emission activity candidates are generated based on semantic similarity, temporal overlap, facility matching degree, unit of measurement matching degree, and business process relationship. Since the same piece of evidence may contain information on multiple activities, and the same candidate may require multiple pieces of evidence to support it, a multi-dimensional matching method is needed for mapping and judgment. The candidate associations can be used to characterize the degree of correspondence between evidence fragments and candidate objects in terms of semantic expression, time interval, facility affiliation, unit of measurement, and business process location.

[0131] In one embodiment, the strength of the association between the evidence and the candidate emission activity can be calculated as follows:

[0132]

[0133] in, Indicates the first Fragments of evidence With the Candidates for emission activities The strength of the correlation between them;

[0134] Indicates semantic similarity;

[0135] Indicates the degree of time overlap;

[0136] Indicates the degree of facility matching;

[0137] Indicates the degree of matching of units of measurement;

[0138] Indicates the degree of matching between business process relationships;

[0139] to Each represents a weight.

[0140] Among them, evidence fragments The data comes from the structured analysis results generated in step 204, and the candidate emission activities. This is derived from the object generated in step 302. The business process relationship matching degree can be determined based on the preset correspondence between the business stage where the evidence is located and the type of emission activity, such as the sequential or supporting relationship of procurement evidence, warehousing evidence, production consumption records, and settlement documents in the business process. Association strength. This is used to establish the attribution relationship between evidence and candidate objects in step 402, and serves as the basis for constructing supporting edges or attribution edges in the subsequent conflict graph. Optionally, the candidate associations can also be further modified by incorporating organizational consistency.

[0141] Step 402: Filter the candidate associations and generate an evidence-emission activity association index.

[0142] In step 402, the candidate associations are screened based on the candidate associations and their association strengths obtained in step 401, and an evidence-emission activity association index is generated.

[0143] Optionally, when the correlation strength between a certain piece of evidence and a certain candidate object reaches a preset condition, the piece of evidence is attributed to the corresponding candidate object; when the same piece of evidence has a high degree of correlation with multiple candidate objects, the piece of evidence can be marked as a multi-attribution candidate evidence, and its multi-association state can be retained in the subsequent conflict diagram.

[0144] Furthermore, when generating the evidence-emission activity association index, the attribution basis can be recorded, including the main triggering matches, association strength, corresponding object attributes, and original evidence reference information. Evidence with low attribution strength but still supporting significance can be recorded as weak supporting evidence for subsequent credibility assessment as an auxiliary reference. This step enables the formation of an evidence set organized around candidate emission activities, providing a data foundation for subsequent multi-source evidence conflict analysis under the same candidate activity.

[0145] In one embodiment, steps 501-505 are described as follows:

[0146] Step 501: Aggregate the corresponding multi-source candidate evidence for the same emission activity candidate.

[0147] In step 501, for each candidate emission activity, the multi-source candidate evidence attributed in step 402 is aggregated to form an object-level evidence set. This object-level evidence set may include candidate data items from multiple original pieces of evidence, multiple pages, multiple field fragments, or multiple system records. Since these candidate data items may support or conflict with each other, this step first aggregates evidence surrounding the same activity object, providing input for the subsequent construction of an evidence conflict graph.

[0148] Optionally, during aggregation, primary evidence, secondary evidence, weak supporting evidence, and multi-attribution evidence can be distinguished. Primary evidence refers to evidence with a high degree of correlation with the candidate object and good field completeness; secondary evidence is used to supplement information such as time, unit, boundary, or facility; multi-attribution evidence is related to multiple objects simultaneously and its scope needs to be further confirmed in subsequent relationship judgments. Through this step, candidate evidence under the same activity object can be organized into the basic input set for conflict analysis.

[0149] Step 502: Construct an evidence conflict diagram.

[0150] In step 502, an evidence conflict graph is constructed based on the object-level evidence set formed in step 501. The evidence conflict graph can be represented using a graph structure, where nodes represent candidate evidence fragments and candidate emission activities, and edges represent conflict relationships, supporting relationships, derivative relationships, and temporal inheritance relationships between nodes. This graph structure representation allows disparate evidence relationships, inheritance relationships, supporting relationships, and conflict relationships to be unified within the same relational framework, facilitating subsequent classification, identification, and conflict resolution.

[0151] Furthermore, based on the field comparison results, source relationships, temporal relationships, and business process relationships among candidate evidence, the edges in the graph can be assigned relationship labels and strengths. For multiple field fragments split from the same original evidence, their common source can be identified through derived edges; for two candidate pieces of evidence that have different sources but consistent expressions, they can be represented by supporting edges to indicate mutual corroboration; for inconsistencies in key attributes, conflict edges are formed; for different candidate pieces of evidence that have relationships in terms of continuous time intervals, statistical periods, or historical record continuity, they can be represented by time inheritance edges. Through this step, subsequent conflict identification can be transformed from item-by-item comparison to relationship judgment based on graph structure.

[0152] Step 503: Identify numerical conflicts, time conflicts, unit conflicts, attribution conflicts, boundary conflicts, and conflicts in the applicable conditions of emission factors.

[0153] In step 503, a relationship analysis is performed on the evidence conflict diagram constructed in step 502 to identify different types of conflicts. Numerical conflict refers to different activity data values ​​for the same activity object under the same or corresponding conditions; temporal conflict refers to inconsistencies in the occurrence time or statistical period of the evidence records; unit conflict refers to incompatible numerical units or invalid conversion relationships; attribution conflict refers to inconsistencies in the attribution of facilities, organizational entities, or business operations; boundary conflict refers to inconsistencies in organizational boundaries or accounting boundaries; and emission factor application condition conflict refers to inconsistencies between the activity conditions reflected in the evidence and the conditions of the proposed applicable emission factor.

[0154] In one embodiment, the degree of conflict between any two candidate evidence nodes can be calculated to assist in conflict classification and subsequent resolution. Specifically, the following conflict degree expression can be used:

[0155] .

[0156] in, Indicates candidate evidence nodes With candidate evidence nodes The overall degree of conflict between them;

[0157] Indicates the degree of numerical difference;

[0158] Indicates the degree of time difference;

[0159] Indicates the degree of inconsistency in units;

[0160] Indicates the degree of attribution difference;

[0161] Indicates the degree of boundary difference;

[0162] Indicates the degree of difference in the applicable conditions for emission factors;

[0163] to These represent the weights of each conflicting term.

[0164] The degree of difference can be derived from the field parsing results in step 204, the activity attribute information in step 303, and the attribution results in step 402. The degree of difference in the applicability conditions of emission factors can be determined based on the consistency between conditions such as activity type, region, time, energy type, supplier information, and factor version and the applicability conditions of candidate emission factors. Overall conflict degree. Used to form the conflict edge strength in step 504 and participate in the comprehensive credibility assessment in step 602.

[0165] Step 504: Assign type labels and strength information to the conflict relationships in the evidence conflict diagram.

[0166] In step 504, based on the identification results of step 503, conflict edges in the evidence conflict map are assigned conflict type labels and conflict intensity information. The conflict type label is used to clarify that a conflict edge belongs to one or more of the following: numerical conflict, temporal conflict, unit conflict, attribution conflict, boundary conflict, or emission factor application condition conflict; the conflict intensity information is used to describe the degree of impact of the conflict on the determination of target activity data.

[0167] Optionally, minor conflicts involving only non-critical auxiliary fields can be assigned a lower conflict intensity; critical conflicts that directly affect the determination of activity data or the judgment of factor applicability can be assigned a higher conflict intensity.

[0168] Furthermore, the strength of conflict edges can be adjusted based on whether candidate evidence has supporting relationships, a common source, a temporal inheritance relationship, or an upstream / downstream business process relationship, in order to reduce the possibility of interpretable differences being misjudged as high-intensity conflicts. Through this step, the conflict diagram can not only represent the existence of conflict, but also the type and degree of impact of the conflict.

[0169] Step 505: Output the structured conflict analysis results.

[0170] In step 505, the conflict identification results obtained in steps 501 to 504 are organized into structured conflict analysis results. These structured conflict analysis results may include object identifiers, candidate evidence sets, conflict node pairs, conflict types, conflict intensity, support relationships, derivation relationships, temporal inheritance relationships, and relevant original evidence reference information. These structured conflict analysis results serve as the basic input for comprehensive credibility assessment and conflict resolution in steps 602 to 605.

[0171] Furthermore, the structured conflict analysis results can also include information such as whether the conflict can be corrected through rule adjustments, whether supplementary evidence is needed, and whether manual review is recommended. This step reduces multi-evidence relationships to standardized conflict analysis outputs, providing a unified interface for subsequent automated decision-making and audit retrospectives.

[0172] In one embodiment, steps 601-605 are described as follows:

[0173] Step 601: Extract the multidimensional evaluation factors required for comprehensive credibility assessment.

[0174] In step 601, multi-dimensional evaluation factors for comprehensive credibility assessment are extracted from the various results generated in the preceding steps. These evaluation factors include at least source reliability, resolution confidence, spatiotemporal consistency, equipment operational capability constraints, boundary matching degree, and evidence tampering risk. Source reliability can be determined based on the evidence source channel, access method, forming entity, and historical stability; resolution confidence can be determined based on the field-level and object-level confidence information output in steps 202 and 303; spatiotemporal consistency can be determined based on the consistency between evidence time and activity time, and evidence location and facility location; equipment operational capability constraints can be determined based on equipment rated parameters, historical operation records, production plan data, or preset capability thresholds, used to determine whether candidate data exceeds the reasonable operating range of equipment or facilities in the corresponding time period; boundary matching degree is used to determine whether the evidence falls within the target organization boundary and accounting boundary; evidence tampering risk can be characterized based on the consistency of the original evidence content summary, metadata anomalies, version change records, timestamp verification results, or signature verification results.

[0175] Furthermore, the aforementioned evaluation factors can be organized at the evidence level, field level, and object level to support subsequent scoring of individual candidate evidence as well as comparative analysis of multiple evidence sets under the same object. This step transforms the results of preceding multimodal analysis, attribution, and conflict analysis into a unified evaluation input.

[0176] Step 602: Calculate the overall credibility of the candidate evidence.

[0177] In step 602, based on the multidimensional evaluation factors extracted in step 601, the comprehensive credibility of each candidate piece of evidence under the same emission activity candidate is calculated. Since in a corporate carbon inventory scenario, a candidate piece of evidence may be reliable in its source but of average analytical quality; or it may have high analytical confidence but not perfectly match boundary conditions or equipment capacity constraints, a comprehensive credibility needs to be formed through multi-factor joint calculation, rather than relying on a single condition for decision-making.

[0178] In one embodiment, candidate evidence The overall credibility can be calculated as follows:

[0179]

[0180] in, Indicates the first The overall credibility of the candidate evidence;

[0181] Indicates the source reliability score;

[0182] This indicates the confidence score.

[0183] Indicates the spatiotemporal consistency score;

[0184] This indicates the equipment's operational capability constraint score;

[0185] Indicates the boundary matching score;

[0186] This indicates a risk score indicating evidence tampering;

[0187] Indicates the relationship with the first The set of other candidate evidence that conflict with each candidate piece of evidence;

[0188] Indicate candidate evidence With candidate evidence The degree of conflict between them;

[0189] This represents the influence coefficient of the corresponding conflict relationship;

[0190] , , , , , and Each of these represents a weight parameter. These weight parameters can be determined based on historical samples, preset rules, business experience, or model training results.

[0191] in, , , , , and The data are derived from the results of steps 103, 202, 303, 402, 503, and 601, respectively, with conflicting items. The conflict analysis results are derived from steps 503 and 504. Overall reliability. Used to sort and filter candidate evidence in step 603, and as a basis for conflict resolution.

[0192] Step 603: Execute conflict resolution and determine target activity data based on comprehensive credibility.

[0193] In step 603, conflict resolution is performed on multi-source candidate evidence under the same emission activity candidate based on the comprehensive confidence level obtained in step 602, thereby determining the target activity data. In one embodiment, corresponding conflict resolution strategies can be invoked for numerical conflicts, temporal conflicts, unit conflicts, attribution conflicts, boundary conflicts, and emission factor applicability condition conflicts.

[0194] For example, for unit conflicts, unit standardization and conversion verification can be performed first; for time conflicts, the judgment can be made based on the overlapping relationship of statistical periods and the time inheritance relationship; for ownership conflicts and boundary conflicts, the selection can be made based on organizational ownership, facility identification and accounting boundary rules; for emission factor applicability condition conflicts, the applicability of candidate factors can be re-evaluated based on conditions such as activity type, region, energy type, supplier information and factor version.

[0195] Optionally, for multiple candidate pieces of evidence within the same conflict group, they can be sorted from highest to lowest overall credibility, and further filtered based on conflict type, supporting relationships, temporal inheritance relationships, and object completeness. If a candidate piece of evidence has high credibility in key fields and is mutually supported by most auxiliary evidence, it can be given priority as the primary source of target activity data. If different pieces of evidence are more advantageous in different fields, they can be combined using a field-level selection method to form target activity data, while retaining the source identifier and corresponding decision basis for each field.

[0196] Furthermore, when there is no clearly dominant candidate evidence in the same conflict group, or when highly credible evidence still has key boundary conflicts, attribution conflicts, or factor applicability conflicts, the object can be marked as an object to be supplemented or an object to be manually reviewed.

[0197] Optionally, when the overall credibility of the target activity data is lower than a first preset threshold, a supplementary evidence request and a manual review mark are generated; when the overall credibility is lower than a second preset threshold, a conservative calculation path mark is generated; wherein, the second preset threshold is lower than the first preset threshold. Through this step, multi-source inconsistent evidence can be converged into target activity data that can be used for calculation, and corresponding processing marks can be generated for data with insufficient credibility.

[0198] Step 604: Record the basis for conflict resolution decisions and alternative evidence paths.

[0199] In step 604, the conflict resolution decision formed in step 603 is recorded. The record may include information such as which candidate evidence the selected target activity data originated from, why unadopted candidate evidence was excluded, the type of conflict, the overall credibility comparison result, and whether there are any requests for supplementary evidence, manual review markers, or conservative accounting path markers. This step is used to structurally preserve the decision-making process so that the data formation chain can be presented when generating the auditable evidence package in subsequent step 80.

[0200] Furthermore, candidate evidence that was not adopted but still has reference value can be recorded as alternative evidence paths. This allows for comparison of the differences between the currently used path and the alternative paths during subsequent audit reviews or evidence updates, without having to reconstruct the entire conflict scenario from the original evidence. This step enhances the interpretability of the conflict resolution results and the convenience of subsequent reviews.

[0201] Step 605: Output the target activity data and its credibility results.

[0202] In step 605, the target activity data and its corresponding credibility results are output. The target activity data may include activity values, units, time intervals, facility affiliations, organizational boundaries, and key conditions related to emission factor matching; the corresponding credibility results may include the overall credibility of the target activity data, a list of key adopted evidence, a conflict resolution summary, supplementary evidence request information, manual review markers, and conservative calculation path markers. This output serves as the input for emission factor matching and carbon emission calculation in step 70.

[0203] Furthermore, source references at the field level can be retained for the target activity data, enabling subsequent emissions calculations to clearly identify which candidate evidence each key field originates from and why it was adopted. Through this step, the enterprise carbon inventory processing workflow completes the transformation from multimodal raw evidence to target activity data.

[0204] In one embodiment, steps 701-702 are described as follows:

[0205] Step 701: Match emission factors and calculate carbon emissions based on target activity data.

[0206] In step 701, the corresponding emission factors are matched based on the target activity data output in step 605, and carbon emissions are calculated accordingly. The target activity data has undergone preliminary multimodal evidence analysis, activity object generation, evidence attribution, conflict identification, and conflict resolution processing, and can be used as input for calculation. To adapt the calculation results to specific activity scenarios, in this step, emission factors are retrieved, filtered, and matched based on conditions such as activity type, time range, regional attributes, energy type, supplier information, and factor version corresponding to the target activity data.

[0207] Optionally, the emission factor matching process can first limit the range of candidate factors based on activity type, and then further filter them based on conditions such as time, region, energy type, supplier information, and factor version. If multiple candidate factors still exist, the auxiliary attribute information in the target activity data and the confidence result output in step 605 can be combined to select the factor with a higher degree of fit as the target emission factor. If the applicable conditions corresponding to the current target activity data are insufficient to complete a unique factor match, multiple candidate factors can be retained as alternative paths, and the basis for factor selection and the reason for not using a factor can be recorded in the output results.

[0208] In one embodiment, for the first The emissions corresponding to each target activity data point can be calculated using the following method:

[0209]

[0210] in, Indicates the first Carbon emissions corresponding to each target activity data;

[0211] Indicates the first Data on individual target activities;

[0212] Indicates the relationship with the first Target emission factors matched with target activity data;

[0213] This represents the unit normalization coefficient, used to convert the current units of target activity data to the applicable units of emission factors.

[0214] in, The target activity data is derived from the output of step 605;

[0215] Derived from the factor matching results in this step;

[0216] The unit normalization rules derived from step 203 and the unit information applicable to the current factor.

[0217] Calculated Used for summarizing, organizing results, and encapsulating subsequent evidence packages in step 702.

[0218] Furthermore, for corporate carbon inventory tasks involving multiple target activities, the aforementioned factor matching and emission calculations can be performed separately for each target activity data point, retaining the activity object identifier, factor identifier, factor version, matching basis, and calculation path for each calculation record. This allows for subsequent verification of individual activity objects or audit retrospective review of the overall corporate inventory results, ensuring the identification of the specific calculation unit and its preceding evidentiary basis.

[0219] Step 702: Summarize the carbon emission calculation results and generate a calculation result record.

[0220] In step 702, the carbon emission calculation results generated in step 701 are summarized, and a calculation result record is generated. The calculation result record may include at least the target activity object identifier, activity data value, activity data unit, matching emission factor identifier, emission factor version information, calculation result, calculation path identifier, accounting boundary information corresponding to the target activity data, and credibility summary information. This step organizes the emission results scattered across different activity objects into a unified and manageable result set, providing input for the subsequent generation of an auditable evidence package.

[0221] Optionally, during the aggregation, the data can be categorized and organized according to organizational boundaries, facility dimensions, activity categories, time intervals, or accounting standards to adapt to the output needs of different corporate carbon audit scenarios.

[0222] Furthermore, for target activity data marked with supplementary evidence requests, manual review flags, or conservative accounting path flags, the corresponding identifiers can be retained simultaneously when generating calculation result records. This allows subsequent auditors or reviewers to clearly identify which calculation results originate from a stable chain of evidence and which still require further strengthening. Through this step, both enterprise-level carbon emission results can be output, and the basis for each result and its processing status can be attached.

[0223] In one embodiment, steps 801-803 are described as follows:

[0224] Step 801: Gather the original evidence index, parsing results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results to construct the content set of the auditable evidence package.

[0225] In step 801, the various intermediate and final results generated in the preceding steps are aggregated to construct a content set for an auditable evidence package. This content set may include the original evidence index set established in step 104, the structured parsing results formed in step 204, the activity attribution relationships established in step 402, the conflict resolution decisions recorded in step 604, the emission factor version used in step 701, the calculation path recorded in step 701, and the calculation results generated in step 702. By unifying the above content into a single evidence package content set, key processing nodes in the enterprise's carbon inventory process can have a traceable, locatable, and interpretable record basis.

[0226] Optionally, when constructing the content set, the data can be organized in a logical order of original evidence layer, parsing layer, active object layer, conflict resolution layer and calculation result layer, so that the internal structure of the evidence package corresponds to the actual processing flow.

[0227] Furthermore, for key objects in each layer, reference relationships can be established with objects in the upstream and downstream layers. For example, field parsing results reference their original evidence fragments, emission activity candidate objects reference their constituent fields, target activity data reference their adoption evidence and conflict resolution decisions, and calculation result records reference the target activity data and emission factors on which they are based.

[0228] Through this chain-like organizational method, subsequent audits or reviews can trace back from the final calculation result to the original evidence step by step, or trace the path of the original evidence to its impact on the final result.

[0229] Furthermore, when constructing the evidence package, supplementary information such as unadopted candidate evidence, reasons for exclusion, alternative evidence pathways, and suggestions for manual review can be retained. In this way, the evidence package not only records the final adopted content but also the corresponding reasons for adoption and the reasons for not adopting other candidates, thereby enhancing the interpretability of the processing and the convenience of review.

[0230] Step 802: Structure and encapsulate the auditable evidence package and establish version associations to generate audit traceability evidence dependencies.

[0231] In step 802, the content set formed in step 801 is structurally encapsulated and version relationships are established to generate an auditable evidence package. The structural encapsulation may include evidence package identifier generation, internal object numbering, hierarchical directory organization, reference relationship encoding, and content summary recording. The version relationships are used to record version information of the evidence package at different generation times, different supplementary evidence states, or different review states, enabling subsequent differentiation of data sources, conflict resolution decisions, and differences in calculation results between different versions.

[0232] Optionally, evidence packages can be constructed at the object level, task level, or period level. Object-level evidence packages can be organized for a single emission activity and are suitable for fine-grained problem investigation; task-level evidence packages can be organized for a complete carbon inventory task and are suitable for overall verification; period-level evidence packages can be centrally packaged for the results of multiple tasks during a certain accounting period and are suitable for phased archiving.

[0233] Furthermore, in cases where corporate carbon inventory tasks involve multiple updates, multiple supplementary certifications, or multiple reviews, version associations can be established for evidence packages to describe the inheritance, revision, and substitution relationships between different versions of evidence packages.

[0234] Furthermore, audit traceability evidence dependencies can be generated based on the structured encapsulation and version association relationships. These dependencies describe the relationships between original evidence, parsed fields, emission activity candidates, target activity data, conflict resolution decisions, emission factor versions, calculation paths, and calculation results, enabling auditors to trace back from any calculation result to its corresponding original evidence and processing decisions. This step ensures that the evidence package possesses good organization, hierarchy, and historical traceability.

[0235] Step 803: Generate timestamps, digest hashes, and signature information for the auditable evidence package, and output the audit traceability interface.

[0236] In step 803, a timestamp, digest hash, and signature information are generated for the auditable evidence package generated in step 802, and an access interface or query result that can be used for audit traceability is output. The timestamp is used to identify the generation time or sealing time of the evidence package; the digest hash is used to characterize the digest value of the evidence package content so as to determine whether the content of the evidence package has changed; the signature information is used to sign the evidence package or its digest hash to support evidence package integrity verification and sealing status verification.

[0237] Optionally, the sealed content may include at least the evidence package identifier, version identifier, generation time, sealing subject identifier, original evidence citation relationship, analysis result citation relationship, conflict resolution decision record, emission factor version, and calculation result record.

[0238] Furthermore, the digest hash can be generated based on at least one of the following: the original evidence index, parsing results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results in the evidence package. It should be noted that the timestamp, digest hash, and signature information in this step are primarily used to enhance the integrity verification and historical consistency management of the evidence package, and their specific implementation methods are not limited.

[0239] Furthermore, when outputting the audit traceability interface, it can support access methods such as result-based backtracking, evidence-based querying, activity-based querying, or version comparison. For example, auditors can start from a certain calculation result and view its corresponding target activity data, conflict resolution decisions, and original evidence level by level; they can also start from a certain original evidence and trace which activity objects it participated in and which carbon emission results it ultimately affected. Through this step, the present invention completes a closed loop of processing from multimodal original evidence access to structured accounting, and then to audit encapsulation and traceability output.

[0240] Optionally, refer to Figure 2 , Figure 2 This is a schematic diagram of the structure of the enterprise carbon inventory and auditable evidence generation system based on multimodal AI provided by the present invention. The enterprise carbon inventory and auditable evidence generation system based on multimodal AI includes...

[0241] Module 210 is used to acquire multimodal evidence data from enterprise carbon inventory checks;

[0242] The parsing module 220 is used to parse the multimodal evidence data using a multimodal AI model and extract information related to the carbon inventory, including activity data, emission sources, units of measurement, time intervals, facility affiliation, organizational boundaries, and applicable conditions for emission factors.

[0243] The candidate object generation module 230 is used to generate emission activity candidate objects based on the analysis results, and to establish evidence source identification, data confidence level and activity attribute information for each emission activity candidate object;

[0244] The association module 240 is used to construct the association between evidence and emission activity candidates, and to attribute evidence from different sources and different modalities to the corresponding emission activity candidates.

[0245] Module 250 is used to construct an evidence conflict map for multi-source candidate evidence corresponding to the same emission activity candidate object, and to identify numerical conflicts, time conflicts, unit conflicts, attribution conflicts, boundary conflicts and emission factor applicability condition conflicts.

[0246] The conflict resolution module 260 is used to comprehensively evaluate the credibility of the multi-source candidate evidence based on source reliability, resolution confidence, spatiotemporal consistency, equipment operating capability constraints, boundary matching degree and evidence tampering risk, and to perform conflict resolution on the evidence conflict graph based on the comprehensive credibility to determine the target activity data.

[0247] Calculation module 270 is used to calculate the enterprise's carbon emissions based on the target activity data and the matched emission factors;

[0248] The evidence package generation module 280 is used to generate an auditable evidence package, which includes an original evidence index, parsing results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results.

[0249] Please see Figure 3 , Figure 3 An embodiment diagram of an electronic device provided in accordance with the present invention. For example... Figure 3 As shown, this embodiment of the invention provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, it performs the following steps:

[0250] Obtain multimodal evidence data for corporate carbon inventory;

[0251] The multimodal evidence data is analyzed using a multimodal AI model to extract information related to the carbon inventory.

[0252] Based on the analysis results, candidate emission activities are generated, and evidence source identification, data confidence level, and activity attribute information are established for each candidate emission activity.

[0253] Establish the correlation between evidence and emission activity candidates, and attribute evidence from different sources and modalities to the corresponding emission activity candidates;

[0254] For multiple candidate evidences corresponding to the same emission activity candidate, construct an evidence conflict map and identify numerical conflicts, temporal conflicts, unit conflicts, attribution conflicts, boundary conflicts and emission factor applicability condition conflicts.

[0255] Based on source reliability, resolution confidence, spatiotemporal consistency, equipment operating capability constraints, boundary matching degree, and evidence tampering risk, a comprehensive credibility assessment is performed on the multi-source candidate evidence, and conflict resolution is performed on the evidence conflict graph according to the comprehensive credibility to determine the target activity data;

[0256] Calculate the company's carbon emissions based on the target activity data and the matched emission factors;

[0257] Generate an auditable evidence package, which includes an original evidence index, parsing results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results.

[0258] Please see Figure 4 , Figure 4 An embodiment diagram of a computer-readable storage medium provided in accordance with an embodiment of the present invention is shown. Figure 4As shown, this embodiment provides a computer-readable storage medium 400 on which a computer program 311 is stored. When the computer program 311 is executed by a processor, it performs the following steps:

[0259] Obtain multimodal evidence data for corporate carbon inventory;

[0260] The multimodal evidence data is analyzed using a multimodal AI model to extract information related to the carbon inventory.

[0261] Based on the analysis results, candidate emission activities are generated, and evidence source identification, data confidence level, and activity attribute information are established for each candidate emission activity.

[0262] Establish the correlation between evidence and emission activity candidates, and attribute evidence from different sources and modalities to the corresponding emission activity candidates;

[0263] For multiple candidate evidences corresponding to the same emission activity candidate, construct an evidence conflict map and identify numerical conflicts, temporal conflicts, unit conflicts, attribution conflicts, boundary conflicts and emission factor applicability condition conflicts.

[0264] Based on source reliability, resolution confidence, spatiotemporal consistency, equipment operating capability constraints, boundary matching degree, and evidence tampering risk, a comprehensive credibility assessment is performed on the multi-source candidate evidence, and conflict resolution is performed on the evidence conflict graph according to the comprehensive credibility to determine the target activity data;

[0265] Calculate the company's carbon emissions based on the target activity data and the matched emission factors;

[0266] Generate an auditable evidence package, which includes an original evidence index, parsing results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results.

[0267] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0268] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0269] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for enterprise carbon inventory and auditable evidence generation based on multimodal AI, characterized in that, include: Obtain multimodal evidence data for corporate carbon inventory; The multimodal evidence data is analyzed using a multimodal AI model to extract information related to carbon inventory, including activity data, emission sources, units of measurement, time intervals, facility affiliation, organizational boundaries, and applicable conditions for emission factors. Based on the analysis results, candidate emission activities are generated, and evidence source identification, data confidence level, and activity attribute information are established for each candidate emission activity. Establish the correlation between evidence and emission activity candidates, and attribute evidence from different sources and modalities to the corresponding emission activity candidates; For multiple candidate evidences corresponding to the same emission activity candidate, construct an evidence conflict map and identify numerical conflicts, temporal conflicts, unit conflicts, attribution conflicts, boundary conflicts and emission factor applicability condition conflicts. Based on source reliability, resolution confidence, spatiotemporal consistency, equipment operating capability constraints, boundary matching degree, and evidence tampering risk, a comprehensive credibility assessment is performed on the multi-source candidate evidence, and conflict resolution is performed on the evidence conflict graph according to the comprehensive credibility to determine the target activity data; Calculate the company's carbon emissions based on the target activity data and the matched emission factors; Generate an auditable evidence package, which includes an original evidence index, parsing results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results.

2. The method for enterprise carbon inventory and auditable evidence generation based on multimodal AI according to claim 1, characterized in that, The step of using a multimodal AI model to analyze the multimodal evidence data includes: The multimodal evidence data is subjected to layout type identification to obtain layout type identification results corresponding to plain text pages, table pages, mixed pages, or scanned pages; Based on the layout type recognition result, the corresponding text parsing pipeline, table parsing pipeline, or image parsing pipeline is invoked; Textual and visual features are fused and encoded to extract entity fields related to carbon inventory, and semantic disambiguation and standardization are performed on the entity fields based on a carbon accounting terminology library.

3. The method for enterprise carbon inventory and auditable evidence generation based on multimodal AI according to claim 1, characterized in that, The activity attribute information includes activity type, activity data value, unit of measurement, time interval, facility identifier, organization affiliation, spatial location, data source, candidate emission factor conditions, and evidence summary identifier.

4. The method for enterprise carbon inventory and auditable evidence generation based on multimodal AI according to claim 1, characterized in that, The construction of the correlation between evidence and emission activity candidates, attributing evidence from different sources and modalities to corresponding emission activity candidates, includes: Candidate associations between evidence and emission activity candidates are generated based on semantic similarity, temporal overlap, facility matching, unit of measurement matching, and business process relationships. The candidate associations are filtered to generate an evidence-emission activity association index.

5. The method for enterprise carbon inventory and auditable evidence generation based on multimodal AI according to claim 1, characterized in that, In the evidence conflict graph, nodes represent candidate evidence fragments and candidate emission activities, and edges represent conflict relationships, support relationships, derivation relationships, and temporal inheritance relationships between nodes.

6. The method for enterprise carbon inventory and auditable evidence generation based on multimodal AI according to claim 5, characterized in that, The multi-source candidate evidence is comprehensively evaluated for credibility based on source reliability, resolution confidence, spatiotemporal consistency, equipment operational capability constraints, boundary matching degree, and evidence tampering risk. Conflict resolution is then performed on the evidence conflict graph based on the comprehensive credibility to determine the target activity data, including: The corresponding conflict resolution strategies are invoked based on numerical conflicts, time conflicts, unit conflicts, attribution conflicts, boundary conflicts, and conflicts of applicable conditions for emission factors. For each candidate piece of evidence, calculate the source reliability score, analytical confidence score, spatiotemporal consistency score, equipment operating capability constraint score, boundary matching score, and evidence tampering risk score respectively. A comprehensive credibility score is generated based on each score. Based on the comprehensive credibility, target activity data is determined, and conflict resolution decision records are generated.

7. The method for enterprise carbon inventory and auditable evidence generation based on multimodal AI according to claim 6, characterized in that, When the overall credibility of the target activity data is lower than the first preset threshold, a supplementary certification request and a manual review mark are generated. When the overall credibility of the target activity data is lower than the second preset threshold, a conservative accounting path marker is generated. Wherein, the second preset threshold is lower than the first preset threshold.

8. The method for enterprise carbon inventory and auditable evidence generation based on multimodal AI according to claim 1, characterized in that, The calculation of corporate carbon emissions based on the target activity data and matched emission factors includes: Match the corresponding emission factors based on activity type, time, region, energy type, supplier information, and factor version; Carbon emission calculation results are generated based on the target activity data and the emission factors.

9. The method for enterprise carbon inventory and auditable evidence generation based on multimodal AI according to claim 1, characterized in that, The steps for generating an auditable evidence package include: The original evidence index, analysis results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results are versioned and archived. Generate audit traceability evidence dependencies based on versioned archive results; Generate timestamps, digest hashes, and signature information for the auditable evidence package.

10. A system for enterprise carbon inventory and auditable evidence generation based on multimodal AI, characterized in that, The method for enterprise carbon inventory and auditable evidence generation based on multimodal AI as described in any one of claims 1 to 9 includes: The acquisition module is used to acquire multimodal evidence data for enterprise carbon inventory checks; The analysis module is used to analyze the multimodal evidence data using a multimodal AI model and extract information related to the carbon inventory, including activity data, emission sources, units of measurement, time intervals, facility affiliation, organizational boundaries, and applicable conditions for emission factors. The candidate object generation module is used to generate emission activity candidate objects based on the analysis results, and to establish evidence source identifiers, data confidence levels and activity attribute information for each emission activity candidate object; The association module is used to build the association between evidence and emission activity candidates, attributing evidence from different sources and modalities to the corresponding emission activity candidates. The module is used to construct evidence conflict diagrams for multiple source candidate evidence corresponding to the same emission activity candidate, and to identify numerical conflicts, temporal conflicts, unit conflicts, attribution conflicts, boundary conflicts, and emission factor applicability condition conflicts. The conflict resolution module is used to comprehensively evaluate the credibility of the multi-source candidate evidence based on source reliability, resolution confidence, spatiotemporal consistency, equipment operating capability constraints, boundary matching degree and evidence tampering risk, and to perform conflict resolution on the evidence conflict graph based on the comprehensive credibility to determine the target activity data. The calculation module is used to calculate the enterprise's carbon emissions based on the target activity data and the matched emission factors; The evidence package generation module is used to generate an auditable evidence package, which includes the original evidence index, parsing results, activity attribution relationships, conflict resolution decisions, emission factor versions, calculation paths, and calculation results.

Images